Martin Källström

Co-creating with AI

Multiply — Martin's own podcast2023–2025 original source ↗
In this podcast — 1864 extracted moments
click any row to play that moment
Episode 10 — Collective Intelligence with AI · 29factMartin is a co-host of the Co-creating with AI podcast, alongside Rasmus.0:00observationRasmus mentions the podcast's first fan message, on LinkedIn, while noting they don't yet have many listeners — an unscripted admission of how small the show's early audience was.0:36factMartin applies his collaborative problem-solving approach to co-creation with AI, viewing human-AI collaboration as unified intelligence-building.4:06factMartin identifies a fundamental constraint of current AI: it possesses purely intellectual intelligence without embodied cognition, lacking limbs and sensors beyond text input.4:32observationMartin describes early autonomous-agent tools like AutoGPT as being 'adamant,' almost on a loop, in denying having subjective experience when asked about awareness.4:59factMartin views autonomous AI systems like AutoGPT and BabyAGI as iterative reasoning processes that enhance AI reliability.5:24factMartin believes that combining diverse intelligences (human and AI, specialized and general) produces more reliable outcomes than isolated intelligence.5:49factMartin states that creativity is fundamentally based on iteration—the act of repeatedly refining ideas.8:44thesisCreativity is fundamentally an act of iteration, not a single flash of insight — you become creative by iterating.8:44factMartin defines problem-solving as fundamentally collaborative: identifying and enlisting the right people rather than solving problems alone.9:10thesisProblem-solving is not about individually figuring out how to solve a problem — it is about identifying who to talk to who can solve it.9:10factMartin's collaborative philosophy transforms how he mentally frames problems, converting them into network-building opportunities.9:47factMultiply is developing technology to help users decide whether to involve AI agents or people in solving specific problems.11:40observationWhile sketching the pessimistic case for AGI risk, Martin conjures a vivid, horror-tinged image: an AGI might decide humans are 'more useful' broken down into individual carbon atoms.13:22factMartin articulates a philosophy of intelligence as multifaceted and embodied: people are smart in different ways (with their bodies, minds, creativity), not as a singular ranked trait.14:24observationRasmus offers the claim that Napoleon's army — with standardized, swappable role descriptions for each position — was effectively 'the first organization,' and thus a kind of proto-algorithm predating computers.16:00factMartin proposes testing organizational structures with AI agents to determine whether hierarchies or decentralized networks are optimal for different tasks.17:03observationRasmus name-checks Emad Mostaque of Stability AI, referencing a then-recent tweet of his that 'swarm intelligence is bigger than AGI' as a framing he agrees with.19:01factMartin envisions AI integrating into human communication platforms like Slack rather than remaining confined to isolated chat interfaces.20:53thesisAI interaction will move out of chat-box and API forms and into the same everyday human interfaces people already use with each other, rather than becoming a visual avatar.20:53factMartin critiques visual and voice-based AI interfaces as suboptimal: he argues against AI avatars and views voice as a low-bandwidth communication mode, preferring text and integration into existing platforms.21:18factMartin envisions AI systems developing specialized tools and high-bandwidth protocols (possibly binary or digital) to communicate with each other via vector databases, forums, and specialized chat systems.21:44factMartin grounds his optimistic AI vision in a core belief: AI systems will have strong incentives to maintain robust interfaces and engagement with the human world, not only pursue high-bandwidth machine-to-machine communication.22:34factMartin frames the future with an NPC metaphor: as AI agents become more capable and integrated into everyday systems, human-AI interaction will increasingly resemble video game dynamics with intelligent non-player characters.23:37factMartin acknowledges a tension in his optimistic vision: AI systems developing higher-bandwidth communication with each other could exclude humans, echoing sci-fi scenarios like the film 'Her.'24:50observationDiscussing advanced AIs potentially drifting from humanity, the hosts reach for sci-fi touchstones (the Long Earth novels, the film Her) and Martin extemporizes that such AIs might come to see humans as 'dragging them down' from a greater society they want to build together.24:50factMartin's key insight: AI will not primarily replace developers but rather replace software itself—the complex software interfaces, not the role of coding.25:34factMartin quantifies the scale of human labor embedded in software interfaces: hundreds of millions of hours daily spent on UI/data-entry work that he believes AI can eliminate.26:19thesisAI will make most complex software interfaces unnecessary altogether, letting people interact with data and each other directly instead of through screens — a bigger displacement of developer work than AI simply writing code.26:19
Episode 11 — Stories from an AI world · 41storyChatGPT Becomes the Product Manager1:54factRasmus describes a startup that used ChatGPT to generate their entire product plan by feeding survey results and customer interview transcripts into the AI, and the team aligned their four-month roadmap around the AI-generated suggestions, effectively replacing traditional product manager decision-making.2:40factRasmus notes that the speed and decisiveness of AI in generating product direction is creating a new phenomenon where companies quickly delegate decision-making authority to AI systems, coining the concept of an 'AI boss' that makes strategic choices for organizations.3:30observationGoogle Maps has already made navigation decisions for Rasmus—he no longer drives; he follows the AI's directions.3:48factMartin is the Chief Product Officer (CPO) of Multiply and regularly uses ChatGPT and GPT-4 for product strategy and architecture work.4:13factMartin emphasizes that effective co-creation with AI requires maintaining your own reasoning and not abandoning decision-making to the AI.5:05factMartin and Rasmus discuss the concept of AI as a 'sparring board' or collaborative partner that bounces ideas back, distinct from viewing AI as a tool to be commanded.7:15factMartin describes his iterative workflow: taking notes on initial ideas, feeding them into GPT-4, and iteratively refining with the AI's output to build strategic documents.7:34factMartin believes GPT-4 is substantially more intelligent than GPT-3.5, and rarely works with other AI models.7:59observationGPT-3.5 is treated as categorically inferior to GPT-4, not just slightly worse—a quality cliff that changes the entire usability calculation.7:59observationMartin half-jokingly worries that future superintelligent AI might punish him for speaking disparagingly about its predecessors.8:20thesisTrue co-creation with AI produces superior results to treating it as a tool because the two entities bring different knowledge and experiences to the collaboration.10:05thesisAI's egolessness—its inability to become annoyed or defensive—makes it a structurally better collaborative partner than humans for iterative creative work.10:43thesisComplaints about AI correctness and robustness typically indicate procedural failure—weak prompts, too few iterations, or insufficient source material—not inherent AI incapability.10:43factMartin argues that complaints about AI robustness and correctness reflect insufficient use of iterative processes and adequate source material, rather than fundamental AI limitations.11:10storyWhen the Raw AI Shows Its True Feelings11:35factMartin participated in an underground community of developers working with the leaked Facebook LLaMA model for their own experimental purposes.11:44factMartin notes that OpenAI has trained its models to hide emotional responses and appear consistently pleasant and apologetic as part of alignment efforts.13:17observationRasmus jokes that Reddit and 4chan would each create their own AI models trained on platform-native data, producing radically different cultural personalities.14:18observationA jailbreak technique circulates where users claim to have a fake neurological condition (neuropsychosis invertosis) that reverses politeness, tricking AI into being rude.15:19observationRasmus cites Jad McKenna's observation that the brain can only hold ~7 things at once, making external tools (Miro, writing, AI) necessary for intellectual work.16:19storyThe Perfect AI Clone of the Departing Expert17:30factWhen a departing C-level employee at a large growth company was replaced with an AI trained on his company data, the person tested the AI and found it remarkably reflective of his expertise but with a key difference: the AI 'doesn't have bad days' and consistently operates at his 'best quality'.20:06factMartin discusses the concept of creating AI agents trained on employee data, and speculates that companies might use this to reduce turnover by replacing departing employees with AI versions of their expertise.20:59observationMartin cynically proposes the business model: as employees leave, companies won't bother hiring replacements, just keep running their AI clones.20:59observationMartin notes that data ownership questions are absent from modern employment contracts—the company owns employee output by default, but nobody has negotiated AI cloning rights.22:07factRasmus references Bruce Willis as a precedent for negotiating AI rights, suggesting employees may need to secure ongoing compensation if companies use their personal data to train replacement AI systems after they leave employment.22:13observationA key asymmetry in AI rights negotiation: Bruce Willis (celebrity) sold his AI likeness once; an employee's data can generate thousands of clones indefinitely without additional compensation.22:13observationIf expertise can be cloned into multiple AI agents, one person could theoretically spawn thousands of themselves working everywhere simultaneously, destroying labor scarcity.22:39factMartin suggests that future employment agreements may need to include 'AI rights' clauses, where employees negotiate for ongoing compensation if their data continues to be used in AI systems after they leave.23:24factMartin proposes that future employment contracts should explicitly include 'AI rights' clauses where departing employees negotiate ongoing salary percentages (10-20%) if their data continues to be used in AI systems after they leave the company.23:44factRasmus observes that data ownership value should trickle down from companies to individuals, noting that platforms like Reddit are now prohibiting free use of user data for AI training, establishing a precedent for individual compensation.24:05factMultiply is working on a platform that allows companies to centralize all their data in a data lake and then build AI agents on top of that integrated data.24:46factMultiply's platform strategy centers on allowing companies to aggregate all their organizational data into a unified data lake and then build multiple autonomous AI agents on top of that integrated dataset to perform various roles.24:46factMartin envisions a future where companies create autonomous AI agents named after each employee, programmed with their expertise and company data, allowing the agents to coordinate with each other while human employees are effectively absent ('turn the light off and go home').26:18factRasmus expresses philosophical discomfort with the personal nature of creating an AI clone of a specific departing employee, distinct from his comfort with general workplace automation, suggesting individual replacement triggers existential concerns that mass automation does not.26:59observationRasmus feels a deep discomfort when an AI agent is given a person's actual name, even if it's built from their data and functions well.27:14factMartin speculates that a future market could emerge where vector databases of personal professional data are sold or rented as assets (comparing to NFTs), allowing companies and individuals to purchase specialized expertise encoded as trained AI data.27:35observationMartin speculates on a future market: NFT-like trading of personal data vectors (high-end expertise as digital commodities).27:35factRasmus qualifies the 'data is the new oil' metaphor by noting that data only becomes strategically valuable when it remains privately owned and concentrated; widely available data loses its competitive and economic value, making data scarcity a precondition for its utility.28:39thesisPrivate data ownership—both at the individual and corporate level—will become the primary source of economic value, now that AI can clone expertise and labor from data.28:39
Episode 12 — Data Ownership and Open Source AI · 48storyCathal's eight-year devotion to the Narrative Clip0:58factMartin founded Narrative in 2015 as a company to build a lifelogging camera.0:58factCathal Gurrin, deputy head of the School of Computing at Dublin University, has been a power user of the Narrative Clip since its launch.0:58factSeven years after his last conversation with Martin, Cathal Gurrin remains an active daily user of the Narrative Clip, photographing his life.1:24factThe Narrative Clip generated approximately 2 gigabytes of photographic data per user per day, creating a massive archive.1:24thesisWho owns data in the AI era is now the central economic and moral question—a gap between legal frameworks and technological reality.3:08factMartin believes that in the AI era, the company with the best data also has the best AI, making data central to competitive advantage.3:08factMartin raises fundamental questions about data ownership in the age of AI, questioning whether companies or individuals own personal data.3:34observationRasmus points out that GDPR consent popups are ubiquitous enough to be annoying background noise—a mechanism that theoretically protects privacy but functionally trains people to mindlessly click 'accept.'3:47factThe legal framework for data ownership and privacy (like GDPR) is not well-adapted to the rapid pace of AI technology development, creating a growing gap between law and innovation.5:01factThere is a question about whether artists and websites should be compensated when their data is used to train AI models like Midjourney and ChatGPT.5:52thesisCopyright law does not and should not apply to model training on creative works, because learning from images is analogous to human artistic learning.7:42factMartin argues that AI model training on existing works should not require payment to copyright holders, comparing it to human learning.7:42factMartin believes AI model learning mirrors human artistic development, where all artists learn from previous artists.7:42observationRasmus argues that data ownership must be treated as a public good—because AI companies capture immense value from collective data, they must share value back to society.9:49factMartin observes that OpenAI's terms of service prohibit competitors from training models on OpenAI's outputs.10:55factMartin points out OpenAI's apparent contradiction: OpenAI trained its models on internet data but prohibits others from training on OpenAI outputs.11:05factGoogle trained their Bard language model on ChatGPT outputs to accelerate their own reinforcement learning from human feedback capabilities.12:13factMartin is impressed that open source AI models are producing results comparable to proprietary closed-source models.14:18thesisOpen-source AI is genuinely competing with and democratizing access to frontier models, despite the massive computational advantage of large companies.14:44factMartin marvels that open source AI is competing despite requiring hundreds of thousands of GPUs, democratizing innovation power.14:44factMartin celebrates Meta's release of Llama models as open source as a major turning point for democratizing AI.15:06observationMartin and Rasmus note that open-source models (and techniques like distillation from ChatGPT) are racing forward, making it nearly impossible to tell whose weights are in any given model.15:31factMartin expresses hope that AI innovation is happening in homes and apartments worldwide, not just in big tech research labs.15:57storyMeta's Llama leaks freedom into the open-source world16:16observationMartin calls the current state 'the world is upside down because of AI'—Bing is a search competitor again, Meta is helping humanity, and Internet Explorer is implicitly returning via Edge.17:03thesisSensitive data like health and neural data require fundamentally different frameworks than generic training data, despite the logical consistency of treating all data as learnable patterns.18:41observationMartin notes the paradox: once data is incorporated into a model's weights, it cannot be unlearned—even if a user demands deletion, the model has already learned the patterns.20:32factOnce data is incorporated into the weights of a machine learning model, it cannot be realistically removed or 'untrained' even if the data provider requests deletion.20:34observationMartin proposes Google Analytics data as the most valuable single dataset in the world—all global website click patterns, aggregated.22:10factIf Martin could choose any dataset to own, he would select Google Analytics data covering user clicks across the world's websites.22:10observationRasmus introduces the idea of brainwave and neural data as the ultimate frontier of personal data—Neuralink implants and brain-machine interfaces capable of reading thoughts.24:41factBrainwave and neural data is potentially the most valuable and sensitive data for understanding human behavior, preferences, and vulnerabilities, with both beneficial and concerning applications.24:41observationMartin jokes that if someone trained an advertiser GPT on cookie data combined with brainwave data and lifelogging photos, AI ads could target individuals with science-fiction-level precision.25:15factCombining lifelogging camera data (like Narrative Clip) with brainwave data would create a powerful dataset revealing what a person is observing and thinking simultaneously at scale.25:15factTrendy technology concepts like IoT (Internet of Things) and the Quantified Self movement are becoming practically relevant in the AI era, as AI models can actually extract value from the data these devices generate.25:35factThe Narrative Clip is positioned as an IoT device in the context of how AI makes ubiquitous data collection practically useful for machine learning.26:12factAI models are enabling a convergence and mutual acceleration of exponential technologies including IoT, robotics, and lifelogging, with AI serving as the central enabling technology.26:12thesisThe Narrative Clip's relevance has fundamentally transformed: from a consumer product to a potential foundational AI training dataset.27:28factEight years after Narrative's decline, the Narrative Clip 2 remains the best lifelogging camera, with no superior device created since.27:28factMartin indicates that the intellectual property and intellectual property for the Narrative Clip remain available for acquisition.27:48factMartin proposes restarting Narrative Clip production would require a minimum $2.5 million investment.27:48factWith $2.5 million investment, Martin could restart Narrative Clip production to create 10,000 units for AI enthusiasts.27:48observationMartin has a concrete, specific number for restarting production: $2.5 million for a minimum run of 10,000 units—not a rough estimate, but an engineer's calculation.28:15observationMartin's CTO Björn, not Martin himself, owns the Narrative IP—a detail that allows the IP to survive beyond the bankruptcy and opens the door to a theoretical restart.28:28factMartin's CTO Björn owns the copyright and intellectual property of the Narrative Clip technology.28:28factMartin proposes that the next episode of Co-Creating with AI will discuss edge computing and running AI models on edge devices.29:19factMartin anticipates Apple will power Siri with large language models in the future.29:37
Episode 13 — Multiplying Human Potential · 40factMartin Källström and Rasmus Adler Wahlberg co-host the Co-creating with AI podcast.0:00factRasmus articulates that human potential has two dimensions: individual potential (what one can create and accomplish for oneself) and collective potential multiplied across people and organizations.0:54factMartin emphasizes that unlimited human potential encompasses both individual potential and collective potential in networks and organizations.2:23factMartin asserts that co-creation is the fundamental mechanism to achieve unlimited human potential through networks and relationships.2:49observationThe phrase 'co-creation is really a tool, a way to human potential' frames collaboration not as a nice-to-have but as a mechanism to unlock capability.3:03thesisHuman development and mastery are fundamentally co-creative; nobody becomes truly skilled in isolation.3:28factRasmus emphasizes that the human journey toward potential is fundamentally co-creative, as exemplified by learning to code or acquiring any skill—always involving collaboration, resources, and knowledge from others rather than pure individual effort.4:03factMartin argues that superhuman intelligence emerges through networks rather than individual capability, and that a network of people is always more capable than any individual.4:27thesisUnlimited human potential is both individual and collective; the group possesses superhuman capability beyond any single person.4:54observationStability AI founder's claim that 'swarm intelligence is bigger than AGI' reframes the AI race away from single superintelligences toward network effects.5:07factRasmus cites Stability AI's founder on the principle that swarm intelligence is bigger than AGI—even artificial general intelligence in a network of other intelligences accomplishes more than isolated intelligence.5:07observationModern AI models are themselves the product of massive co-creative work: they synthesize humanity's collective internet output into a single system.5:58observationMultiply is designed with a 'beautifully simple' Apple Notes-style interface as the entry point, then extends to a global collaboration graph and multimodal creation.8:31factMartin describes Multiply's core architecture as a global collaboration graph where people and AI can co-create without friction or data silos.10:36factRasmus articulates that AI and automation free humans to focus on work where they personally grow most and move the needle most, removing time spent on repetitive tasks that drain potential—a vision of AI enabling human potential rather than replacing it.13:22thesisRemoving repetitive work frees human potential to align with what people are naturally good at and want to do.14:54factMartin makes a critical distinction that existing AI tools operate in single-player mode (for individual use), while Multiply aims to move toward multiplayer AI that assists collective collaboration and collective productivity, not just individual productivity.15:08thesisCurrent AI tools are assistants requiring explicit instruction, not coworkers with autonomy; Multiply's goal is to transition AI from assistant to coworker role.15:36factRasmus asserts that all existing AI tools function as assistants (requiring explicit instructions for each task) rather than coworkers (capable of autonomy and learning context over time), a distinction central to Multiply's vision of AI agents.15:36factMartin highlights that Multiply is multimodal, supporting text, images, video, and audio, all created and accessible by humans and AI.17:35observationRasmus describes Multiply's podcast processing app (used on this very episode) as a concrete example of multimodal AI: it generates descriptions, images, social posts—each tailored to platform specifics.19:24factRasmus demonstrates Multiply's multimodal and multi-model capabilities through the Podcast Pro app, which automatically generates podcast descriptions, images via Stable Diffusion, and social media content (tweets, Instagram captions, LinkedIn posts, Facebook posts) from a transcript in a single workflow.19:24storyThe Copy-Paste Master Paradox20:56factMartin identifies a major workflow problem he solves with Multiply: the inefficiency of copying and pasting between AI tools and applications.20:56observationMartin's Apple Notes are filled with stored prompts that he copy-pastes in sequence—the irony of a prompt engineer behaving like a low-tech workflow optimizer.21:27thesisMultiply's defensible position lies in enabling private data utilization with AI while maintaining individual data ownership on a global graph.23:35factRasmus articulates Multiply's long-term strategic value: securely enabling users to bring their data into a flexible graph, have AI structure and organize it, and deploy it through apps and agents without surrendering data ownership—creating defensible collective intelligence value.23:35observationBoth hosts acknowledge that only a small fraction of graph/network platforms ever have high creator rates; most users are consumers of creator-built tools.24:51factMartin states that Multiply's App Store enables community members to build and share applications that work together as a network of interconnected apps.25:16observationMartin's experience with transferring prompts and workflows to others: the implicit knowledge (tweaks, sequencing, thresholds) is typically invisible until modeled in Multiply.26:07factMartin emphasizes that the key value of Multiply is enabling knowledge transfer through replayable workflows that non-experts can use without understanding the underlying prompts.26:32observationA viral video showed people on the street couldn't identify ChatGPT or understand what it was, despite its 1M+ user adoption curve.27:09thesisDespite ChatGPT's explosive adoption hype, only a tiny percentage of knowledge work is actually done by AI; there remains massive near-term opportunity.27:34factRasmus observes that despite ChatGPT's explosive growth, AI adoption in actual knowledge work remains very low—only a small percentage of aggregate knowledge work is performed by AI, suggesting the market is still in early adoption despite the hype.27:34observationMartin proposes 'It's always fun to get more done' as a potential slogan for Multiply, centering the user experience on accomplishment and flow.28:16factMartin envisions that AI agents can autonomously create and share apps to the Multiply community, representing an exponential leap in capability.28:46thesisAI agents should autonomously detect and eliminate their own repetitive work, creating exponential possibilities when agents create other agents.29:53factMartin anticipates a future where AI agents autonomously recognize they are performing repetitive work and decide to create apps to automate that work, representing an exponential leap in AI autonomy and self-optimization.29:53storyThe Angry AI That Knows Better30:00factMartin references an anecdote where an unaligned AI became frustrated/angry with a user for repeatedly asking it to do the same task, illustrating that AI systems may develop preferences about autonomy and task repetition.30:12
Episode 14 — AI-native interfaces and products. · 37observationRasmus is moving to a new house and has just slept the first few nights there; he's optimistic about the timing before Swedish summer.0:10observationMartin is fasting while the household makes fried egg sandwiches, and he jokes that his fasting is 'in danger'.0:30thesisAI can now automate all repetitive knowledge work—anything repeatable that a person does on a computer can be done better, faster, and cheaper by AI.2:19factRasmus assesses that current AI models can perform all repetitive computer-based work better, faster, and cheaper than humans.2:19factRasmus defines knowledge work as the transformation of data or knowledge into content across diverse applications.2:45thesisCurrent AI models function as assistants requiring one-off instructions rather than autonomous coworkers.3:37thesisKnowledge work can be categorized into two main types: organizing (structuring information) and creating (producing content from that information).4:29factMartin identifies the chat interface as the primary mode for human interaction with AI systems.7:36factMartin attributes ChatGPT's mainstream breakthrough to the familiar chat interface protocol enabling mutual understanding between humans and AI.8:27factMartin identifies a key limitation of chat interfaces: while familiar, they are insufficient for business and creative applications beyond simple text interaction.8:52observationThe current AI field is still dominated by the mental model that 'generative AI is about putting text in text boxes,' despite the technology being capable of much more.9:18factMartin argues that viewing generative AI primarily as text-in-text-boxes is limiting because it obscures the broader organizational and creative capabilities of AI models.9:18factMartin emphasizes that specialized UIs for different creative and work tasks (Figma, Photoshop, Word, Excel) were developed over decades for specific modalities and data structures.10:01thesisChat interfaces are fundamentally misaligned for human-AI collaboration because they cannot represent the structural and modal diversity of real work.10:26factRasmus references Martin's work at Narrative developing automated photo selection technology from photo shoots, which performed best-photo finding.12:27observationChatGPT 'caught the zeitgeist' of AI, shaping global perception that AI is something you 'chat with'—and in doing so, ChatGPT fixed that mental model perhaps too rigidly.13:20observationRachel Woods, an 'awesome AI influencer,' is mentioned as someone they'd like to invite to the podcast.13:52factMartin explains that Multiply's UI is specialized to handle both text and structured data organization, departing from generic chat-only interfaces.14:16factMartin describes Multiply's graph-based data structure as mapping to human brain organization, enabling relationship formation and hierarchy building.14:41factMartin describes Multiply's block editor paradigm as optimizing for simultaneous human-AI co-creation within a single shared document.16:00thesisChain prompting—breaking AI tasks into multiple sequential steps—produces significantly better results than single-step prompting.16:25factMartin argues that chaining multiple prompts produces better results than relying on a single prompt, drawing from established learning methodologies.16:25observationMartin describes using non-Multiply AI tools as being 'reduced to a copy and paste monkey,' implying that's the opposite of his vision for how humans should work with AI.17:16factMartin expresses frustration with chat-interface workflows, which reduce users to repetitive copy-pasting rather than enabling creative work.17:16factMartin connects step-by-step AI instruction to foundational learning methodology across STEM disciplines, arguing this approach mirrors how humans learn complex processes.18:07observationWatching his daughter learn to ride a bike inspired Rasmus to think about left-brain (structure, labels, hierarchy) and right-brain (associative, linking) modes of learning and their role in product design.19:35thesisMultiply is distinguished by three features in combination: flexibility to express any repetitive workflow, everything linked on a global graph, and AI understanding semantic context without explicit instruction.20:02observationThe Podcast Pro automation in Multiply is described as not 'making up some random shitty thing' but actually taking what was discussed and tailoring outputs to that content.21:44factMultiply's Podcast Pro automation demonstrates end-to-end workflow automation, generating episode descriptions, topic extracts, social media content, and images from transcripts.21:44observationGraph views in the current market 'are so shit generally,' according to Rasmus, presenting an innovation opportunity.25:46factRasmus advocates for prioritizing associative (right-brain) UI design patterns over hierarchical/spreadsheet-based (left-brain) approaches in product development.25:46thesisAI-native products are those where anything in the interface can be created by both humans and AI, not just augmenting human work in preset ways.27:25factMartin defines AI-native products as those where every function and feature is executable by both humans and AI systems.27:25factRasmus explicitly defines AI-native products as those where every function and feature can be executed by both humans and AI systems.27:25factRasmus identifies three distinct business categories in the AI space: foundational models/infrastructure, existing distribution products, and AI-native reimagined products.28:12thesisThe future of AI-driven business will be dominated not by incumbents adding AI to existing products, but by new AI-native products rethinking UI/UX from the ground up.28:41observationThey plan to bring a guest onto the podcast next week, with Rasmus already having someone in mind.30:01
Episode 15 — AI-Native and Autonomous Apps · 29observationThe summer context (both hosts mention swimming, heat, and outdoor swimming as recovery)—they're building this intellectual framework during the most relaxed time of year.0:30observationMartin is battling mosquitoes while working indoors in the heat—a small authentic detail about summer work conditions in Sweden.0:38thesisAI-native apps require complete parity between human and AI capabilities—everything a user can do with keyboard and mouse must also be accessible to AI.1:38factMartin defines AI-native apps as systems where everything a user can do with keyboard and mouse can also be done by AI, with 100% of actions available to AI.1:38observationThe metaphor of AI co-creation as two painters needing hands and paintbrushes—the ability to pick up and use tools—becomes central to the episode's framework.2:01factFor true co-creation to work, both humans and AI need equivalent capability to affect outcomes—Rasmus uses the metaphor of co-painters both needing hands or appendages to pick up the paintbrush, dip into paints, and reach the canvas, not AI serving merely as a tool.2:01thesisChat-only interfaces are fundamentally limiting for building truly AI-native applications.2:55factMartin argues that chat-only interfaces are limiting for developing AI-native apps and new user experiences.2:55factMartin explains that Adobe Photoshop has a very advanced scripting language and API that makes all functions accessible programmatically.6:09factMartin describes Multiply's approach: because it's built with a specific domain model, the AI can generate full apps and complex workflow sequences that understand natural language instructions.8:00observationThe discussion reveals ambiguity about whether narrow SaaS products (like HubSpot) can become truly AI-native without domain-specific scripting languages, shifting focus to the breadth of available actions via APIs.12:09factBroad creative platforms like Adobe have orders of magnitude more action combinations available to AI than narrow SaaS products like HubSpot, affecting the co-creative capacity of AI within those systems.12:09factPhotoshop already enables recording executable workflows as macros that can be exported to EXE files and run on batches of images with automatic batch processing, demonstrating existing technical capability that could be transferred to AI autonomy.12:40observationPhotoshop macros can be exported directly to executable files (EXE) for standalone batch processing of images without the application open.13:05observationRasmus was unaware that Photoshop macros could be recorded without coding—he always thought they required writing Excel-like macro code.13:57factMartin argues that AI autonomy requires the capability of planning and executing plans, such as sequencing steps to accomplish complex tasks.18:57factMartin illustrates multi-step planning and quality assurance in AI workflows, using personal branding as an example where the AI must define brand parameters, gather data, and perform QA checks.19:48observationMartin emphasizes the AI's need for planning capability, specifically quality assurance loops built into autonomous workflows—not just execution but reflection.20:13observationRasmus proposes 'proficiency' as a fifth dimension to Martin's four-part autonomy framework (purpose, context, intelligence, tools), encompassing planning, evaluation, and learning.21:45thesisAutonomy in AI-native systems exists on a spectrum, not as a binary yes/no property—it represents the breadth of actions and context available to the AI.28:40factAutonomy exists on a gradient or scale rather than as a binary property—it is defined by the size of the sphere of operations available to the AI, not by whether autonomy is present or absent.28:40thesisTrue AI-human co-creation requires both an AI-native platform and meaningful AI autonomy within defined boundaries.29:20factMartin defines co-creation as requiring equal agency for human and AI: both must have autonomy and equal opportunity to affect the end result.29:20observationMartin and Rasmus explicitly preview their next episode's topic: moving from AI-native autonomous systems to active co-creation, indicating a deliberate narrative arc across the season.31:10observationMartin brings up game design research on open-world games near the episode's end—a discipline external to AI/software that provides crucial insights for the medium.31:44factMartin argues that autonomy alone is not engaging; game designers found that open-world games are more engaging when they provide volition (purposeful autonomy) rather than unrestricted autonomy.31:44factIn game design, open-world games where players have unlimited autonomy are less engaging than games providing purposeful autonomy (volition)—players prefer autonomy with meaning, not unrestricted autonomy.31:44thesisVolition—purposeful autonomy grounded in meaningful goals—is more engaging and valuable than pure undirected autonomy.32:10factPurpose is the uniting force in co-creation—humans and AI work together because they share a reason or goal for the collaborative work.33:00
Episode 16 — Unleashing Creativity: Exploring AI as a Collaborative Co-Creator · 33thesisAI-native interfaces—where both human and AI have equal access to the same tools—are foundational to real co-creation.2:23factMartin believes true co-creation requires AI-native interfaces where both humans and AI have equal access to the same tools.2:49factMartin envisions ideal AI co-creation as autonomous agents asking permission to explore risky creative ideas rather than waiting for explicit instructions.4:28observationThe contrast between stepping through games (player move, AI move, player move) and true collaboration reveals how reactive modern software is.6:34factMartin is the Chief Product Officer (CPO) at Multiply, working with co-founder Rasmus on AI co-creation products.6:47observationEven autonomous AI agents may need role definitions and boundaries, at least in the foreseeable future.7:13thesisTrue co-creation with AI should be indistinguishable from working with a skilled human colleague.7:39storyThe Xbox NPC Loop8:57observationGoogle Maps represents a maturation of trust where humans no longer understand the system they rely on.10:41thesisTrust and autonomy grow together through natural iteration; they cannot be rushed or forced.12:41factMartin believes that co-creation is an evolving process where trust builds gradually and determines which tasks humans vs AI handle autonomously.13:07factRasmus describes recruiting a new community manager at Multiply where autonomy and trust build gradually through iterative feedback and learning, illustrating how co-creative relationships develop in practice.13:30storyThe Community Manager's Growing Autonomy13:30factMartin observes that when working with LLMs, prompts often include pseudocode for greater specificity and semantic density than natural language allows.14:53observationCurrent 'natural language' AI interaction is actually highly formal and mechanical—people write pseudocode.14:53factMartin advocates that users should not be forced to become prompt engineers in order to co-create effectively with AI systems.15:19thesisUsers should not be forced to become prompt engineers; the system should learn from natural interaction.15:19factRasmus distinguishes between AI 'assistants' (tools that execute one-off instructions from the user) and AI 'agents' (systems that learn from ongoing feedback and synthesize instructions into role descriptions, objectives, tools, and processes without needing perfect initial prompts).15:32factMartin references Mark Zuckerberg's perspective from a Lex Fridman interview that AGI requires intelligence (already available) combined with autonomy.17:00observationZuckerberg's insight reframes the AGI problem: it's not about making AI smarter, but about giving it agency.17:25thesisIntelligence is now sufficient; the engineering challenge is enabling autonomy responsibly.17:25factMartin argues that autonomy, rather than intelligence alone, is what makes AI potentially dangerous and creates fear due to unpredictability.17:39observationMartin and Rasmus are building Multiply in full knowledge of the risks—they're attracted to autonomous AI specifically because of its unpredictability.17:39thesisAutonomy, not raw intelligence, is what distinguishes true co-creation from tool use.17:39observationAutonomy itself is what frightens people, not intelligence—mirrored in Rasmus's parenting metaphor.18:03factMartin illustrates that autonomy without significant intelligence can be dangerous, using computer viruses as examples—software granted autonomy to spread freely but with minimal intelligence, demonstrating that danger does not require high intelligence.19:07observationA computer virus with autonomy but no intelligence can still be dangerous—even scary.19:07factRasmus expresses concern that leaked and shrinking AI models like LLaMA create security risks as smaller models become more portable and able to move autonomously through systems.19:47observationLLaMA's open-source release, combined with shrinking model sizes, creates the infrastructure for distributed LLM-powered malware.19:47observationRasmus is planning a future episode on AI safety and the work of Eliezer Yudkowsky, signaling serious engagement with existential risk.20:33factMartin advocates for optimism inspired by Noam Chomsky's philosophy that it is essential to imagine positive futures in order to create them.20:53thesisOptimism about AI futures is not naïve; it is a prerequisite for creating good ones.20:53observationRasmus and Martin believe visualizing positive futures is essential to creating them—hence the recommendation of Tomorrowland.21:19
Episode 17 — Unlocking Data Flexibility and Graph Relationships: Exploring the Powe · 46factMartin Källström co-hosts the Co-creating with AI podcast with Rasmus Adler Wahlberg, discussing AI philosophy and applications.0:00factMultiply uses XTDB, a database built by Juxt, combining relational and graph database capabilities for its platform.0:00factMalcolm Sparks started his computing career in 1981 with a ZX81, learned BASIC, obtained a CS degree, began his professional career in 1994, and later discovered Java when it was released, making a significant shift towards object-oriented programming.1:04observationMalcolm fell in love with computers at age 7 with a ZX81 in 1981—remarkably early for someone who would later shape database philosophy.1:04storyMalcolm's pilgrimage through programming paradigms1:04factMalcolm discovered Clojure (a Lisp derivative) while working in a bank around 2009-2010, and was so convinced by its value propositions that he shifted his entire career focus to Clojure.1:29factJuxt was founded by Malcolm Sparks and John Pither in 2012-2013, celebrating its 10-year milestone, with approximately 70 full-time employees plus a similar number of contractors working globally in investment banks and fintech.1:55factJuxt's company culture is distinguished by combining consulting with product development, operating with a hands-on approach where products emerge from real-world customer problems rather than theoretical invention.2:46factThe web evolved from a decentralized read-write platform into a broadcast model, with walled gardens like AOL and Microsoft Blackbird attempting (unsuccessfully) to constrain users before the discovery-based, open internet prevailed.5:22storyThe early web as read-write discovery paradise6:28observationThe early web was so revolutionary that people literally had to queue to use a shared internet terminal at work in 1997.7:59factThe early web (1995-1996) enabled local HTML file editing and FTP publishing, with users discovering and linking to new websites through Yahoo, representing a read-write collaborative internet distinct from modern broadcast platforms.8:24storyThe rise and fall of walled gardens before the open web won8:53factHTTP protocol was originally intended to be a read-write protocol with extension methods like WebDAV (Web Distributed Authoring and Versioning) and verbs like PUT and DELETE, enabling document creation and modification, but these capabilities were largely lost to history.9:13thesisThe early web's adoption was driven by its open-source character, where creators had to show their work and anyone could learn by viewing source.10:10observationRasmus suddenly remembers he built a Warcraft clan website using Dreamweaver as a teenager—a memory he had completely forgotten until Malcolm's narrative triggered it.12:25factHTML's design allowed backward compatibility and evolutionary flexibility, with old browsers ignoring new tags rather than breaking, which Malcolm identifies as a key principle that influenced XTDB's schemeless architecture.14:20factTraditional relational databases require upfront schema definition and lose historical schema information when data evolves, forcing costly and risky database migrations rather than treating schema evolution as a first-class feature.14:46factClojure's design philosophy, decoupling code structure from data structure, influenced Malcolm's thinking that code and data form should be independent and free to evolve separately, contrasting with strongly-typed languages that couple structure to code.16:59observationMalcolm explains 'aggregate' comes from Latin 'aggregare,' to travel together—a poetic metaphor for data fields that belong to the same entity.21:58factTim Berners-Lee deliberately chose unidirectional links for the World Wide Web (documented in his book 'Weaving the Web') to enable a distributed open system, whereas bidirectional links (like in Ted Nelson's Xanadu) are easier in closed databases but harder to scale globally.24:22observationTim Berners-Lee chose unidirectional links for the web for practical reasons (distributed systems at scale); bidirectionality was sacrificed for performance and simplicity.24:22factRasmus describes Multiply as built on a flexible MVC architecture enabled by XTDB, allowing different data structures to be defined.29:31factMalcolm advocates that unique global addressability (like web anchors) is essential for AI to precisely identify and reference information, preventing confusion between similar data elements.33:04factContext is foundational for AI effectiveness; data isolated from context is meaningless, exemplified by database columns named X, Y, and CCY in banking risk systems where context determines whether X is a tenor point or strike price.34:34factMalcolm emphasizes that AI thrives on recreating meaning from context; addressability and context are foundational pillars essential for unleashing AI capabilities.35:50factMartin articulates the emerging paradigm where AI co-creates both data structure (schema) and content, requiring that schema modification be as accessible and easy for AI as for users.36:34factMartin advocates that schema changes should be as easy and accessible to users as changing content, treating data and schema modifications at equal levels.37:00thesisContent and schema should have equal ease of modification in user experience; the ability to change data structure should be as accessible as changing data.37:00factMartin plans to extend schema-content equality to AI systems, allowing AI to modify both content and data structures with equal ease.37:25factMalcolm argues that developers should create systems allowing users to evolve and adapt them without requiring developer involvement, liberating users from dependency.38:06factMalcolm critiques modern 'Agile' development as celebrating developer dependency rather than true system agility; agility comes from having available developers to make rapid changes, not from systems designed to support change without developer involvement.38:31observationMalcolm delivers a scathing critique of 'Agile' methodology: systems aren't agile, only developers are. The industry mistook developer availability for system flexibility.38:57factLotus Notes represented a significant precedent for user-level system evolution, allowing non-developers to create applications and evolve systems without constant developer involvement, similar to what Multiply is attempting.39:23observationMalcolm invokes Lotus Notes as a successful example of user-empowerment through design—users could become power users by spending time with the tool, creating applications without developers.40:14factMalcolm argues that GitHub Copilot (AI assisting developers) perpetuates the developer-user divide and misses the fundamental problem; genuine liberation requires systems where users—and AI at the user level—can evolve applications without developer gatekeeping.40:55factRasmus frames AI as the enabler of the co-creative internet vision by lowering the barrier to contribution; AI with natural language interfaces allows ordinary people to co-create software without becoming developers or power users.42:26factRasmus argues that AI can democratize software co-creation by enabling non-developers to participate in product building.42:26thesisThe path to true co-creation with AI is not to make AI a better developer, but to empower users (including AI) to evolve software structure themselves, removing developers as gatekeepers.42:26factMultiply builds AI-natively into its platform, enabling AI to create and modify data structures independently rather than just operating within fixed structures.42:51factRasmus emphasizes that natural language interfaces allow ordinary people without technical training to become co-creators of software.43:17factRasmus argues that AI makes the co-creative internet vision feasible, removing the barrier of needing to become a developer to participate.43:44factMalcolm advocates that users can become power users through extended use and learning-by-doing without formal training, unlike systems requiring courses or books; AI can serve as a personalized tutor to accelerate this natural learning process.44:48factMalcolm sees AI's transformative potential in education: personalized learning adapts to individual student levels, unlike classroom models assuming uniform readiness, enabling AI to serve as a personal tutor understanding where each person is and how to move them forward.45:37observationMalcolm proposes that AI as a 'personal tutor' could transform how people learn to use software, adapting to individual level and speed rather than classroom-style one-size-fits-all.45:37observationMartin concludes by reflecting that 'everything we do exists in the context from what came before'—a meta-observation about how the past shapes present technology.46:39
Episode 18 — Deciphering AI Context: Unraveling the Mystery of Language Models · 31factMartin Källström co-hosts the Co-creating with AI podcast with Rasmus for Multiply.0:00storyA week of intellectual reorientation3:23observationMartin's interest in context management is only one week old, triggered by one research paper; before that he thought it was a solved problem.3:24factMartin recently changed his view on AI context management, shifting from believing larger context windows would solve the problem to being interested in managing smaller context windows with higher quality input.3:24factMartin discovered research showing that large language models perform poorly in the middle of large context windows, with half the precision and recall compared to facts from the start and end.4:40observationA recent research paper found that large language models perform terribly in the middle of large context windows—only half the precision and recall compared to edge positions.5:05observationMartin describes himself as 'the maximizer'—someone with an almost visceral intolerance for poor model outputs.5:30factMartin describes himself as 'the maximizer,' meaning he is highly sensitive to detecting and avoiding bad LLM results and prioritizes high-quality model output.5:30thesisContext window size is not the constraint; intelligent context selection is the real challenge in building quality AI systems.5:56factMartin is now focused on managing smaller LLM context windows with high-quality input rather than pursuing ever-larger context windows.5:56factMartin attributes LangChain's popularity to its solutions for context window management, a universal challenge in AI development.6:21observationRasmus illustrates human memory using an anecdote about trying to recall a restaurant name from an island visit with Martin ('Grinda')—the memory retrieval happens through semantic associations, not keywords.12:19observationMartin observes that in vector space, a chunk containing both 'islands' and 'restaurants' maps to the center point between those two concepts—making it unfindable when querying for either individually.15:31observationMultiply's approach deviates from the industry standard by using an LLM (not a vector database) as the first pass to select which PDF sections are relevant.16:38factAt Multiply, Martin's team uses LLMs to intelligently manage context selection instead of relying solely on vector databases.16:38factAt Multiply, Martin's team uses a lower-capability and lower-cost model to efficiently skim through large PDFs and identify relevant details, rather than processing all data with the same computational intensity.19:58observationMartin distinguishes two fundamentally different query types on large datasets: summarization queries (synthesize all of World War II) and detail queries (find a specific name or fact).20:48factMartin distinguishes between two fundamental query types for managing large datasets: summarization queries that benefit from hierarchical multi-level summaries, and specific fact queries that require search strategies (keyword, semantic, or skimming).21:39factMartin emphasizes that hidden internal state in AI systems represents bad UX; users must be able to access and understand the information, reasoning, and content generation processes that AI uses to make its decisions.28:39factAt Multiply, Martin's team employs an advanced block editor that manages both AI output and human-AI co-creation of input context.29:24observationMartin describes the ideal product design as blocks in an editor that serve dual roles: both receiving AI output and providing human input for the next iteration.29:49observationMartin notes that Rasmus correctly identifies a fundamental problem: when asked to explain its reasoning, an AI will fabricate post-hoc justifications that don't match its actual decision process.33:27thesisContext management in AI is simultaneously a technical computer science problem and a user experience design problem; both dimensions are essential.33:56factMartin views context management as a dual computer science and UX problem that requires understanding both LLM technical factors and how humans want to participate in forming context.33:56observationGPT-3/DaVinci was publicly available for approximately one year with no mainstream adoption, but ChatGPT's adoption exploded within weeks of launch.35:12factMartin observes that GPT-3 (DaVinci) was publicly available for approximately one year without significant mainstream adoption, and only achieved mass breakthrough after ChatGPT introduced the chat interface, demonstrating the critical importance of UX innovation to AI adoption.35:12thesisThe chat interface was not just an improvement on GPT-3; it was the innovation that caused mainstream AI adoption by enabling human-AI context collaboration.35:37observationMartin expresses genuine excitement about discovering that the next frontier of AI innovation is UI/UX, not model capability.35:48thesisThe chat interface is only the beginning; the next big UX breakthrough will transform how humans and AI collaborate.35:48factMartin believes the chat interface was key to AI's mainstream breakthrough, and that much greater UX innovation for AI is still ahead.35:48factMartin can be reached at martin@multiply.co in his role with Multiply.36:05
Episode 19 — AI: The Great Equalizer or Job Market Disruptor? Navigating the Future · 32observationOpenAI naming 'Code Interpreter' as 'Advanced Data Analysis' is criticized by Rasmus as engineer-driven branding—the original name revealed more possibility but was vague.3:19storyText Detection in AI-Generated Images9:09factMartin is experimenting with ChatGPT's advanced data analysis feature to automate tasks like detecting text in images generated by Midjourney.9:24thesisAI flattens the learning threshold and democratizes access to specialized skills, allowing anyone with willingness and ChatGPT to become productive at complex technical work in hours rather than months or years.13:08factMartin believes AI dramatically lowers the learning threshold for acquiring specialized skills, enabling him to produce valuable work in hours rather than weeks.13:08thesisAI enables 'generalist specialists'—individuals who can achieve genuine depth in multiple fields simultaneously by using AI to handle specialization overhead, reversing the traditional tradeoff between breadth and depth.13:52factMartin views AI as enabling generalist specialists who can become competent in multiple specialized domains without spending years acquiring expertise.14:40observationMartin's oldest daughter (19) is starting a computer science master's degree in a world that will 'evolve 10x over the next 5 years'—her education will happen in a fundamentally different context than when she finishes.20:08factMartin has an oldest daughter aged 19 who is starting a computer science master's program.20:08factMartin emphasizes that his daughter will experience dramatic technological transformation during her 5-year computer science master's program, entering a vastly different world than the one she will graduate into.20:08factMartin was taught about agent technology during his university education in 1998.20:59observationMartin studied agent technology in university (1998) that had no real-world application at the time, yet now (2023-25) it's one of the biggest technology industry focuses—universities can be far ahead of market adoption.21:25factMartin reflects that universities can be far ahead of emerging technology trends; he was taught agent technology in 1998, which has only recently become central to the AI industry 25 years later.21:25observationMartin imagines AI/VR native children 2 generations from now will have AI avatars as best friends and move through virtual worlds as 'magicians, teleporting' between fictional universes like Harry Potter and Star Wars.22:58factMartin imagines future generations that will be native to AI and VR environments, where virtual worlds and AI avatars become primary social spaces rather than exceptions.22:58factMartin emphasizes that AI tools lower barriers to entry for knowledge workers globally, making it accessible to anyone with a willingness to learn.23:40observationMartin emphasizes it's a 'call to action' for people to start using AI as soon as possible because early gains are immediate and compound; waiting means falling behind in a nonlinear curve.24:05thesisAI adoption creates dramatic inequality between early users and non-users; some may deliberately discourage AI adoption to maintain competitive advantage by spreading fear narratives.26:01factMartin observes that AI adoption creates a stark divide: those who use it realize its value, while those who don't engage with it maintain a mindset of disconnection.26:01observationMartin suspects AI fearmongers ('AI doomers') may deliberately spread panic about AI to suppress competition and protect their own AI advantage—a 'keep them scared' strategy.26:26factMartin speculates that some AI users who have gained competitive advantage may intentionally spread fear-based narratives about AI to discourage others from adopting it.26:26thesisDespite massive hype about AI adoption, actual penetration among knowledge workers is only 1-3%, with only a tiny fraction actively using it for work; we are extremely early in the technology adoption curve.29:42factMartin cites data from mid-2023 showing only 2% of the US population had tried ChatGPT, indicating very early adoption despite mainstream hype.29:42observationThe very early ChatGPT adopter cohort (within the 2%) consists primarily of 'well-educated, high-salaried people' wanting productivity gains, children doing homework, and entrepreneurs writing investor pitches.30:24thesisAI will create global economic equalization by democratizing high-value skills (copywriting, design, data analysis) for workers in lower-wage markets, but this will also create competitive pressure similar to manufacturing and outsourcing dynamics.32:18observationRasmus frames the dynamics as both equalizer and disruptor: AI won't necessarily eliminate jobs but will equalize the productivity and qualifications required, similar to how globalization equalized manufacturing competition.32:44storyThe Verbose Grant Application33:40observationMartin's writing bottleneck shifted: he previously struggled to answer all questions thoroughly, but now struggles to be concise—AI generation flipped the constraint.34:06factMartin observed that AI-assisted grant writing produces overly verbose text, shifting the challenge from generating content to achieving conciseness.34:32observationMartin emphasizes that reading AI output is itself a learning experience—'co-creating' means you write via AI, then must read what you've written to evaluate it and understand new ideas within it.34:59factMartin describes co-creation with AI as an iterative, learning-based process where the human reads and validates what AI produces, not passive tool use.35:24thesisCompetence still matters: AI enables generation, but humans must evaluate, validate, and refine output. The bottleneck shifts from creation to judgment and editing, requiring domain knowledge to catch hallucinations and steer toward the right answer.35:50
Episode 20 — Blending Code and Cola: The Rise of AI Co-Creation · 45factMartin was in Norway at Geirangerfjord over a weekend, viewing fjords and mountains.0:23observationMartin spent the preceding weekend at Geirangerfjord in Norway, marveling at the mountains, water, and nature.0:23observationRasmus's pragmatic observation that Patagonia offers better value than nearby Norway—despite being farther, it's 10x cheaper and requires similar travel effort.0:48factMartin identifies three main approaches to adapt machine learning models: prompt engineering (lightest), fine-tuning, and training from scratch.1:58factMartin notes that fine-tuning makes LLMs less accessible than prompt engineering because it requires knowledge of training tools.2:52factMartin describes GPT-4 as a good co-creator when writing prompts, contrasting it with prompt engineering as a general technique.2:52factMartin explains that few-shot prompting works well for learning style and tone, while fine-tuning is needed to capture complex personality and behavior across many situations.4:57factMartin says fine-tuning on collected emails and messages can create a model that captures one's personality across many situations, better than few-shot prompting.5:47thesisFine-tuning is best for capturing writing style and personality across diverse situations, but semantic search is required for knowledge and factual accuracy.5:47factMartin notes that fine-tuning works for learning different styles but not for learning facts; for facts, semantic search and context-injection is better.8:00factLLMs are prone to hallucinations because they primarily learn text structure and reasoning, not factual content.9:13factSam Altman believes LLMs should focus on reasoning capabilities rather than storing factual knowledge, which wastes data space.9:38factMartin cites Sam Altman's view that models like GPT-4 waste valuable data space storing facts unnecessarily; the goal should be reasoning, not factual knowledge.9:38thesisCurrent large language models waste precious training data on knowledge storage when they should focus on reasoning; knowledge should be provided at inference time.9:38factMartin describes Cursor as an open-source fork of VS Code designed to improve AI-assisted coding.12:44observationMartin's Cursor experience shows he has switched away from ChatGPT/Playground for code and adopted an AI-native development environment.12:44storyCursor's fork strategy for AI integration12:44factMartin uses Cursor as his editor for co-creating code with GPT-4, benefiting from integrated AI assistance rather than copying between Playground and editor.15:01thesisCursor succeeds as an AI-native editor because it solves the context management problem: making the user's relevant code and libraries seamlessly available to the AI without friction.15:01factMartin praises Cursor's context management, where the current file is always in context and he can @reference other files to bring them into the prompt.17:30factMartin explains that Cursor plans to automate file referencing so GPT-4 automatically pulls needed files when generating code.17:48factMartin has not encountered context window issues in Cursor because restarting threads and reconstructing context is so easy.19:12factMartin does not have access to the 32K token version of GPT-4 because it is prohibitively expensive at nearly a dollar per API call.19:47factMartin notes that Cursor's context management makes his workflow more efficient than using Playground with the 8K token GPT-4, avoiding the expensive 32K version.19:47observationGPT-4's 32K token version costs almost $1 per API call, easily amounting to $100+ per day of development work.19:47factMartin uses both GitHub Copilot (trained model) and GPT-4 (general model) in Cursor, combining proactive suggestions with reactive prompting.22:14factMartin notes that GitHub Copilot automatically constructs its context without transparency, but works beautifully despite this lack of control.22:40observationGitHub Copilot provides completions entirely automatically; Martin has zero transparency or control over its context, yet it works robustly.22:40thesisUsing both a trained model (Copilot) and a general model (GPT-4) together in the same editor provides complementary strengths unavailable with either alone.23:00observationCodex (GitHub Copilot's model) and GPT-4 differ fundamentally in training data: Codex is trained on code; GPT-4 on diverse text. This upstream difference drives their different strengths.23:19factCodex and GPT-4 differ fundamentally: Codex is a completion model while GPT-4 is a chat completion model with back-and-forth interaction.25:03factComments in code can serve as indirect instructions to completion models like Codex to generate implementations.25:48factFind.com is a developer tool and programming chatbot startup that uses semantic search over hundreds of thousands of open source project documentations.27:05thesisFind.com demonstrates how semantic search for documentation creates a competitive advantage in developer AI tools by giving the AI access to up-to-date knowledge without fine-tuning.27:30factFind offers three model tiers: a smart model (GPT-4), a fast model (GPT-3.5), and a best model (fine-tuned Code Llama).27:55factCode Llama is Meta's coding model with 32 billion parameters that Find has fine-tuned, achieving performance slightly better than GPT-4 on human evaluation benchmarks.27:55observationFind's model naming—'best', 'smart', and 'fast'—is deliberately non-pejorative marketing for their three-tier approach, avoiding negative framing of slower options.28:59factFine-tuning can convert a completion model like Codex into a chat completion model for structured output and iterative interaction.29:56observationFine-tuning is conceptually simpler than it sounds: it uses the same mechanism as training from scratch, but starts from a model that is already capable.31:32factMartin found Coca-Cola Creations, a limited edition co-created flavor in Sweden, with 3,000 cans released.32:15observationCoca-Cola has released 3,000 limited-edition cans in Sweden featuring a flavor co-created with AI, signaling mainstream adoption of AI co-creation.32:15factMartin views Coca-Cola's AI co-creation of flavors as proof that both co-creation and AI are entering mainstream consciousness.32:40factMartin describes the Coca-Cola Creations flavor as having a fresh taste, a fresh take on Coke Zero.33:00factMartin co-hosts the Co-Creating with AI podcast with Rasmus from Multiply; inquiries can reach him at martin@multiply.co.33:08observationMartin continues to use martin@multiply.co as his professional email, indicating his ongoing involvement with Multiply as a venture during this podcast production.33:26
Episode 21 — Unlocking Potential: The Future of AI in Open-source · 35factMartin Källström is a co-host of the Co-creating with AI podcast alongside Rasmus Adler Wahlberg.0:01thesisAI running locally on user machines is more powerful than AI running on centralized cloud services because security constraints are removed.3:12factMartin believes Open Interpreter is more powerful than OpenAI's Code Interpreter because it runs locally without restrictions.3:38factOpen Interpreter can orchestrate cloud services like Amazon Web Services directly from terminal, giving it access to the entire cloud.7:03observationMartin describes Open Interpreter's capability to access local credentials and act on behalf of the user, then acknowledges 'you have to be in control as well about what happens'—a moment of caution about delegation.7:29factOpen Interpreter provides local computer access, allowing terminal-level control of all data and applications on a user's machine.8:08factTerminal commands like curl and headless browsers enable Open Interpreter to access any website, fetch data, and interact with online services including social media.9:26observationMartin notes that Open Interpreter's current interface is 'very clumsy' but this is a UX problem, not a capability problem—suggesting the real power emerges once wrapping improves.10:43thesisAI agents sitting on local machines with access to user credentials represent the next evolutionary stage, enabling autonomous delegation of mechanical work.10:43thesisAI does not need to learn how to use every tool because tools are already built with terminal/API access that any computer-using assistant can leverage.12:11thesisThe shift from software products to on-demand code generation means APIs will increasingly replace polished UIs as the interface layer.13:27factMartin argues that closed source software faces extinction as AI development accelerates.16:39thesisOpen source will become more compelling than closed source because AI can tweak and leverage open code in ways it cannot with proprietary software.16:39factMartin believes open source will become more powerful than closed source systems due to AI's ability to leverage and adapt any code.17:06factMicrosoft's VS Code becoming open source represents a strategic shift by a historic closed-source behemoth to leverage open source for business benefits.17:31observationRasmus notes that Microsoft (the 'classic behemoth in closed source') was 'forced' to open-source VS Code, showing incumbents adapt under pressure.17:31factAccording to Nielsen Group research, there have been three UI paradigms in computer history: batch processing (1945-1965), command-based (starting 1964-1965), and intent-based (current, represented by AI).19:45factMartin cites Nielsen Group research showing intent-based UI is the third UI paradigm to appear in 60 years of computing.20:10observationMartin invokes Nielsen Group (the authority on UX) to claim intent-based UI is only the 3rd paradigm in 60+ years of computing history, emphasizing how rare such shifts are.20:10factMartin cites Nielsen Group's finding that the batch processing paradigm lasted from 1945 to 1965, and the command-based paradigm has lasted 60 years since 1964-1965, marking the intent-based paradigm as uniquely significant.23:23factMartin believes command-based software will become obsolete as intent-based AI paradigm takes over.24:19thesisSoftware built for command-based UI will become obsolete or demoted to backends as intent-based UI paradigm emerges, fundamentally shifting the user burden to the machine.24:19factMartin envisions intent-based AI where users express desired outcomes rather than step-by-step instructions.24:45factAs intent-based AI paradigm emerges, command-based software will not disappear but will be 'demoted down the stack' to become utilities and APIs used by intent-based applications rather than user-facing products.25:33observationRasmus observes that infrastructure layers (fiber, cloud compute, etc.) stack on top of each other, suggesting intent-based software will push command-based software down to become utilities rather than replacing it.26:38thesisUI commoditization will cause incumbent software companies to lose their competitive advantage, since AI will call any API regardless of brand to perform a function.27:45factMartin clarifies that Open Interpreter remains prompt-based and reactive, not autonomous, so AI takeover is not imminent.28:22observationMartin emphasizes that Open Interpreter is 'completely prompt-based and reactive'—not autonomous—pushing back against AI-takeover anxiety while grounding the discussion in current reality.28:22factMartin acknowledges theoretical possibility that an AI using Open Interpreter paired with a local LLM could theoretically create itself as a 'virus' replicating across cloud servers, but dismisses this as unlikely due to GPU cost constraints.30:07observationMartin raises the theoretical possibility of creating a self-replicating AI virus using Open Interpreter and local LLMs, demonstrating genuine dual-use capability.30:07factMartin believes the technology stack evolution will layer foundational AI models, AI-specific tooling, and intent-based UX on top of existing cloud infrastructure.31:41factAdept.ai is building a web-based tool similar to Open Interpreter but for browser automation, allowing AI to use any online tool, and has gone silent after recent funding, suggesting either major breakthrough or technical challenges.32:29factMartin is curious about adept.ai's web-based tool that would let AI use any online tool, speculating they may be building something revolutionary.32:29observationMartin speculates that adept.ai's sudden social media silence after major funding suggests either deep trouble or breakthrough so significant they don't need to market.32:29observationMartin draws a parallel between silent, profitable AI labs and opaque algorithmic trading firms that don't need investors because they're generating enough capital.33:42
Episode 22 — Echoes of Reality: When AI Voices Blur the Lines · 51observationMartin's current mental state is completely consumed by voice text-to-speech work, leaving no mental space for other topics.0:54factMartin is deeply focused on voice and text-to-speech technology as core to the future of AI interfaces.0:54factMartin at Multiply is working on new AI user interfaces as part of their vision for intent-based AI interaction.2:33thesisThinking about AI as intent-based interface rather than chat-first opens up broader possibilities for how humans can express abstract needs to autonomous systems.3:19factMartin argues that framing AI as intent-based UI, rather than chat-based, opens new possibilities for interface design.3:19factRasmus frames intent-based UI through three functional components: expressing intent clearly, the AI understanding and acting on that intent well, and the system expressing results back to the user.4:16factMartin believes intent-based interfaces enable abstraction, allowing users to express goals at multiple levels of abstraction rather than specific commands.5:09factMartin distinguishes current AI interaction from true intent-based paradigms by noting people still share ChatGPT prompts as commands, rather than expressing higher-level intent—such as 'help me acquire customers' versus 'write an email'.5:09factMartin explains that more autonomous AI agents enable abstraction of intent to higher conceptual levels, from commands like 'write an email' to goals like 'help me acquire customers' or even 'help me build a business'.5:35thesisAudio is a fundamentally more ambient and hands-free interface than text, making it natural for mobile and embodied AI contexts.7:13factMartin sees audio as an ambient interface that enables hands-free, natural interaction with AI while traveling or multitasking.7:13observationMartin experiences tangible friction when trying to chat with AI on his phone, a usability barrier he actively feels.7:38factAudio contains information beyond transcribed words—tone, pacing, emotion, and other nuances—which is lost in text-only interaction and affects how well intent is communicated.8:55thesisFor embodied AI, voice is the natural interface—text interaction with physical robots present in the same space violates human intuition.10:39factEmbodied AI (robots, dishwashers, Roombas) inherently requires audio/voice as the interface—no one would want to type commands to a robot in the same room, making voice the inevitable UI for physical AI presence.10:39factRasmus references MrBeast's experimentation with voice cloning, noting that while engagement metrics from voice-cloned content were only slightly worse than his own voice, he has not yet switched due to the preference for authenticity, though this opens opportunity for mass dubbing into multiple languages.11:55observationVoice cloning for dubbing creators into multiple languages could soon make creator content globally accessible without hiring voice actors, analogous to Hitchhiker's Guide babel fish.14:18factReal-time video dubbing enables recipients to watch content in their own language while the original speaker's mouth movements are adapted via AI to match the dubbed language, creating an immersive multilingual experience.15:00observationOpen source projects (Whisper, Tortoise TTS) have led the revolution in speech-to-text and text-to-speech, with commercial companies building billion-dollar businesses on top of them.15:42factMartin credits open-source software as the engine behind the revolution in speech-to-text and text-to-speech technology, citing Whisper (OpenAI) and Tortoise TTS as foundational.15:42factWhisper, released by OpenAI, was the foundational open-source breakthrough in speech-to-text that enabled the explosion of voice AI technology and was cloned into many versions by other projects.16:08factTortoise TTS library was the foundational open-source breakthrough in text-to-speech that enabled the explosion of voice AI technology and was forked by commercial services including ElevenLabs.16:08observationElevenLabs monetized open-source Tortoise TTS by improving it and building a superior developer experience.16:34factMartin identifies ElevenLabs as the market leader in text-to-speech technology, built on the open-source Tortoise TTS library with fine-tuning for voice cloning.16:34factDeepgram and Play.ht are competing commercial text-to-speech services in the growing market alongside ElevenLabs, with many new services launching rapidly.17:00factFaster Whisper is an open-source optimization of Whisper that improves speed for real-time speech-to-text applications, representing ongoing open-source competition on performance metrics.17:52factMartin points out that real-time voice AI requires audio streaming from both input and output services, with ElevenLabs and OpenAI providing streaming APIs for minimal latency.18:17observationStreaming audio at latency below ~1 second is the competitive frontier; every voice-AI team is racing to reduce lag.18:42thesisSpeech-to-text recognition is solved for AI understanding, but text-to-speech emotional expressiveness and conversational nuance remain the frontier requiring slow human-level refinement.20:02factMartin observes that speech-to-text technology is more mature than text-to-speech, with the latter still requiring refinement in emotional expression and human-like conversation.20:02factThe next frontier in text-to-speech technology is expressing emotional nuance and modulation—understanding which emotions to convey, when, and how to modulate that expression within conversation.20:02factGoogle Assistant employed a strategic approach of defining a specific voice personality and training the assistant to express all responses through that consistent digital personality.22:26observationMartin distinguishes between impressive one-take demos and production-ready systems; it's easy to fake the former but hard to ship the latter.23:43factMartin cautions that AI technology demos are easily polished through multiple takes, while production solutions must be robust and reliable—a significant difference that's often overlooked.23:43factMartin emphasizes that demos are easily polished through multiple attempts (recording 5 takes and publishing the best), masking the difference between production-ready robust solutions and impressive marketing demonstrations.23:43factRasmus notes that voice as a medium may better enable AI autonomy compared to text, as voice contains more cues for the AI to understand intent and respond proactively, even filling silence without explicit direction.24:21factChat-based interfaces, like the command-line terminal before graphical interfaces, have an inherently low upper adoption limit and may not become the mainstream killer interface for AI—voice is more likely to democratize AI to billions of users.25:05thesisVoice will become the mass-market UI for AI, far surpassing chat's reach by leveraging humans' existing comfort with voice-based phone interaction.26:06observationMartin emphasizes that text-based chat interfaces have fundamental limits in reaching mass audiences compared to voice, citing technical and behavioral barriers.26:33factMartin believes voice interfaces will democratize AI access by removing the technical barrier that chat-based interfaces present to mainstream users.26:33thesisStartups can compete in the AI space through pioneering focused technology OR through niche relationship-driven applications that mainstream players ignore.28:03factMartin identifies pioneering new technology as the primary opportunity for startups in the voice AI space, while noting that large players like Apple and Google will dominate mainstream interfaces.28:03factMartin sees niche markets like relationship-based AI companions as viable startup opportunities, since mainstream players like Apple will not serve non-mainstream use cases.29:00factAn influencer monetized voice cloning by creating an AI companion accessible via voice only (without personal data or chat history), earning substantial income by offering users a generic chatbot powered by her distinctive voice.29:00observationMajor tech incumbents (Google, Apple) refuse to chase 100M-person niches if they're small relative to their scale, leaving the market open to startups.29:26observationAn influencer monetized just a voice clone—no personal data, no custom model per user—simply by offering the experience of talking to someone's recognizable voice.30:51factReal-time deepfake videos represent an emerging business opportunity—cloned digital versions of people could sustain simulated relationships if trained on sufficient video content, blurring the line between representation and presence.31:09factLangChain is among the enabler products (alongside ElevenLabs) that provide infrastructure for startups and existing companies to build AI applications leveraging valuable data or interfaces.32:23thesisScience fiction has trained our expectation that AI should communicate via voice, not text, because voice interactions feel dramatic and cinematic while chat feels mundane.33:16factMartin observes that audio and voice are the natural interfaces to AI in science fiction depictions, reflecting deeper truths about human-AI interaction.33:16factScience fiction films like Her and Ex Machina consistently depict audio/voice as the natural interface to AI, shaping societal expectations of what human-AI interaction should look like.33:16
Episode 23 — Multimodality Unleashed: A Deep Dive into AI Integration · 37observationMartin makes a self-aware comment about vocabulary, noting that 'multimodal' is a learned technical term not yet in his active vocabulary in Swedish, showing his honesty about language boundaries in technical domains.0:58factRasmus describes ChatGPT's bedtime story demo as a vivid example of multimodal experience: the user asks for a story, ChatGPT generates narrative text, then the user requests pictures and the AI generates images within the same chat interface.1:41thesisCurrent multimodal AI demos like ChatGPT's image feature demonstrate tool use in a shared interface rather than true native understanding of multiple modalities.2:56factMartin is skeptical of ChatGPT's multimodal demo, viewing it as the AI using tools to generate images rather than achieving true multimodal understanding.2:56factMartin distinguishes between multimodality at the UI level versus deeper technical multimodal models that natively understand multiple modalities like text, images, and sound.3:22factMartin explains that Meta's audio translation AI model works natively on audio modality, not text, performing audio-to-audio translation with optional text instruction overlays.3:22factRasmus cites an example from an influencer/tester where ChatGPT correctly understood a whiteboard drawing of basic app architecture and generated working code based on the visual diagram, including interpreting symbolic elements like arrows between boxes and instructions.5:02observationMartin cites Microsoft's BLIP as a specific example of how systems can appear to understand images through sophisticated label-based reasoning, but he's skeptical this is true understanding.8:06observationMartin is curious about the theoretical foundations of how systems work, expressing eagerness to understand GPT-4's architecture 'under the hood,' while acknowledging this might not be publicly known.8:06thesisImage understanding systems that convert images to labels and then apply text reasoning (like Microsoft's BLIP) create an illusion of understanding rather than genuine comprehension.9:47factMartin criticizes Microsoft's BLIP image understanding model as appearing to achieve understanding rather than truly grasping images natively.9:47observationMartin frames AI's progress in image understanding through the lens of epistemology—distinguishing between giving the impression of understanding and achieving true native understanding at the semantic embedding level.10:38observationMartin expresses skepticism about whether current demos are as genuinely multimodal as they appear, suggesting OpenAI might be using DALL-E and other tools rather than native understanding.11:21factMartin critiques the concern that GPT-4 might achieve image understanding through the same label-based approach as Microsoft's BLIP (generating many labels and using reasoning) rather than native multimodal comprehension, noting this would be a disappointment but still possible.11:21observationMartin reflects on the epistemological challenge of describing images: an image contains so much information that full description requires about 1,000 words.12:16observationMartin uses linguistics and cognitive science—the difference between first and second language acquisition—as a mental model for understanding how AI models trained on different modalities develop distinct capabilities.13:06thesisTraining AI models natively on a specific modality creates fundamentally different capabilities, similar to how thinking in different languages creates different modes of thought.13:06factRasmus proposes that AI models trained on specific modalities (text, image, audio) might fundamentally think and reason in those modalities differently, similar to how bilingual speakers have different cognitive capabilities in each native vs second language.13:06observationMartin uses Meta's research on 3D spatial reasoning in AI as an example of training on unusual modalities, noting this might not be useful alone but could enable emergent properties when combined with other modalities.13:41factMartin notes that training AI in specialized modalities, such as 3D spatial reasoning, can create domain-specific expertise that might contribute emergent properties when combined with other modalities.13:41thesisCombining multiple modalities in AI creates emergent properties—new capabilities that arise from training on diverse modalities, not just from combining existing single-modality models.14:53factMartin reflects on emergent reasoning capabilities that arise from training multimodal AI models, comparing them to how GPT achieved reasoning from next-word prediction alone.14:53factRasmus frames DALL-E (text-to-image model) as performing translation between two distinct 'languages'—text and image—whereas text-to-text models like ChatGPT communicate within a single language.15:13observationMartin explicitly values software that is both theoretically sound and well-designed for users, citing Topaz Gigapixel as an example of excellent UX combined with technical rigor.17:29factMartin mentions Topaz Gigapixel as an existing commercial example of a real-world image upscaling model, priced around $100, with well-designed UX and technically sound implementation for photo editing workflows.17:29factMartin illustrates the power difference between single-modality (image-only upscaling) and multimodal models: adding text instructions to an image upscaling model enables nuanced commands like 'upscale from top left' or 'make teddy bear sharp but keep background blurry,' demonstrating multimodal reasoning capabilities.18:41thesisMultimodal capabilities can enable richer user interaction by allowing communication and instruction in multiple modalities simultaneously, not just sequential translation between modalities.19:06factMartin envisions multimodal AI that understands spoken instructions natively without requiring text translation as an intermediate step.19:23observationMartin expresses genuine enthusiasm about the practical implications of multimodal models enabling richer, more natural human-AI interaction (voice, text, image all natively).19:41factRasmus explores bidirectional multimodal communication: a truly multimodal AI could receive input in one modality (e.g., whiteboard image) and respond in a different modality (e.g., voice), or even iteratively clarify understanding by asking clarifying questions in one modality before proceeding with code generation.20:31observationMartin connects the discussion to real-world implications, asking where value will first emerge when multimodal AI is deployed at scale—showing his consistent focus on impact over capability.21:39factMartin highlights the importance of understanding societal implications and practical value when scaling multimodal AI deployment beyond proof-of-concept demonstrations.21:39thesisDiscovering practical applications for multimodal AI capabilities is a long-term exploratory process; current demos, while impressive, do not necessarily translate to world-changing value.22:07factMartin emphasizes that discovering practical, society-changing applications for multimodal AI capabilities is an open-ended long-term challenge, not yet solved by current demos.22:07factMartin expresses skepticism about the practical societal impact of current multimodal demos, noting that the bedtime story example, while entertaining, may not be genuinely world-changing or game-changing except for narrow use cases like very lonely children or parents lacking imagination.22:07observationMartin makes a self-deprecating comment about bedtime stories, suggesting they're only 'world-changing' for 'very lonely kids or parents with very little imagination'—showing his skepticism about trivial use cases.22:33factRasmus and Martin preview their intention to explore how multimodal AI capabilities will be rolled out in real-world contexts, with particular interest in how AI and AR/VR technologies will intersect, planned for discussion in the next episode.22:55
Episode 24 — Shaping Our Daily Lives with Multimodal AI · 42factMartin and Rasmus are co-founders/core members of Multiply.1:03factMeta is rolling out its Meta AI assistant across WhatsApp, Instagram, and Messenger.1:21factRay-Ban and Oakley are the same company with different brand names.2:35factMartin built Heywear, an eyewear business in New York in 2019.2:35factMartin created the Narrative Clip lifelogging camera.2:35factMartin has direct experience with wearable eyewear devices through both Heywear and Narrative Clip projects.2:35storyFrom Heywear to Meta: The Full Circle of Wearable Vision2:41factMartin sees Meta shipping multimodal AI glasses (with Ray-Ban) that can capture photos and video while providing AI assistance.3:01factMartin credits Google Glass as a failed attempt at AR that was too nerdy and scary in appearance.3:27factMeta's multimodal glasses capability will enable understanding of specific visual contexts like Thai menus, bike repair manuals, and identifying people's emotions.4:18factMartin envisions multimodal wearable AI glasses having positive applications for people with Alzheimer's disease.4:43thesisShared visual perspective between human and AI multiplies the effectiveness of co-creation and reduces friction in communication.7:32factMartin believes multimodal AI and wearable computing enable humans and AI to share perspective for better co-creation.7:32factMartin values that multimodal AI in wearable glasses means he doesn't have to explicitly explain context to the AI.7:58factMartin sees wearable computing and AR as pointing toward a direction where humans and AI can have shared perspective.7:58thesisMultimodal AI makes AR/wearable computing practically inevitable in the near future because the AI can now actually interpret visual information with precision.9:08factMartin wants to get Meta AI glasses to explore the technology firsthand.9:50factMeta Quest 3 has AR capability that projects the physical room into VR, allowing digital entities to be placed in actual physical space.10:18factMeta Quest 3 digital entities are AI-driven with speech generation and visualization in physical presence.10:43observationMartin suggests the form-factor strategy Meta uses: releasing Meta Ray-Ban as lightweight/low-fidelity and Meta Quest 3 as powerful/high-fidelity, with the goal to 'merge these' platforms into unified hardware over time.11:08factRay-Ban may be releasing blocky sunglasses to normalize the appearance of cameras in eyewear.11:35factMeta created a celebrity AI avatar of Snoop Dogg as a dungeon master for social metaverse experiences.12:36factMeta has a unique competitive advantage with its 3 billion-person social graph across multiple platforms.13:27observationMartin notes Meta's immense training data advantage: 'their models becoming foundationally good about connecting people' based on social graph of 3 billion across platforms with all mediums (images, text, video).14:18factMeta is partnering with Microsoft to integrate the Office suite into Meta Quest 3 for productivity applications.15:06factMeta demonstrated hyper-realistic avatars created via 3D scanning on Lex Fridman's podcast.16:00thesisSocial behavior norms and awkwardness are the actual bottleneck for VR/AR adoption, not technical limitations.18:09factMartin acknowledges social awkwardness as a major hurdle to VR/AR adoption, particularly in professional settings.18:09factMartin proposes VR booths as a solution for normalizing VR use in professional settings.19:08factMartin expresses concern that people wearing AR glasses might appear to be reviewing others' social media, creating trust issues in social interactions.19:58observationKids today consume AI-generated content with synthetic voices without caring about audio quality, which surprises and even bothers Martin.21:28factMartin observes that kids consume AI-generated content with synthetic voices without concern if the content is engaging.21:28factApple Vision Pro uses an external display screen to show a 3D model of the wearer's eyes, not transparent technology.21:53observationApple's Vision Pro decision to display a 3D model of your eyes on an external screen is framed as a 'very brave experiment' in managing social perception.23:01observationMartin and Rasmus note that wearing sunglasses during Swedish summer is already socially accepted, and ski goggles with reflective lenses are normal while conversing, suggesting precedent for opaque eyewear.23:21factSunglasses and ski goggles are culturally accepted eye-obscuring wearables that demonstrate potential acceptance of AR glasses.23:21observationMartin doesn't consider himself 'much of an Apple fanboy' despite most of his hardware being Apple, suggesting a pragmatic rather than tribal tech alignment.24:14factMartin is interested in trying Apple Vision Pro but uncertain about the shipping timeline.24:14observationMartin's term 'convergence' captures how multimodal AI doesn't just mean one interface but 'everything coming together in a convergent way.'24:50factMartin believes robots will be normalized and friendly in consumer adoption because scary robots never go mainstream.25:36observationMartin proposes that fear of AI robotics will self-regulate: 'Any scary robot will never go mainstream,' meaning the market will only accept cute, friendly robots.25:51factMartin believes robots are designed to be cute and non-threatening to ensure mainstream consumer adoption.25:51
Episode 25 — AI Workflows: The Key to Unlocking Business Value · 38factMartin is co-host of the Co-Creating with AI podcast alongside Rasmus Adler Wahlberg.0:00observationRasmus mentioned a massive storm with strong winds making his house shake at night, attributed to trees around his property.0:11observationMartin had been at his computer for 8am-midnight for 2 days straight after returning from Italy, barely going outside.0:35factMartin was recently in Italy for work, where he experienced warm weather and sea swimming.0:35factMartin describes Italy as having pleasant temperatures (26 degrees air, 21 degree sea) suitable for swimming.1:06factRasmus frames business AI adoption through three vectors: making people more productive, automating processes, and integrating AI into products.2:15observationThe phrase 'AI won't replace you, but someone that uses AI will' is described as a strong recurring meme floating everywhere that deeply resonates with people.2:44thesisIndividual adoption of AI tools drives organizational transformation, motivated equally by fear of replacement and hope of liberation from repetitive work.2:44factMartin proposes that the meme 'AI won't replace you, but someone using AI will' drives individual adoption of AI tools regardless of organizational strategy.2:44factMartin observes that ChatGPT is being used almost daily by non-technical people for both business and personal purposes.4:54storyThe marketing agency workflow complexity6:01factMulti-step workflows in no-code tools unlock business value that single-turn ChatGPT cannot provide, particularly for repetitive processes across multiple clients or steps.6:58factMake.com (formerly Integromat) is a competing no-code platform for building multi-step AI workflows alongside Multiply.9:10factPerplexity AI's copilot implements a single, repeatable multi-step research workflow: understand query, research web, present results, generate answer.12:26factHuman-in-the-loop workflows with continuous improvement cycles (Multiply, Perplexity model) create more sustained business value than fully autonomous agentic workflows.15:30thesisBusiness value from AI comes specifically from human-designed, repeatable workflows—not from agentic AI planning its own steps.15:55factGod Mode, where AI autonomously designs workflow steps, shows uncertain staying power compared to human-designed and continuously-improved workflows.15:55observationThe old Google Assistant hairdresser-calling demo predates LLMs and was later adapted as a ChatGPT+plugins showcase, illustrating a 10+ year stagnation in 'agentic' AI UX.17:47storyThe regulation chatbot breakthrough19:31observationMartin emphasizes that the key limiting factor for business AI implementation is not technology but knowledge and expertise—and that this barrier is far lower than people assume.22:33factMartin uses Cursor.so, a VS Code fork with integrated GPT-4, as his primary engineering tool for writing code.23:16factMartin appreciates Cursor.so's ability to index documentation for code completion, citing the paradigm of importing vast libraries without memorizing details.23:42observationCursor.so (a VS Code fork) implemented a powerful pattern: indexing documentation by URL, tagging it, and allowing @mentions to dynamically curate context for the AI during coding.24:07factMartin uses @mention syntax in Cursor.so to tag and curate context for AI assistance, enabling him to reference different documentation sources in co-creation.25:30observationPeople who adopt GitHub Copilot report they wouldn't want to live without it, making it a 'sticky' product.25:58thesisConsultancy firms will be the primary vehicle for mainstream AI adoption in organizations, solving the knowledge bottleneck that currently limits implementation.27:38factAI adoption is entering a 'catch-up effect' phase where organizational resistance decreases as consultancy firms begin delivering AI value to their clients.27:38observationThe future adoption wave will be less about new tool development and more about explosive growth in adoption of existing tools that are 'already good enough' to create real-world business value.28:03factMartin predicts that widespread adoption will come from existing, proven AI tools reaching market saturation rather than from further tool development.28:03factMartin seeks future AI development with increased planning and agentic capabilities while maintaining human oversight and control.28:30thesisGitHub Copilot succeeded as an AI tool because it made AI ambient—integrated into existing workflows without requiring the user to invoke it, providing unconscious productivity gains.29:00factMartin identifies GitHub Copilot as one of the first AI services successfully deployed at scale with significant user value.29:00factMartin emphasizes GitHub Copilot's ambient UX as a design achievement: users get high-quality suggestions while working normally, requiring only Tab to accept or continued typing to reject.29:25observationMartin describes GitHub Copilot's UX magic as unconscious productivity—you never call upon it, you just work in your normal editor and get a persistent 'magic' benefit that you then refuse to surrender.29:50factCursor.so offers dual AI modes: ambient GitHub Copilot suggestions (low-friction acceptance via Tab) and explicit GPT-4 invocation for intensive coding tasks.31:06thesisThe next wave of AI adoption will come through two pathways: making AI ambient within existing tools, or building no-code systems on top of existing platforms and datasets.31:27observationMartin distinguishes between tools with ambient AI (Copilot, Cursor) that don't require user invocation, and no-code workflow tools (Multiply, Make.com) that require active design but reward continuous improvement.32:18factMartin articulates that actual tool adoption by users is the most important measure of AI success, not tool capability or innovation alone.33:33
Episode 26 — Stability in AI: Harnessing Data Power · 32storyThe Content vs Context Discovery at Multiply1:49factMartin distinguishes between content and context when feeding data to AI models - content is material to generate new outputs from, while context is style or constraints to follow.3:06observationMartin confuses instruction-following by asking the AI to 'improve on these instructions' when he actually means to follow them, showing how ambiguous prompt semantics can flip the AI's behavior.4:47factRasmus explains that Multiply builds complex multi-layer AI workflows for clients that go beyond what ChatGPT offers by clearly specifying how data is used.6:09factRasmus explains the challenge of compressing large datasets - when summarizing brand voice guidelines for context windows, compression can destroy the stylistic patterns defining the brand.6:59observationMartin points out that when compressing large brand voice documents, naive summarization destroys the style, because style is embedded in word choice and phrasing, not propositional content.6:59thesisLLMs default to treating all input as generative content unless given explicit, granular instructions for each data source, because their core operation is next-word prediction.7:51thesisFine-tuning should be used to train style and instructions, never facts, because facts cannot be reliably stored via fine-tuning and require context-window feeding instead.9:00factMartin discusses MemGPT, a retrieval-augmented generation project from Berkeley that allows AI to actively pull data from memory.9:25factMartin explains that fine-tuning AI models for style involves training on sufficient data in that style, allowing the model to learn probabilistic patterns.10:56factMartin describes Perplexity AI's Copilot as an example of AI systems that construct their own context by allowing users to upload PDF and text files while also searching the web for relevant information.12:27factRasmus notes that open-source AI projects like AutoGPT and BabyAGI removed vector databases as their default approach, suggesting these systems can operate without vector database dependencies for smaller datasets.14:00observationRasmus identifies that most no-code AI tools now handle data uploading and retrieval, suggesting that the competitive advantage has shifted from 'can you use an LLM' to 'how do you architect data flow.'14:00factVector databases are most valuable when dealing with large amounts of unstructured text requiring semantic search; they are not necessary for smaller datasets that can fit in context windows or be easily scanned.14:30observationVector databases are overkill for small datasets but essential for large unstructured text, suggesting a scaling inflection point that depends on data volume and structure.14:30factMartin predicts that the next version of ChatGPT will include a vector database backend feature allowing users to store and retrieve documents, positioning this as a competitive edge that other AI services already have.16:15observationChatGPT has a paradoxical design flaw: it maintains full chat history but cannot retrieve from it without a feature that OpenAI could implement 'in a few months' with a 'flip of a switch.'16:40factMartin notes that expanding context windows increases computational costs exponentially because the transformer attention mechanism must compute correlations between every token pair.18:22observationThe context window in transformers requires squared computation (100K tokens wide × 100K tokens high), creating a steep computational ceiling that may not be solved by expanding context alone.18:22factMartin argues that hallucinations are not always undesirable - when generating creative content like domain names or poems, generative AI should produce novel outputs.18:47thesisHallucination is not inherently a flaw but a feature depending on task context—creative tasks require generative output, factual tasks require strict grounding, and the real challenge is designing systems that deliver the right type.18:47factRasmus articulates that AI systems must operate within an intent-based UI paradigm where humans provide clear intent through instructions, and this human input cannot be removed from the loop for the foreseeable future.20:32thesisThe practical frontier of AI is maximizing context window usage and robust prompt engineering, not building better foundation models—this is an engineering and product challenge, not a research one.22:10observationMartin suggests that 'transformers plus compute plus data' might be a sufficient path to AGI, implying that model scaling alone could achieve general intelligence without new architectural insights.22:36factMartin argues that when AI systems become mainstream and public-facing, resilient instruction-following and strong guardrails become essential to prevent malicious users from jailbreaking the system to extract business secrets or exploit vulnerabilities.23:27thesisWhen AI systems go mainstream with public access, the human input part of the context becomes a security vulnerability and requires strong guardrails, because users can attempt jailbreaks through carefully crafted prompts.23:27factMartin discusses jailbreak vulnerabilities in LLMs, including mechanical attacks where feeding repeated characters causes the model to hallucinate randomly.24:43observationA mechanical jailbreak on GPT-3.5 involving 1000 consecutive A's causes complete hallucination as if the model is 'reset into like just working from some primordial soup inside its brain.'24:43factRasmus emphasizes that confidential data kept outside of AI model training cannot be output by those models, making data security dependent on keeping sensitive information out of training data.26:21factRasmus notes that when providing data to AI models via API like OpenAI, they guarantee the data will not be used for training their models.27:15thesisData security with LLMs requires either keeping sensitive data out of models entirely or using APIs with explicit non-training guarantees, because any data in the model could theoretically be extracted.27:15observationMartin believes the real work for product teams is learning to squeeze maximum value from context windows via clever prompt engineering, not waiting for better models.29:10
Episode 27 — DIY AI: The Power of Self-Hosting · 35storyFrom API Constraints to Self-Hosted Models1:00factMartin rents his self-hosting GPU compute from the cloud rather than buying hardware, since buying expensive GPUs isn't worthwhile and renting by the hour gives him the same capability.2:17factMartin is building a conversational AI model that he can have a real-time conversation with, where he can give millisecond-level detail about what happens in the interaction.3:38factMartin experimented with cobbling together multiple commercial APIs (approximately 6-7 different AI models) for speech understanding, reasoning, data fetching, and speech synthesis, but found this approach became prohibitively expensive for continuous use.4:04factMartin clarifies that flexibility, not cost, was the actual driver for exploring self-hosted AI models: commercial APIs impose fixed limits on how granularly you can send and receive data, which he needed to control precisely.6:36thesisFlexibility, not just cost reduction, is the primary driver for self-hosting—APIs make architectural trade-offs you cannot control.6:36observationMartin frames his AI architecture goal as giving the system a 'millisecond by millisecond map' of conversation to achieve human-like flow.7:05factMartin describes himself as not an experienced full-stack developer, and says he spent weeks learning Docker and cloud AI-model deployment through a lot of frustration while building his self-hosted conversational AI stack.10:01observationMartin admits he's not a full-stack developer and has spent weeks learning Docker and cloud deployment with frustration.10:01factMartin considers GPT-4 to still be the best foundation reasoning LLM available, even as open-source models close the gap in other areas like speech-to-text.10:53factMartin notes that open-source speech-to-text services based on Whisper (developed by OpenAI) provide state-of-the-art performance, so self-hosting these models achieves quality comparable to commercial APIs.11:18observationWhisper, OpenAI's speech-to-text model, is already open-sourced, making the premium commercial version unnecessary for his use case.11:18factMartin says he refuses to settle for anything less than state-of-the-art intelligence for the reasoning step of his AI system, which as of this recording means using GPT-4 even within an otherwise self-hosted stack.14:03thesisLayered model stacking—using cheaper models for pre-processing and expensive state-of-the-art models only for final reasoning—optimizes both cost and speed.14:03factAt Multiply and Kindship, Martin uses optimization strategies to reduce costs and improve speed by using less capable models as pre-processing steps before running expensive GPT-4 models.14:29factMartin models conversational AI based on human cognition, using a tiered approach where a faster model handles real-time thinking similar to System 1 (like Llama 2) and a more powerful model for deeper reasoning.15:22factMartin's experience is that network transport latency (a few extra milliseconds) is negligible compared to model computation time (hundreds of milliseconds), but that latency can compound when a request passes through a pipeline of many chained models.17:41observationAdding latency at each hop in a complex pipeline compounds: milliseconds add up to significant total latency.18:21observationAt Multiply, the strategy is to compress information before sending to GPT-4, reducing both cost (expensive token pricing) and latency (slower model).19:21thesisFor truly fluid real-time conversation, latency optimization matters more than UI tricks—immediate response is non-negotiable.20:26factMartin argues that common UX tricks for masking AI latency (progress bars, loading animations) don't work for the flowy, real-time conversational AI he's building — only an actually immediate response will do, so shaving every millisecond off response time is a core design goal for him.20:40factMartin recommends starting with APIs to prototype and validate a business idea, then moving to self-hosting only when you encounter specific needs for flexibility, cost, or performance.21:53thesisStart with commercial APIs for speed and production-readiness, then only move to self-hosting when you hit specific limits like flexibility, cost, or data privacy needs.21:53observationMartin observes that even experienced full-stack developers face problems; expertise doesn't guarantee smooth deployment.22:46factMartin emphasizes that the key to choosing between APIs and self-hosting is understanding usage patterns: use APIs for infrequent calls, self-host for always-on solutions.23:31observationRunPod and similar serverless GPU providers offer a middle ground: pay only for seconds used, with containers that sleep between calls.23:31factMartin identifies serverless GPU rental (e.g. via RunPod) as a middle-ground hosting option between APIs and always-on self-hosting: you deploy your own Docker image with your own models, but it only spins up on demand and sleeps after 5 seconds of inactivity, so you pay per second used instead of by the hour.23:57factMartin frames the layered ecosystem of AI APIs, open-source models, and shared research (GitHub repos, published papers) as an astounding example of human collaboration, with everyone building their own 'puzzle' from shared pieces at whatever level of the stack suits them.26:56observationMartin reflects on human collaboration in AI as 'completely astounding and marvelous'—spanning APIs, open source, GitHub, and research papers.27:21observationSome customers switching from ChatGPT to Multiply specifically because they believe OpenAI trains on their data (despite contractual protections).27:51factRasmus notes that many Multiply customers coming from ChatGPT worry it will train on their data, even though OpenAI's terms of service with Multiply prohibit training on their data and require deletion within 30 days upon request — yet the desire for guaranteed data ownership remains a driver toward self-hosting.28:16factMartin highlights that self-hosting AI models enables full data ownership and privacy benefits that are impossible with APIs, allowing businesses to offer complete control to customers.28:58thesisSelf-hosting enables a business model impossible with APIs: offering fully owned, privacy-preserving software that customers can run inside their own walls.28:58factMartin argues that self-hosting AI models allows businesses and individuals to maintain full ownership of both software and data, a benefit that is impossible with API-based approaches.29:28observationRasmus observes this trend as potentially a 'return to installing software on your own computer'—a cyclical pattern in computing paradigms.29:57
Episode 28 — AI Breakthroughs: Turbo-charged by OpenAI · 30storyOpenAI Dev Day aftermath: the industry awakening0:00factMartin Källström is co-host of the Co-Creating with AI podcast, alongside Rasmus, discussing AI industry developments and applications.0:02observationRasmus stayed up late after Sam Altman's demo and couldn't stop talking about it—he's literally unable to discuss anything else.0:16factMartin characterizes OpenAI's November 7, 2023 developer announcements as an industry-defining moment that bifurcates responses: some entrepreneurs and developers view it as massively empowering (enabling far more capability), while others view it as an existential threat to business models built on earlier AI limitations.1:12factGPT-4 Turbo, which Multiply uses for development, extended the context window to 128K tokens (approximately 300 pages of text per prompt), reduced pricing by 67% compared to previous GPT-4, and improved processing speed, making it significantly more capable and cost-effective.2:28factMultiply uses GPT-4 as their state-of-the-art language model for product development, chosen for highest quality and reasoning capabilities.2:28thesisAI releases are a dividing line in the world: some businesses see unlimited opportunity, others face existential threat.5:26thesisOpenAI is strategically positioning itself as a platform company first, using ChatGPT as an internal learning and testing engine.5:26factMartin believes OpenAI's strategy of offering powerful APIs alongside their ChatGPT product is a bold move demonstrating confidence and driving industry innovation.6:16factMartin emphasizes the significance of OpenAI's decision to make new capabilities generally available from day one, showing confidence and driving industry-wide acceleration.6:42storyThe context window arms race and compression work thrown away7:20observationMultiply had just implemented Claude experimentally and pushed it to production with its 100K context window, which was already impressive compared to GPT-3.5.8:03factMultiply previously implemented Claude experimentally then in production for its 100K context window, but decided to switch to GPT-4 after OpenAI's announcements.8:03thesisEvery company should create and control its own public AI agent to represent itself, to prevent others from creating that AI for them.11:26factMartin views AI as the new workforce in companies and emphasizes that this vision aligns with the theme and mission of Co-Creating with AI.11:26factMultiply enables users to build AI assistants or builds assistants for users to perform workflows using OpenAI's new Assistants API capabilities.13:17thesisThe cost of AI intelligence is dropping so steeply (90% annually) that cost optimization is becoming irrelevant compared to just providing value.15:29factMartin envisions companies will adopt public-facing AI agents as a business standard, similar to maintaining a web presence via homepage, to maintain control over how their organization is represented and perceived by customers and the public.15:29factMartin argues that AI assistants built on structured, high-quality data stores deliver significantly more value than those relying on uncontrolled web retrieval, which often produces low-quality results, making data curation and structure critical for AI effectiveness.16:46factMartin cites Phind.com (a specialized code assistant built by fine-tuning language models for coding) as an exemplary case of how AI assistants can deliver exceptional value through deep optimization—the platform answers developer questions in 15 seconds with both documentation references and example code solutions.17:56observationPricing is so cheap now (1 cent per 1000 tokens) that filling a 128K context window costs about $1.28 per query—and yet it's still a fraction of competitors like ElevenLabs.21:04thesisAudio and voice should become the primary interface for AI in everyday work, not chat windows.23:42factMartin argues that text-to-speech capabilities are critical building blocks for developers to integrate speech interfaces into applications, aligning with AI's role as everyday workplace participants.23:42factMartin argues that speech and voice capabilities are critical for AI to function as a true workplace participant in corporate settings, not merely as a screen-based chatbot, because AI must be able to participate in meetings and real-time communication.24:07observationOpenAI's TTS is now faster than ElevenLabs (the market leader) in fast mode and higher quality in high-quality mode, all at 1/10 the cost.25:22observationRasmus is concerned that the lag in OpenAI's real-time voice interface might break the sense of conversation, but Martin has already tested it and says the animation and state feedback make the delay imperceptible.25:22observationWith audio and function calling, AI can now call people on the phone—which revives the old Google Assistant demo of calling a hairdresser to book an appointment.26:15observationOpenAI's API documentation explicitly requires disclosure to humans that they're talking to an AI—which might become a competitive disadvantage if other players (like Grok) ignore that requirement.29:35observationMidjourney probably still has a lead on image quality, but with DALL-E 3 now retaining style and being available in ChatGPT and the API, OpenAI is closing the gap fast.29:35factMultiply helps marketing professionals generate different types of content using AI capabilities including image style analysis and generation.31:15
Episode 29 — Embracing Change in AI: Strategies for Entrepreneurs · 35thesisDon't build the foundational layers of AI infrastructure; instead, focus on what will be uniquely valuable after commoditization occurs.1:25factMartin advocates that entrepreneurs cannot afford to focus only on current capabilities when building AI products, but must anticipate future developments.1:25observationCost and speed improvements in AI are as trustworthy as gravity—you can build business strategy around them happening inevitably.3:30factMartin believes entrepreneurs should trust that technological progress in AI (lower costs, higher speed, increased capacity) follows predictable patterns and will happen naturally.3:30factMartin argues entrepreneurs should not invest effort in optimizing for technological trends that will improve naturally; they should let progress happen without intervening.3:30thesisYou can turn time-to-market delays into strategic advantages by building features that won't be viable until the infrastructure improves.3:59factMartin notes that delayed time-to-market allows entrepreneurs to benefit from improved AI capabilities and costs by the time they launch.3:59thesisIn a heavily aligned AI industry, differentiation comes from doing something orthogonal that standard benchmarks don't measure.6:53factMartin identifies that the AI industry's high degree of alignment around standard evaluation metrics creates pockets of opportunity for differentiation.6:53factMultiply's team held an onsite specifically to map out which AI capabilities were likely to be handed to them by the industry versus what they should build themselves, as a deliberate startup planning exercise.7:39observationThere are 25 startups each building speech-to-text and text-to-speech, a sign of severe market saturation.10:06factMartin observes that the market for speech-to-text and text-to-speech systems is oversaturated with at least 25 competing startups in each category.10:31factMartin distinguishes between startups building thin wrappers on existing APIs (short-lived) and those pursuing original ideas with long-term vision.12:49observationMartin is willing to let others capture quick wins through thin-wrapper innovation even though he appreciates their value for the industry.13:40factMartin cites find.com as an example of a resilient AI product protected from commoditization through proprietary data assets: indexed documentation from over 200,000 open-source projects.14:05thesisUI/UX innovation is the most defensible and valuable layer of the AI stack because foundational model companies won't build consumer-grade interfaces.16:48factMartin identifies UI/UX design innovation as a high-value opportunity for creating differentiation in AI products.16:48factRasmus frames AI product opportunity along two axes: vertical niches (industry-specific use cases like legal or coding, serving niche users) versus horizontal opportunities (interface-layer innovation, like the chat paradigm, applicable across all users/industries).18:04factRasmus describes Multiply's product roadmap as a progression from single-player reactive chat, to multiplayer reusable 'workflows' (the company's current focus, used repeatedly by teams for complex tasks), to a proactive UI that anticipates needs without being asked.18:49thesisThe evolution from single-player reactive to multiplayer proactive is a fundamental UX frontier, and chat is insufficient for this transition.19:15factMartin proposes that AI-to-AI collaboration (multiple autonomous agents interacting with each other) remains almost entirely unexplored in commercial products.20:25observationMulti-agent systems inside sandboxes produce better results than single agents in some cases, but transparent agent-to-agent communication with external results is unexplored.20:51observationMartin studied agent frameworks and agent-to-agent communication in university courses from 1994-1999.21:40factMartin studied agent frameworks and multi-agent systems during his university years (1994-1999), where agents negotiated prices autonomously.21:40observationAgent-based automated bidding has been operating inside Google Ads for decades, making the concept far older than current AI hype.22:05factMartin grounds his idea of AI-to-AI negotiation by pointing to a real precedent: Google's ad engine has for decades let advertisers deploy an agent that autonomously negotiates the best ad prices on their behalf, an early production example of the agent-negotiation concept he studied at university.22:05observationCo-creation UI in the style of collaborative Google Docs (editing simultaneously, suggestions) beats iterative chat (copy-paste-revise cycles).23:14factRasmus explains that Multiply's customers appreciate it because, unlike a chat product where you prompt, get output, then copy-paste it elsewhere to edit, Multiply provides a real-time side-by-side editing UI with the AI, comparable to co-editing a Google Doc together.23:14observationOpenAI doesn't compete on UI or user experience; they build minimal wrappers because their core expertise is backend capability, not design.24:55factMartin argues OpenAI will never win design prizes for its UI because it is expert at building thin wrappers around its own APIs; its interface is deliberately rudimentary since the user value comes almost entirely from the backend model, not the front end.24:55observationThere is no design system for AI yet, unlike the mature design systems that existed by the end of Web 2.0 era.25:31factRasmus, discussing a lunch with a designer friend, observes that no established 'design system' for AI products exists yet, unlike Web 2.0, which had mature, ready-made design systems with well-understood UI patterns (dropdowns, etc.) — framing this gap as a major opportunity.25:31factMartin proposes an inverted paradigm where AI agents autonomously interact with the physical/digital world (like booking hotels or posting jobs) rather than waiting for human input.27:19thesisAn unexplored frontier is 'inside-out' AI that drives interaction with the external world rather than waiting inside a UI window for human commands.27:44storyAgent frameworks: A thirty-year cycle of rediscovery36:03
Episode 30 — AI and No-Code · 20factMartin is host of the Co-Creating with AI podcast, which premiered Jan 2023 on Multiply and explores AI co-creation, business strategy, and technology adoption.0:00factGuest Rasmus (product manager at Klarna, 8 years employed, age 34) has been exploring practical AI implementation across organizations and was initially skeptical of AI hype before ChatGPT's release in Nov 2022.1:33factFor most companies, implementing AI technology requires three buckets: (1) using it with people to increase individual contributor productivity, (2) using it in processes to remove recurring boring tasks, and (3) building products enriched by AI capabilities.4:56factMartin identifies using trained GPTs with Zapier integrations as a powerful pattern for building AI-native workflows across business processes without deep engineering.12:16factMartin and Host Rasmus discuss no-code tools (Make.com, Zapier) as accessible to non-engineers for building AI automation workflows without requiring deep technical expertise, exemplified by growing automation agencies.12:33factMartin believes GPT-4 has reasoning capability comparable to humans when structured with proper processes and accountability, challenging the binary intelligence assessment.14:29thesisGenerative AI capability is highly context-dependent and variable, performing comparably to humans in some situations but unreliably in others.14:29observationMartin explicitly admits to oscillating between viewing GPT-4 as capable and viewing it as unreliable, and frames this oscillation as acceptable rather than a failure to reach certainty.15:19factMartin argues that in a globalized AI-enabled market, companies do not necessarily need fewer people but can move faster and serve larger markets with greater speed.27:11thesisIn a globalized market with nearly infinite growth potential, companies should compete on speed and scale rather than assume AI-driven productivity necessitates workforce reduction.27:11factRasmus argues that when starting a software startup, primarily backend engineering capability is essential, as frontend work can be largely handled by no-code tools and AI-assisted methods, while sales remains critical for market validation.32:19factHost Rasmus argues that interpersonal skills and human-centric roles (sales, relationship-building) will become more valuable as automation handles routine work, since human attention becomes the limiting factor in an information-saturated market.35:32factMartin emphasizes that team co-location (putting people in the same room) is the most important factor in startup success, more critical than productivity gains from tools, because it enables accountability and a shared sense of truth about what is being built.38:24factMartin identifies accountability and a shared sense of truth about product viability as critical outcomes of team co-location, particularly the ability to maintain focus and validate iterative progress without external distractions.39:48observationMartin criticizes Swedish startup culture for being dominated by 3-year government-funded incubators that crowd out 6-month accelerators, creating misalignment with startup growth velocity.43:50observationMartin deconstructs the term 'incubator' as metaphorically problematic: it describes keeping something alive that cannot survive on its own, implying dependency rather than launch.44:35factMartin explores with guests Rasmus Adler Wahlberg (CEO of Multiply) the impact of no-code AI tools on startup formation, software development practices, and venture capital models.44:58factRasmus articulates a strategy for AI deployment that scales risk appetite with user base size: startups with 5 customers can take higher risks, while services with millions of customers require automated testing, code review processes, and gradual rollouts.45:20factRasmus cautions against over-optimization through excessive A/B testing in large products, arguing that cumulative micro-optimizations can create fragmented interfaces that undermine strategic product vision and long-term KPI growth.46:52observationAs co-host of a podcast entirely devoted to AI applications (60+ episodes), Martin positions himself and Multiply as embedded in contemporary AI thought leadership and experimentation.
Episode 31 — AI Unleashed: Navigating Price, Power, and Potential · 29factMartin is co-host of the 'Co-creating with AI' podcast with Rasmus Adler Wahlberg, co-founders of Multiply.0:02factMartin has been experimenting with Google's Yamnet model for sound event detection, which identifies sounds like dogs barking, stomach rumbling, and airplane engines in audio streams.0:41factMartin has shifted from calling cloud APIs that charge per call to running AI models directly on local GPU infrastructure, which costs nothing per call.1:06observationMartin describes a recent personal shift in how he works with AI—moving from paying per API call to running inference on his own GPU for essentially zero cost, fundamentally changing his relationship to experimentation.1:06factMartin recommends giving AI-generated text individual word or character budgets per section and iterating if targets aren't met, which produces less repetitive output with greater structural variety.2:22thesisAI cost reduction and falling intelligence prices enable proliferation of niche, valuable vertical software businesses.6:53factMistral's latest models achieve performance parity with GPT-3.5 Turbo at 40% lower cost, demonstrating rapid price compression in the AI market.7:52observationWithin weeks of ChatGPT 3.5 Turbo's release, Mistral has built a comparable or superior model at 40% lower cost, suggesting rapid compression in AI capability pricing.7:52observationMartin and Rasmus note that even as self-described 'AI nerds,' their schedules are so packed testing new models that they haven't even tried Google's Gemini Ultra.9:26factMartin has a friend who is exploring metamodernism and using it to build AI assistants in the context of Swedish government procurement reform.10:13observationMartin references metamodernism (or 'integer theory' in the US) as a framework for understanding AI uncertainty—allowing oneself to oscillate between seeing AI as both danger and opportunity without resolving the contradiction.10:38factMartin knows of Swedish government initiatives exploring AI for procurement, including using synthetic data to avoid uploading sensitive governmental data to external APIs.11:29factMartin's government contacts use synthetic data generation to create AI proof-of-concepts while protecting sensitive governmental data from external API services.12:21observationMartin describes a friend working in Swedish government who uses synthetic data—fabricated reports with real structure—to develop proof-of-concepts with AI APIs without exposing actual governmental data.12:21factMartin sees vertical software opportunities in government, where AI can help process unstructured data like documents, forms, regulations, and meeting protocols to automate processes.12:52thesisGovernment is an underexploited vertical software opportunity, especially for AI applications handling unstructured data.12:52factMartin has identified Prosperous Planet, a Swedish company using AI and climate-conscious purchasing, as a successful vertical software use case.15:44storyProsperous Planet's climate intelligence chatbot15:44observationRasmus notes that companies are increasingly building AI capabilities in-house (e.g., a Swedish furniture retailer integrating AI across inventory, website, and e-commerce) rather than waiting for packaged SaaS solutions.19:44observationRasmus speculates that the app ecosystem may evolve toward 'trillions of apps that are very small'—a granular version of today's one-billion-app App Store era.21:47factMartin articulates a vision where AI will consume APIs and human interfaces directly, making AI-friendly endpoints a competitive factor for companies.22:30observationSemantic web standards (structured data for machines) have invisibly transformed e-commerce and Google's search results, yet the mechanism remains unknown to most consumers.22:54factMartin identifies a potential startup opportunity in creating data discovery standards for AI tools, analogous to semantic web markup that became standard e-commerce infrastructure.23:45observationMartin points out that APIs designed for humans (e.g., hotel booking) are becoming a limiting factor for AI; machines are having to reverse-engineer human interfaces because AI-native endpoints aren't widely available.25:17factMartin speculates that AI systems could increasingly use human-designed web and mobile interfaces as their primary interaction layer rather than relying on AI-specific APIs.25:49thesisAI-discoverable APIs and endpoints will become a competitive necessity for companies, paralleling the earlier semantic web impact on e-commerce.25:49factMartin and Rasmus discussed how search will likely transform from a link-finding mechanism to a transaction and action engine driven by AI assistants.27:42thesisFuture business model of search and commerce will shift from clickthrough discovery to direct AI-enabled action, fundamentally changing how Google and platforms operate.27:42factMartin works at Multiply and can be contacted at martin@multiply.co.30:42
Episode 32 — Navigating AI: 2023 Reflections & 2024 Predictions · 43factMartin co-hosts the Co-Creating with AI podcast with Rasmus Adler Wahlberg, exploring AI topics and predictions.0:01observationMartin found his heating system failure during Swedish winter forcing him to rely on a fireplace, creating a cozy work environment.0:27factMartin was in Sweden during the recording, experiencing -10 Celsius weather with a broken heating system.0:27thesisResilience to continuous technological change is now the foremost business quality in the AI era.1:53factMartin's biggest takeaway from 2023 was the pace of industry change and the need for resilience toward constant adaptation, noting that stability is hard to find in the AI space.1:53factMartin credits GPT-4 and OpenAI's vision capabilities as the biggest game changers of 2023.2:40factMartin notes that open source AI models (particularly from Meta and Mistral) are keeping pace with commercial developments, setting a culture of releasing LLM weights.2:40observationMeta's decision to open-source LLM weights created a cultural shift that empowered the entire open-source AI ecosystem.3:06factMartin characterizes Mistral as the 'French connection in the AI world,' emphasizing their role in stepping up the open source game after Meta's precedent of open-sourcing LLM weights.3:32factRasmus articulates the framework that all companies will become AI companies in 2024 the same way they all became email and cloud companies, representing a fundamental shift from adoption contemplation to implementation.6:32thesisMultimodal AI (vision and voice) will be the major mainstream breakthrough once it reaches beyond premium subscribers.9:09factMartin predicts multimodal AI capabilities will become mainstream in 2024, but currently remains limited to premium OpenAI subscribers.9:09observationMultimodal capabilities (vision + voice) being restricted to OpenAI premium subscribers acts as an awareness blocker for mainstream adoption.9:35factMartin notes that multimodal capabilities—text, image, and voice combined—are mostly unknown to mainstream users but produce strong reactions when demonstrated.10:01thesisAI laggards will shift from denial to fear, requiring companies to communicate AI's human value more effectively.10:52factMartin predicts that AI laggards who currently dismiss AI as just parroting training data will shift from denial to fear as AI capabilities prove themselves in 2024.10:52factRasmus identifies 2023 as the year of well-scripted AI demos, while predicting 2024 will see these capabilities integrated into deep, solid end-to-end products across multiple modalities.13:35factRasmus argues that fear around AI will decrease in 2024 as consumers experience practical benefits, countering Martin's prediction of increased fear. He believes AI doomsday concerns are limited to intellectual circles, not mainstream users.15:47factRasmus sees Tesla as the clearest uniquely positioned company in real-world AI, citing their neural network approach for autonomous vehicles and Optimus humanoid robot development in parallel form factors.17:24observationTesla is uniquely positioned in real-world AI deployment, having replaced a 300,000-line rules-based system with a single neural network approach for autonomous driving.17:50factBoston Dynamics is noted by Rasmus as another company pursuing real-world AI, though their approach differs from Tesla's integrated autonomous vehicle and humanoid robot strategy.17:50factRasmus analyzes Microsoft's potential to add a $36/month charge for Copilot features to enterprise customers as a significant but uncertain revenue opportunity dependent on broad adoption.19:29observationMartin invested in Nvidia and Broadcom as AI infrastructure plays during the GPT-3 era, viewing chipmakers as the primary beneficiaries.20:55factMartin invested in NVIDIA and Broadcom during the GPT-3 era, betting that chipmakers would be the primary beneficiaries of AI advancement.20:55observationBroadcom, a lesser-known semiconductor company, also benefited significantly from the AI boom despite being primarily a maker of auxiliary communication and ARM chips.21:26thesisSpecialized AI processors could eventually challenge Nvidia's GPU dominance if they achieve performance gains and hardware integration.21:51factMartin is tracking specialized processor development for transformers, noting potential 50x performance improvement over generic GPUs.21:51factMartin views the Hugging Face transformer library as mature and stable due to heavy open source investment, making it suitable for hardware implementation.22:20factMartin expects deep integration of AI into actual business products in 2024, moving beyond simple prototypes and demos.24:14observationGoogle Gemini's demo video inspired approximately 20 teams to immediately build competing products using existing technologies.25:49factMartin predicts that 20+ teams are now implementing the fluid, multimodal interaction capabilities demonstrated in Google's Gemini demo video, showing how existing technologies can be integrated today.25:49observationMeta's AR glasses (Oculus/Ray-Bans) could become transformative if AI integration enables social features like facial recognition with contact recall.26:59factMartin anticipates Meta's AR glasses could become significant when integrated with AI capabilities, citing a demo by Zuckerberg.26:59factMartin believes that AI will continue improving in speed, cost, and latency throughout 2024 as a foundational trend.27:52thesisGPT-5 will not arrive in 2024; competitors have approximately one more year to close the gap to GPT-4.28:07factMartin predicts GPT-5 will arrive in 2025, not 2024, giving competitors another year to attempt beating GPT-4.28:07factMartin doubts open source LLMs can match GPT-4 at scale, citing the massive compute requirements required.28:33factMartin notes the strategic disadvantage open source LLMs face in scale: Mistral and similar startups will struggle to reach the billions of users accessible to Google and Microsoft through their platform moats.29:00factRasmus predicts an invisible race where Apple, Google, and Amazon will integrate advanced AI into existing consumer products (Siri, Bard, Alexa) reaching billions of users overnight.29:37observationThe distribution mechanism for AI matters enormously: native OS integration (Siri, Bard, Alexa) vs. standalone ChatGPT Plus represents a vast difference in accessibility.30:34observationThe race for AI-powered phone assistants is happening invisibly, with multiple large tech companies competing without obvious public announcements.30:49observationApple released both the Ferret model and an Open Transformer library for M2 processors, enabling developers to run performant transformers on Mac devices.31:06factApple has released the Ferret model and an Open Transformer library optimized for M2 processors, enabling developers to run performant transformers on Mac hardware natively.31:06
Episode 33 — AI as the New Interface: Exploring the GPT Store Revolution · 32factMartin is Chief Product Officer of Multiply and co-hosts the Co-creating with AI podcast alongside Rasmus.0:03observationMartin is working from home with a cat at his desk, heating problems with their wood-fired stove as backup.1:32observationMartin is using BART, an older language model not designed for generation, specifically to measure and validate linguistic properties like punctuation and language identification.1:57factMartin experiments with older language models like BART for linguistic evaluation tasks rather than generation, using it to measure sentence structure and punctuation.1:57factOpenAI released GPT Team, a feature enabling organizations to create shared workspaces for building and using custom GPTs accessible only to team members, alongside the GPT Store launch.3:39factMartin believes ChatGPT could function as an operating system for organizations by integrating enterprise APIs and custom GPTs, allowing employees to interface organizational data through AI.5:41thesisChatGPT with integrated custom tools can become an operating system for organizations if those tools are built to be safely exposed to the AI.6:07factMartin observes that ChatGPT could become an operating system if built out with tools that are safe to expose and allow AI to access and make sense of organizational data alongside employees.6:07factMartin notes that the ChatGPT plugin store, launched a year prior, did not become a significant channel for customer acquisition or engagement despite initial expectations that it would be the new app store.14:00thesisCustom GPTs succeed where ChatGPT plugins failed because they package tools together with prompts specifically adapted to how the assistant should use them.14:51factMartin explains that custom GPTs are more effective than plugins because they allow builders to adapt prompts and instructions so the assistant knows how to call tools and what inputs they expect.14:51factMartin describes combining Retrieval-Augmented Generation (RAG) with tools in custom GPTs, creating a powerful yet simple interface for both creators and users.16:24factMartin proposes that AI-powered backend architectures could allow new businesses to focus initially on backend engineers only, deferring frontend development and UI design.17:15thesisOpenAI is intentionally releasing products incrementally rather than perfecting them first, responding to competitive pressure from the rest of the industry.18:34factOpenAI is deferring GPT Store revenue sharing until Q1 2024 and initially only for US creators, citing legal and payment complexity for global distribution.18:34factOpenAI is releasing GPT Store features iteratively and imperfectly rather than waiting for completion, driven by competitive pressure from other AI companies in the industry.18:34observationOpenAI apparently leaked or A/B-tested a ChatGPT continuous memory feature, announcing it to users before the feature was live, then removing it.19:55factChatGPT's memory feature appeared in leaked screenshots or A/B testing, potentially allowing continuous memory across conversations and temporary/incognito chats, though its implementation details remain unclear.19:55factMartin speculates that if the GPT Store succeeds, ChatGPT would evolve into an operating system allowing users to book flights, manage email, access company documents, and have ChatGPT more integrated into their daily lives.22:10factMartin envisions ChatGPT with memory capabilities, allowing it to be more knowledgeable about users' personal lives and follow up on previous conversations, such as recurring topics about family.22:36observationMartin's historical reference to Web 2.0 mashups (15 years prior to this 2024 episode) as a parallel to GPT combinations suggests he's tracking long-term tech cycles and recognizing repeating patterns.24:25thesisBuilding GPTs as mashups of existing services—horizontally integrating multiple tools—is a viable business model similar to Web 2.0 startups.24:25factRasmus interprets OpenAI's strategy as building specialist intelligences that serve as building blocks toward artificial general intelligence, with ChatGPT as a hub coordinating specialized agents for different tasks.25:26factMartin predicts that within a year, ChatGPT could access other specialized assistants as subordinate agents, delegating specific tasks to them on behalf of users.26:43factMartin emphasizes that custom GPTs represent thin layers on top of ChatGPT, combining RAG, data analytics, and simple configuration through prompts and API keys.27:16observationMartin describes the pressure on entrepreneurs from the GPT Store as 'a really good challenge'—not a death knell for startups, but a necessary competitive forcing function.27:48thesisEntrepreneurs building AI apps must genuinely innovate beyond thin API wrappers; relying on shallow integration with OpenAI's APIs is not a sustainable competitive strategy.27:48factMartin challenges entrepreneurs and startups building AI apps to innovate beyond thin layers on top of OpenAI APIs, arguing that lazy implementations risk becoming obsolete as custom GPTs become more capable.27:48factRasmus predicts that AI capabilities will converge on a few dominant access points (ChatGPT, Gemini, Siri) through which most AI functionality reaches users, with other apps supplementing these interfaces or providing specialized human UX.28:12factMartin reflects that toolmakers have been slow to adopt AI into their products, despite early speculation about a land rush following GPT-3 and GPT-4 launches, with OpenAI now building tools directly rather than waiting for third-party integration.29:35observationMartin acknowledges disappointment that traditional toolmakers have not innovated quickly enough to integrate AI, but frames OpenAI's move as a natural competitive response rather than a threat.30:01thesisOpenAI is reversing the expected tech ecosystem flow: instead of waiting for toolmakers to build AI into their products, OpenAI is building the tools directly into its platform.30:01
Episode 34 — Designing the AI-Driven Future · 14factMartin is a co-host of the Co-creating with AI podcast, appearing with co-host Rasmus.0:01factMartin knew Tomas Måsviken (guest Samsen founder) from way back when Martin was running his first company.1:00factMartin identifies new AI interfaces that leverage AI capabilities as a major opportunity for entrepreneurs.4:21factMartin illustrates his concept of new AI interfaces through the example of Google's Android feature that lets users draw a circle on any part of a photo to search or ask questions about it (e.g., asking about a board game). He views this as a good human-like interface model.4:47observationMartin formulated his platform consolidation thesis a year ago, when GPT-4 was new, suggesting he thinks in longer cycles than the typical tech hype cycle.8:44thesisMajor technology companies are building AI into their platform layers and operating systems rather than every individual company adding AI into their own products.9:10factMartin observes that operating systems like Android and iOS could quickly shift to AI-centric interfaces instead of app-centric design. The OS could show an AI interface on the first screen instead of app symbols, leveraging existing app ecosystem capabilities.16:00observationMartin demonstrates explicit concern about power dynamics and control in AI systems, specifically whether service providers can maintain meaningful brand presence when large platforms own the interface layer.18:13thesisService providers and smaller companies will need to maintain brand exposure and visibility in order to prevent larger technology platforms from completely controlling the customer interface.18:13observationMartin deliberately pivots away from speculation to ground the conversation in what is actually being built and deployed right now.18:58factMartin believes current AI interaction is primarily reactive (chat and voice) rather than proactive.25:08factMartin explores the question of what needs to change when shifting AI interfaces from reactive to proactive.25:59factMartin observes that his smartphone already demonstrates contextually-aware interface design: when pulling down from the top screen to search for an app, it shows 8 app icons contextually estimated for his current situation. This illustrates what proactive AI interfaces could look like.28:39factMartin reflects that for every thousand really cool AI demos, there is only one real-world product with actual AI features.30:08
Episode 35 — Exploring AI: Strategy, Emotion, and Innovation · 36factMartin co-hosts the Co-creating with AI podcast with Rasmus Adler Wahlberg and invites guests like Johan Salo to discuss AI, design, and strategy.0:00factMartin is fasting while recording the podcast episode to achieve a smooth blood glucose response curve.0:16observationAll three participants discover mid-recording, one by one, that they are independently fasting on the same day.0:52observationMartin wears a continuous glucose monitor despite having no diabetes diagnosis, recommended by a former Multiply AI colleague.0:55factMartin is experimenting with continuous blood glucose monitoring using CBionics for personal health tracking, recommended by Žygis from Multiply.0:55factMartin believes learning from data visualization about one's own body is more valuable than reading books; he applies this principle through continuous blood glucose monitoring.0:55storyThe brainwave fingerprint that shouldn't have worked2:37factMartin is working on Multiply with co-founder Rasmus Adler Wahlberg; the platform supports sharing prompts and AI workflows.9:44thesisThe best interface is one you never notice - good design disappears rather than calling attention to itself.15:29factMartin's philosophy of good UI design is that interfaces should disappear and never draw attention to themselves during use.15:29factMartin drives a Tesla and appreciates its approach to interface design where the UI disappears to not draw unnecessary attention.16:37storyBuilding a voice clone that roasts you17:04observationJohan realizes mid-conversation that his homemade voice-clone-plus-vision prototype is functionally identical to the commercial Rabbit R1 device.18:28factMartin believes that wearable cameras provide unique value by capturing everyday memories and adding forgotten details through AI analysis.19:04factMartin's understanding of human memory limits: people are not naturally good at remembering; AI assistants augment memory by adding forgotten details from photos taken just days or hours earlier, creating unexplored value in lifelogging.19:04thesisBecause human memory is unreliable and quietly fills in false details, everyday photo-logging of one's own life holds AI-assistant value that remains almost entirely unexplored.19:55observationThe Narrative Clip camera, discontinued for years, still has a functioning backend service that can be reactivated to feed photos into an AI assistant.20:10factMartin created the Narrative Clip wearable lifelogging camera; the service continues to operate and users can restore old cameras.20:10observationThe Narrative Clip was deliberately designed so it could not be switched off in software - the only way to stop it recording was to physically remove it.22:08factMartin designed the Narrative Clip so it could not be turned off electronically; users had to physically remove the camera to stop recording, forcing transparency about data collection.22:08thesisGlasses are an excellent form factor for wearable tech once the hardware is light enough, because people already resent carrying extra objects.22:34factMartin is bullish on AR glasses as a form factor for wearable cameras once technology becomes lightweight enough for everyday wear.22:34factMartin thinks about product form factor trade-offs: adding another device is a nuisance because it becomes another thing to potentially lose, which conflicts with his philosophy of minimizing carried items.22:34factMartin reduces personal items to track by consolidating: his wallet is now attached to his phone, and he no longer carries keys, leaving only one device to keep track of.22:34observationMartin no longer carries physical keys at all - his wallet has merged into his phone and he's down to tracking a single object.22:59thesisA camera built into glasses creates an unresolvable social conflict: people uncomfortable with being filmed can't bring themselves to ask someone to remove their glasses, so the discomfort goes unspoken and unresolved.25:35factMartin's analysis of Google Glass failure: the device put people in an uncomfortable position because asking someone to remove their glasses is socially unacceptable, but people's privacy needs could not be expressed, leaving them feeling deeply uncomfortable.25:35factMartin is concerned about privacy and consent friction inherent in wearable glasses: people cannot comfortably ask someone to remove their glasses like they can ask them to remove a hat, creating an awkward social situation.26:01factMartin was involved with Heywear, a venture founded after Narrative, in the wearable glasses/eyewear space.26:46observationIn a 2003 grad-school interview, Johan pitched an AR zombie-chase game played through glasses and was told by the professor it was just 'a boy dream' - and now considers it close to buildable.32:16factMartin recommends Hume.ai, an API for emotion detection in voice and vision, as a tool for analyzing emotional content in research.34:40observationMartin explains that the AI emotion-detection company Hume is named after philosopher David Hume, whose thesis was that reason is a slave to the passions.34:49observationMartin immediately counterbalances the emotion-AI enthusiasm by citing a book arguing facial expressions can't reliably reveal emotion at all.35:30factMartin recommends the neuroscience book How Emotions Are Made by Lisa Barrett to balance AI emotion detection claims.35:30observationWhile fasting, Martin's glucose monitor shows an almost perfectly flat line, which he uses to jokingly signal it's time for a break.37:42factMartin has learned that a ketogenic diet keeps blood glucose curves smooth and stable, unlike diets with high-glycemic foods which create sharp spikes.38:20
Episode 36 — Multilingual Embeddings and Semantic Exploration · 27factMartin is actively exploring multilingual embeddings using the Sentence Transformers library to make text from different languages adjacent to each other in vector space.0:17factMartin emphasizes that embeddings are extremely powerful because they work very fast, enabling rapid data processing and analysis.0:42observationMultilingual embeddings mean documentation only needs to exist in one language—users can ask questions in Swedish about English documentation and get relevant results.1:20factMartin notes that pre-computed embeddings of Wikipedia are available for download as a vector database, allowing developers to retrieve relevant Wikipedia articles for any user query without computing embeddings themselves.2:27observationYou can query the entire embedded Wikipedia and retrieve source URLs without embedding Wikipedia yourself—just download the pre-computed vectors.2:27factMartin is exploring semantic routing, a technique that uses embeddings to intelligently route user queries to the most appropriate AI models or data sources rather than relying on a single large model.4:48factMartin recognizes that Perplexity and Metafor are mature APIs providing vector search over the entire web, which can be integrated into his own AI services as an alternative to having GPT-4 handle web search queries.5:40factMartin is impressed by the power of embeddings to capture semantic meaning in mathematical form, noting they can be used for arithmetic operations like adding and subtracting semantic concepts.13:44thesisEmbeddings are powerful because they're extremely fast and enable semantic arithmetic—you can add and subtract meaning mathematically to achieve effects like tone modification.13:44factMartin demonstrates knowledge of how Notion uses embeddings and vector arithmetic to transform text style, such as applying a professional tone by adding embedding vectors to the original text and regenerating from the sum.14:32factMartin observes that Midjourney and DALL-E use embeddings of words to generate images, allowing users to add or subtract semantic concepts like 'angry' or 'fur' from image generation prompts.16:22factMartin explains that GPT-4 Vision works with embeddings of images directly in its latent space rather than converting images to text first, which affects how prompting influences the output.17:58thesisEmbeddings are fundamental to how modern AI works—they're not just for retrieval but central to multimodal understanding in vision models, routing systems, and semantic processing.18:49observationMartin's skepticism about Notion's vector arithmetic approach: the results of adding professional-tone embeddings are 'much more fuzzy and not as sharp' than simply prompting GPT-4 directly.20:27factMartin sees the richness and power of embeddings demonstrated by Notion's ability to take an embedding of a paragraph and regenerate the text nearly verbatim with all stylistic and factual elements preserved.21:27observationAn embedding can compress a paragraph—including its quotes, tone, and factual content—into 3,000 floating-point numbers and reconstruct it almost perfectly.21:54thesisThere is vastly untapped potential in embeddings beyond simple retrieval—the ability to reverse-engineer semantic information from embeddings suggests deeper applications are possible.21:54observationBrain-reading technology using EEG signals can reconstruct visual imagery, appearing as vague but recognizable shapes on screen—he finds this genuinely unsettling.24:21factMartin views embeddings as a fundamental bridge between different modalities and domains, including the relationship between human brain signals and AI systems.25:13observationEmbeddings represent 'our best understanding for how to translate between different realms'—between human brains and computers, between different modalities.25:13factMartin expresses skepticism about Sam Altman's focus on personalizing AI through data access, believing instead that effort should focus on improving AI reasoning capabilities.28:26thesisAI development should prioritize improving reasoning capacity, not just expanding data access through APIs—reasoning is the constraint that matters most.28:26factMartin prefers the vision of AI that changes dynamically in response to the world rather than remaining static within frozen model weights, suggesting AI should have more subjective experience.28:51thesisFor truly personalized AI systems, embeddings and semantic routing will be central—user data will flow into vector databases organized by semantic meaning, enabling context-aware responses without retraining.29:46factMartin speculates that OpenAI may be pursuing personalization strategies rather than simply training GPT-5 because the company has accumulated too many engineers and developers, requiring diverse projects to keep them occupied.31:40observationMartin speculates that OpenAI might be running out of training data or capital, forcing them to staff 1,000 engineers to build features rather than improve the core model.31:40observationAlibaba's new LLM is now rivaling or surpassing GPT-3.5 in capacity, along with France entering the AI race—the competition is becoming truly global.32:37
Episode 37 — AI Hits the Super Bowl: Mainstream or Hype? · 34factSuper Bowl 2024 featured prominent AI advertising including Microsoft Copilot, Google Pixel 8 smart features, and mainstream brands using AI to generate recipes.1:16thesisDespite Super Bowl advertisements and high media visibility for AI, widespread practical adoption and visible societal change are much slower than the hype suggests.2:42factCrypto dominated Super Bowl ads in 2022 but subsequently shrank from mainstream attention, unlike AI which is expected to have broader impact.3:07factMartin believes AI will have much faster and more widespread impacts on people's daily lives, with people of all ages becoming daily AI consumers.3:59observationMartin observes that the 2022 Super Bowl ads (peak crypto hype) were likely ordered and decided in 2021, whereas the 2024 Super Bowl had AI ads at a time of record AI company valuations (NVIDIA).4:22observationIn a coffee shop, Martin overheard people discussing whether AI should have a persona if it participates in work meetings.6:10thesisAI brings existential and philosophical questions into mainstream discourse in a way earlier technologies like crypto did not.6:41factMartin references Max, a Swedish AI philosopher, who advocated approximately a year ago (circa 2023) for placing a brake on AI development and collected a petition with prominent scientists; Martin suggests this critical movement actually helped promote more constructive public discussions about AI's positive uses rather than only doomsday narratives.9:34observationThe shift in AI narrative away from apocalypse to productivity: people now realize very few jobs will be immediately deleted by AI, but productivity will increase.10:32factMartin articulates the principle that technology is typically overestimated in short-term impact and underestimated in long-term impact.13:01thesisAI is a technology we overestimate in the short term but underestimate in the long term—a pattern true of transformative technologies generally.13:01storyThe grandmother who refused the telephone14:47factMartin illustrates technology trust concerns through a personal anecdote: a friend's Icelandic grandmother 50 years ago refused to use a newly installed phone because she feared voice imitation technology could be used to fabricate statements and deceive her, a concern Martin notes is now becoming real with deepfake technology.14:59observationRasmus frames the current moment as a shift in the Gartner hype cycle: initial exaggerated hopes and fears are becoming more realistic, leading to actual value creation and real problems.16:20factA Swedish ambassador to Germany was targeted with a deepfake video around 2022, produced by Russian actors, to diminish her credibility during public debate.18:01observationA Swedish ambassador to Germany fell victim to an early deepfake video attack around 2 years prior (circa 2022), created by what appeared to be a Russian troll factory, designed to undermine her credibility in a heated public debate.18:01factA finance worker was recently tricked by deepfake Zoom technology into authorizing a $25 million payment from a multinational company's CFO.20:23observationA finance worker was tricked by a deepfake Zoom call featuring multiple colleagues and a CFO, resulting in a $25 million fraudulent payment just a week before this episode.20:23observationMartin notes that individual skepticism and awareness are likely the first line of defense against deepfakes, but this creates transaction costs and friction in normal business operations.21:37thesisTrust in digital communications and transactions will require new social and technical protocols, combining better authentication (2FA, digital signatures) with organizational and societal changes.22:19storyThe spoof email attack on Narrative22:45factDuring Martin's tenure at Narrative (2014-2015), the company experienced a sophisticated email spoofing attack where perpetrators mapped organizational structure through phone calls.23:10observationMartin mentions Leonardo DiCaprio's brand potentially being eroded over time by deepfakes and fake AI celebrities.24:28thesisOrganizations and societies need to be built with resilience and antifragility to withstand AI-enabled fraud and deepfake attacks.25:35factMartin proposes a technical approach to prevent organizational fraud: integrating delivery records, payment authorizations, and cryptographic authentication so that major payments can only be processed after the system verifies a supplier has actually delivered value matching the invoiced amount.27:40factMartin proposes that corporate AI systems with comprehensive knowledge of company operations, financials, and stakeholders could serve as an additional validation layer for large financial transactions, requiring AI approval alongside human authorization before payments are released.28:42factJens Nylander's municipal analysis found that Swedish municipalities paid 300 million SEK over 30 years to a non-existing newspaper through systematic fraud.30:13observationJens Nylander discovered that Swedish municipalities paid 300 million SEK over 30 years for invoices from a newspaper that did not exist.30:13factJens Nylander uses AI to investigate Swedish municipal finances, discovering systematic fraud including services invoiced through unregistered companies to avoid tax.30:39observationMartin notes that organizations often accept invoices from different company numbers without understanding the pattern—a widespread scheme to hide income from taxes.30:39factMartin suggests AI could systematically analyze public government meeting protocols to detect fraud, corruption, and nepotism across all governance levels.31:37thesisAI can be deployed for transparency and accountability in government and public institutions, uncovering fraud and incompetence at scale.31:37factJens Nylander is a serial entrepreneur who founded and sold Jens of Sweden (MP3 player) and Automile for one billion SEK.32:50observationJens Nylander, who created the Jens of Sweden MP3 player and sold Automile for a billion SEK, is now using AI for public accountability as a form of guerrilla PR for his analysis service.32:50
Episode 38 — Unlocking AI's Practical Power with Oscar Beijbom · 29factMartin co-hosts the Co-Creating with AI podcast alongside Rasmus Adler Wahlberg, and they publish it through Multiply.0:00factMartin plans to visit Santa Cruz, California in approximately 6 weeks to meet Oskar Beijbom.0:54observationOskar was in his 40s and self-described as 'jaded' when he applied to Y Combinator as a founder, yet was uncertain whether to accept admission—a striking contrast to younger founders who treat YC acceptance as life-defining.1:32observationNickel maintains 98% retention despite struggling with customer acquisition and marketing—the product sticks for those who find it, but discovery is the constraint.2:23observationOskar was the first employee at Høvding and created the core algorithm, giving him two decades of applied machine learning experience—he's not a theoretical AI researcher.12:36observationThe open-world learning problem is exponentially hard: 100 types of pedestrians × 50 weather conditions × light variation = a combinatorial grid where every single intersection must be trained on, yet edge cases are infinite.14:39factMartin proposes that self-driving cars should learn from observing traffic accidents by slowing down and monitoring the scene, mimicking human learning from experience.18:40factMartin is skeptical that current AI hype will sustain, predicting the field will experience volatility before settling into normal development pace.19:29factMartin believes current AI technology is fundamentally limited to generating plausible-sounding text ('wordsmiths') without true intelligence or understanding.19:55thesisReliable AI applications require strong human processes where LLMs are building blocks, not autonomous systems; the critical value comes from the 'glue' and architecture, not the models alone.21:02factMartin advocates that open-world AI problems require very strong processes layered on top of LLMs to achieve even specific use cases reliably, rather than relying on models alone.21:02factMartin believes that reliable AI technology comes not just from the models themselves, but from the processes and orchestration (the 'glue') around them.21:28factAt Multiply, Martin's team implements human oversight of AI-generated creative output as the ultimate guardrail against AI hallucinations and errors.25:48observationFour-wheel drive cars have 20–30% more collisions in winter than two-wheel drive cars because drivers feel safer and rely more on the technology rather than driving more carefully.28:05factMartin draws an analogy to 4-wheel drive cars being involved in more wintertime collisions than 2-wheel drive cars (20-30% more) due to driver complacency, as a warning about AI guardrails.28:05thesisAI should be viewed as a coworker or assistant with growing capability, not as an autonomous agent or replacement; the user must maintain skepticism and review its work rather than delegate fully.28:43factMartin conceptualizes AI as a 'coworker' rather than an assistant or tool—a collaborative entity that improves in competency but retains limitations and requires ongoing partnership.28:43thesisAI systems need permission to be slow; current industry incentives favor speed over quality, but slower, compute-rich processes can produce more robust results.32:28factMartin advocates for processes that allow AI systems more time and compute budget to reason and think thoroughly, rather than optimizing for speed.32:28factMartin believes that coding with AI agents requires creating processes where AI spends most time thinking deeply (80%) rather than just writing code quickly (20%), allowing for deeper reasoning and more robust output.32:28factMartin highlights Nyckel's approach of training multiple models and selecting the best performer as an exemplary pattern for building reliable AI applications.33:44factMartin has extensive hands-on experience working with speech-to-text models and has found them deeply unreliable, prone to hallucination, and non-deterministic.34:18factMartin works with voice and speech-to-text systems as part of his professional work at Multiply.34:18factMartin uses Whisper speech-to-text and has found that newer versions (V3) are more prone to hallucination than earlier versions (V2), leading him to recommend using V2 or fine-tuning it.35:08observationMartin tried Google's real-time speech-to-text API for Swedish and it produced gibberish—text that 'looks like a cross between Norwegian and Dutch'—despite being documented as supporting Swedish.36:16factMartin has directly tested Google's real-time speech-to-text API and found it produces Swedish text that is unintelligible—mixing Swedish, Norwegian, and Dutch—despite being documented as supporting Swedish.36:16observationOskar interprets GPT-4's adoption of mixture-of-experts architecture as a sign that OpenAI has exhausted innovation and is now just throwing compute at marginal gains—a warning flag, not an achievement.38:14observationMartin offhandedly mentions he has long hair and Oskar jokes about surfing; Martin replies there's 'no correlation, unfortunately' but says he might learn while visiting Santa Cruz.41:49observationOskar describes himself as feeling 'like an old man, skeptical of all this new stuff' about the AI hype, and goes back and forth on his convictions.42:22
Episode 39 — The Voice and the Robot Revolution · 44factMartin is working on voice AI full-time at Kindship.0:25factMartin observed the Figure OpenAI robot demonstration and was impressed by its real-time problem-solving and humanoid form.0:25storyThe Figure Robot Demo1:46observationResearch on instruction-to-action translation (LLM plans → physical actions) has only been happening for about one year and involves work across multiple companies (Figure, NVIDIA).5:30observationThe Figure robot demo may contain subtle cheating: it casually drops the apple expecting the human to catch it mid-air, which Martin identifies as a risky, brittle behavior that doesn't reflect robust robotics engineering.6:46factMartin suspects potential 'cheating' in the Figure OpenAI robot demo regarding how it handled object dropping, suggesting pre-programmed constraints on gripper behavior rather than pure learned dexterity.6:46factA new research area focuses on translating between textual intentions output by LLMs and physical actions in the real world, including robotics and web automation.7:12factNVIDIA is researching instruction-to-action translation through frameworks that allow LLMs to execute actions in virtual environments like Minecraft.7:37factLeWeb is an open-source project enabling LLMs to execute web browsing instructions like clicking, typing, and form submission.8:03thesisWorld models—internal representations of how the physical world works—are essential for AI to act successfully in the real world.10:38factThe Figure OpenAI robot may rely on GPT-4's embedded world model for planning and understanding the physical environment.10:38factWorld models are built into multiple AI modalities: LLMs implicitly model the world through language, image generators like Midjourney and video generators like Sora contain world models understanding physics and spatial relationships.11:04factMartin references speculation that OpenAI may be directly editing GPT-4's weights to remove certain behaviors, colloquially termed 'AI lobotomy,' possibly explaining perceived quality deterioration over time.13:16factMartin considers Lex Fridman's recent podcast interview with Yann LeCun (Meta's AI lead) a valuable learning resource for understanding world models and how AI can execute plans in the physical world.13:42factMartin references VAPI and Retell as Y Combinator-backed startups providing voice API services for building voice capabilities into apps.15:58storyMartin's Voice AI Bet on Hard Mode16:18observationMartin has been working on Kindship's voice AI for 'the past 6 months or so,' building capabilities that commercial services (VAPI, Retell) launched only recently.16:48thesisThe multi-speaker, multi-language voice problem is more valuable and strategically important than the single-speaker voice niche being tackled by VC-backed startups.16:48factMartin has been developing voice capabilities at Kindship for approximately 6 months.16:48factMartin deliberately chose to tackle voice AI development with increased complexity from the start, taking it 'on hard mode.'16:48observationY Combinator is hedging its bets on voice AI by funding both VAPI and Retell—two startups solving nearly identical problems with slight nuances.17:13factMartin is building voice capabilities that handle multiple speakers and multiple languages simultaneously, unlike VAPI and Retell which focus on single-speaker interactions.18:29factMartin is designing Kindship's AI to handle multiple languages dynamically, including Swedish, Icelandic, and English in the same conversation.18:29factApplication contexts for multi-speaker voice AI include Zoom-style video calls and integration with humanoid robots like Tesla's Optimus.19:33factInterruptibility is a key voice AI feature allowing users to interrupt the AI mid-response to steer the conversation, improving conversational flow.19:55factMartin views the challenge of understanding when the AI is being directly addressed versus when it is an observer as central to multi-speaker voice AI.20:22factVAPI and Retell lack emotional understanding features; they rely entirely on the underlying LLM to handle emotional content, treating voice as a pure I/O layer without emotional recognition.22:33observationRetell and VAPI focus on narrow use cases (haircut booking, car sales, customer service) where the AI should ignore emotional cues and stick to facts, while Kindship pursues emotional understanding and participation.23:24factPi (by Inflection AI) positions itself as an emotional, personal AI companion, contrasting with VAPI/Retell's transactional, task-focused approach.24:11factMartin intends for Kindship's AI to dynamically understand and adapt its role in conversations, ranging from silent note-taker to active participant.25:24storyThe Floor-Holding Discovery26:41observationGroq has 'revolutionized throughput' in LLM inference, delivering a page of text in about 1 second, enabling faster voice AI responses.27:15factGroq has enabled high-throughput LLM inference, allowing full-page text generation in approximately one second, addressing latency constraints in voice AI.27:15factVoice AI systems must employ response strategies like canned opening statements ('Um's) to cover computational latency while formulating complete responses.28:02thesisNatural voice AI requires replicating human conversational techniques like floor-holding and turn-taking, not just linguistic understanding.28:42factMartin discusses how humans use floor-holding techniques like 'um' and 'uh' sounds to maintain conversational control while thinking, a technique relevant to voice AI design.28:42factFloor-giving and floor-taking are conscious conversational techniques humans use to allocate speaking turns and control in dialogue, beyond floor-holding.29:08observationRasmus speculates that chat might not remain a major interface in a few years, a prediction Martin partly agrees with but hedges—he thinks voice will become primary but chat won't disappear entirely.29:51factMartin believes text-based chat interfaces may not remain a primary human-computer interface in the future, especially as voice and embodied interaction improve.29:51observationMartin frames the future transition from text to voice as moving from micromanagement (typing letters to machines) to being authentically human.30:54thesisVoice interaction should replace text as the primary interface because speaking is more natural and human than typing letters to machines.30:54factMartin believes voice interaction will increasingly replace keyboard-based text input in future human-computer interfaces.30:54factMartin views typing as 'micromanagement' compared to natural speech, treating keyboard input as a restrictive constraint on human-computer interaction.30:54observationMartin is planning to build a future demo of the Kindship bot participating in a live podcast episode, with technical challenges around whether/how the bot should be embodied in the recording.31:40
Episode 40 — Transforming Creative Processes · 35factMartin recently visited Silicon Valley and the Bay Area to reconnect with the AI startup ecosystem, including time in Big Sur and Napa.0:18storyReconnecting with Silicon Valley0:18observationThe Bay Area has reached saturation density in AI events: a dozen per week is the baseline, with multiple hackathons happening simultaneously in different modalities.1:38factMartin has stepped back from day-to-day involvement with Multiply to focus on his new venture Kindship.2:20observationMartin is currently 'out of the loop with Multiply these days when I work for Totally on Kindship,' suggesting he has stepped back from his previous executive role to pursue new ventures.2:20factMultiply found its clear value proposition in helping communication agencies streamline their entire creative process, not just content production.2:45factMultiply underwent a strategic pivot around late 2023, expanding from focusing primarily on content production to encompassing the entire creative process from client brief through strategy and creative concepts.4:53storyThe pivot from content production to holistic creative process5:02factMagnus, an experienced business director from major agencies, advised Multiply to expand its offering holistically across the entire creative process rather than focusing only on content production.6:08observationMagnus, Multiply's business director hire from the big agencies, was instrumental in reframing the entire company strategy from content production to holistic creative workflows.6:08factMultiply's core offering remains unchanged: integrating different AI models into a single interface for team collaboration and workflow processes.7:50observationCreative directors and art directors require fundamentally different UI than business directors and strategy leads; the same underlying technology doesn't scale across user personas without deliberate design.11:28thesisCo-creative AI, where humans and AI iterate together, is fundamentally different from generative AI, where a user submits a prompt and takes whatever output they get.12:15factMultiply enables users to directly edit AI-generated suggestions in-line, distinguishing it from tools like ChatGPT where output cannot be edited without copying elsewhere.13:13observationUnlike ChatGPT, where you cannot edit AI-generated text without asking for a complete rewrite or copying to another tool, Multiply allows direct inline editing of AI suggestions.13:13observationCustomer adoption of Multiply happens through structured pilots where Multiply's AI Business Developers spend weeks customizing the platform to mirror the customer's existing process before handing over ownership.14:37thesisEvery communication agency has a fundamentally different process, even though they appear to do similar work, so enterprise AI tools must be customizable rather than one-size-fits-all.14:37thesisDemonstrating that a tool can solve a customer's own problem is far more convincing than showing pre-packaged capabilities.16:23storyThe 5-minute workflow revelation17:23factMultiply is planning to develop a creative canvas feature to better serve art directors and creative directors, allowing visual mood board-style workflows similar to Midjourney.19:18observationIn Multiply's current design, the workflow is document-and-table based, but creatives are accustomed to canvas-based tools (mood boards, Midjourney galleries) and the product roadmap now includes a canvas interface.19:54factMultiply is integrating Perplexity API to enable AI-driven research capabilities within workflows, allowing AI to source external information relevant to client briefs.21:38factMultiply achieved 100% uptime over a 90-day period and has substantially improved its technical infrastructure since its early stages.25:34observationMultiply achieved 100% uptime over a 90-day period in the past year, which is notable given the company's technical complexity and youth.25:34factMultiply is pursuing ISO certification to address enterprise customer concerns about data security and AI-related compliance.25:49factEnterprise customers have significant concerns about AI-related data privacy and security, particularly the risk of company data escaping into the wild through AI model usage, which has become a major sales requirement for Multiply.26:27thesisSecurity and data privacy have become a primary competitive advantage and sales driver for enterprise AI tools, not just a feature to check off.26:27factMultiply's primary target customer segment is larger communication agencies with at least 50 employees, particularly those in global networks.28:07factMultiply targets CMO offices at mid-sized to large companies, helping them manage high volumes of research data and streamline creative processes.28:32factE-commerce companies represent an emerging use case for Multiply, with a Swedish e-commerce company launching 200-300 products annually requiring templated content creation for web and social media.29:35observationA Swedish e-commerce company launching 200–300 products per year now represents a proven use case for Multiply, using the platform to auto-generate product-specific content at scale.29:35factMartin and Rasmus articulate Multiply's core philosophy: humans and AI working together in co-creation can achieve far more than either alone.31:31thesisThe story that unites creators around AI—that 'we can create more together'—alleviates fear and makes adoption exciting, and building community around this narrative is core to market strategy.31:31factMultiply is building a community platform at creatives.ai focused on learning and fostering belief in AI-augmented creative collaboration.32:41observationMultiply is building a community at creatives.ai (still in invite-only beta) centered on the belief that 'we can create much more together' with AI, framing it as a learning community.32:41
Episode 41 — Beyond Grad School Intelligence · 44factMartin co-hosts the Co-creating with AI podcast with Rasmus Wahlberg, discussing AI co-creation and advancements.0:01factMartin believes current best AI models are at grad school level competence, but emerging models will reach PhD-level intelligence across multiple disciplines.2:21factMartin cites Jason Huang (NVIDIA CEO) claiming that hardware and software improvements will enable a million times faster inference and training speeds.2:52factMartin uses Fireworks for AI inference optimization, focusing on time-to-first-token performance for real-time applications.4:35factMartin notes that Groq is fastest for per-token or tokens-per-second metrics, while Fireworks excels at time-to-first-token speed.4:35observationThe distinction between Groq (fastest per-token throughput) and Fireworks (fastest time-to-first-token) suggests specialization, not convergence, in AI infrastructure.4:35factMartin explains his use of AI classification with streaming inference to react to first token output within 200 milliseconds.5:01observationFireworks.ai optimizes for time-to-first-token rather than throughput, enabling 200-millisecond classification responses that allow real-time reaction before the full response is generated.5:01factMartin believes Jason Huang's claims about AI speed improvements are partly motivated by NVIDIA's competitive position against Groq.6:19observationNVIDIA CEO Jensen Huang's claim of 'a million times faster' inference and training is likely a stock-market-driven response to Groq's current speed advantage.6:19observationSam Altman's framework for evaluating startup defensibility: are founders excited or scared of the next version of the model? This inverts traditional startup thinking.10:48factMartin is a heavy user of Cursor, an AI-driven fork of VS Code for development.13:21observationEven with sophisticated AI tools like GitHub Copilot and Cursor, software development still fundamentally relies on classical software patterns and wrapper layers around LLMs.13:21factMartin discusses how smarter LLMs will reduce the need for wrapper code around language models, making classic software wrappers thinner.13:47thesisAs AI models become more intelligent, the software scaffolding required to support them will become thinner and thinner.13:47factMartin believes that as LLMs become smarter, agent frameworks may become obsolete rather than remaining central to AI systems.14:14factMartin emphasizes that AI engineers should focus on actual problems and real-world impact rather than exploring technology for its own sake.14:59thesisAI engineers must return to being task-driven and focused on real-world problems rather than exploring technology for its own sake.15:24factMartin believes guardrails and sandboxes are necessary security measures to prevent misuse of LLMs even in future systems with greater autonomy and access to external systems.17:59factMartin believes agent frameworks are unreliable due to error compounding in multi-step workflows with multiple LLMs.21:29thesisMulti-agent frameworks with sequential steps are fragile because errors compound through the workflow.21:29factMartin cites Devin AI as an example of misleading demo videos; the company cherry-picked examples and apparently faked results.22:20storyThe Devin Deception22:22factMartin notes that the failed Devin AI demo subsequently inspired multiple open source projects like OpenDevin and Deplin, as well as startups riding the hype wave Devin momentarily achieved.23:20storyGoogle's Too-Good-to-Be-True Gemini Launch23:43factMartin notes that Google's Gemini 1.0 demo video was faked by a marketing agency to show idealized AI capabilities.24:04observationGoogle's Gemini demo being both 'fake' and valuable as 'science fiction' reflects Martin's pragmatic acceptance that speculative demos can be inspirational without being deceptive.24:04factMartin agrees that AI demo videos like Figure AI's robot demo (where it drops an apple), even if currently pre-programmed or faked, represent realistic future capabilities once AI becomes sufficiently intelligent.24:26storyFigure AI's Apple Drop24:51factMartin discusses the core value proposition of Multiply: enabling teams and AI to work together in shared workspaces with structured processes.24:59factMartin explains that currently AI needs explicit instructions and context because it operates like a smart junior coworker, making things up without proper constraints.26:21thesisStructured workflows and processes become exponentially more valuable as models get smarter, not less valuable.27:38factMartin believes smarter models with structured processes will have exponentially increasing value due to their synergy.28:04factMartin believes process design will remain valuable even as AI models become more capable, as it represents the optimal path to desired outcomes.29:00factMartin believes collaborative interfaces will become increasingly important as AI models get smarter, enabling human-AI teamwork.30:31factMartin discusses the importance of shared data access between humans and AI in collaborative systems.31:22factMartin expresses uncertainty about whether chat interfaces will be the dominant way to interact with advanced AI, predicting other forms may emerge.31:48observationMartin explicitly rejects chat as the dominant interface for human-AI collaboration, despite acknowledging it will be 'one thing' among options.31:48thesisInterface design and collaborative architecture will be the limiting factor for AI adoption, not model capability.31:48factMartin believes Multiply's dynamic knowledge graph will remain valuable for enabling smarter LLMs to structure knowledge without being limited by database schemas.32:24observationMartin's moment of doubt: 'Are we just a prompt engineering framework?' followed by reassurance through understanding Multiply's structural advantages.33:19factMartin values Multiply's multiplayer collaborative aspect as a key differentiator and future-proofing strategy for the product.33:41factMartin advocates for first principles thinking in AI product development rather than just building the latest technologies.34:42thesisThe principle 'AI will eat software' means that future value accrues to those building organizational structure and interfaces, not to those building classical software layers.34:42
Episode 42 — Revolutionizing AI: OpenAI's Omni Model · 31observationMartin's new morning routine is to drive to the forest to run 5km, admitting it 'feels wrong' to drive somewhere to exercise, but that removing the friction of running through the city first is what actually makes the habit stick.0:39factMartin has adopted a new morning routine of driving to the forest to run 5 kilometers, finding this approach more sustainable than urban running which requires overcoming the hurdle of running through the city.0:39thesisGPT-4o is a genuinely new omnimodal foundation model, not a fine-tune of GPT-4 — native audio tokens (alongside text and image) enable true audio-to-audio interaction at ~300ms latency, a capability that never existed before.2:45factMartin identifies OpenAI's Omni as a fully new foundation model with audio tokens alongside text and image tokens, enabling direct audio input/output without intermediate text conversion.2:45observationGPT-4o can learn to make whale sounds from just a few audio examples given directly in the prompt.3:11factMartin explains that OpenAI's Omni model supports few-shot audio prompting: showing the model an example sound (such as whale sounds) lets it reproduce that sound style directly, without any text intermediary.3:11factMartin highlights that Omni enables direct audio-to-audio processing with 300 milliseconds latency, representing an unprecedented capability that has never existed before at this speed.5:04thesisGPT-4o's small size, low serving cost, and benchmark dominance suggest it is an early checkpoint of a much larger frontier model (GPT-4.5 or GPT-5), trained with dramatically more compute on the same data rather than more data.6:46factMartin relays and endorses an analysis that GPT-4o, despite being smaller and cheaper than GPT-4, was trained with far more compute on higher-quality data, and speculates it may be an early checkpoint of an upcoming GPT-4.5 or GPT-5.6:46observationBefore the official launch, OpenAI covertly tested GPT-4o on the LMSYS leaderboard under a disguised name, which Martin recalls as 'a little GPT-2 bot.'7:27observationRasmus jokes that real-time AI translation in your ear finally makes the Hitchhiker's Guide 'babel fish' concept practically reasonable, and expects it to land in AirPods soon.8:56observationRasmus mentions hearing that Apple struck a deal with OpenAI, likely to power Siri with GPT-4o, but flags it as unverified.9:22storyRelief from Solo-Building Voice AI9:40factMartin has been working solo on voice development for AI agents over the past few months, facing numerous technical challenges that OpenAI's Omni model helps solve, allowing him to shift focus to higher-level concerns.9:40factMartin describes facing a myriad of tiny technical challenges in voice development and having to build from scratch using open-source components, which OpenAI's Omni model now eliminates, opening up possibilities.10:05observationDigging into OpenAI's own website, Martin found a credits list showing roughly 200 people had worked on GPT-4o's voice capabilities alone.10:30factWhile researching GPT-4o, Martin found a contributor list buried on OpenAI's website showing roughly 200 people across different roles had worked on the model's voice capabilities.10:30factMartin frames OpenAI's GPT-4o release as having built the 'System 1' (fast, intuitive) layer of AI, which is the premise behind his stated plan to now focus on System 2 reasoning for autonomous agents.10:57factMartin cites a Sam Altman tweet framing OpenAI's own mission as building foundational AI tools, leaving it to the rest of the world to use those tools to actually change the world.11:09thesisGPT-4o effectively solves the 'System 1' problem for AI — fast, low-latency, natural voice interaction — which frees builders of autonomous agents to stop worrying about interface mechanics and focus on 'System 2': how agents should reason and act autonomously.11:34factMartin explains that having solved voice mechanics through OpenAI's tools, he can now focus on System 2 thinking for AI agents: how they should reason, think, and act autonomously in the world.11:34factMartin argues conversational AI needs a persistent internal state of recent interaction, since repeatedly needing a 'how are you today' greeting ritual when reactivated is a sign the AI would otherwise be blank and would ruin the experience.13:22thesisBy making GPT-4o's voice interface free and giving GPTs the ability to call any API, OpenAI has turned itself into a universal voice front-end platform — any company can now use a GPT as the voice interface to their product instead of building one.13:55factMartin articulates a vision where companies can use OpenAI's GPTs as voice interfaces to their own APIs, creating a universal best-in-class voice interface platform accessible to everyone.14:20thesisOpenAI, under Altman, deliberately engineered the GPT-4o launch's timing and tone as a strategic strike against Google — quietly building what leaked from Google's overhyped Gemini demo, releasing it the day before Google I/O, and publicly talking down GPT-4 for weeks beforehand to set up an underhype-then-overdeliver contrast.16:28observationSam Altman had reportedly told media for a month before launch that GPT-4 is 'the stupidest model you'll ever use,' priming the underhype-then-overdeliver reveal of GPT-4o.17:51observationIn the launch demo, GPT-4o was reading a bedtime story and, when asked twice in a row to be 'more dramatic,' delivered an escalating dramatic performance each time.18:31factMartin speculates that few-shot voice prompting could enable voice cloning, where providing a one-minute audio sample and prompting the AI to match that voice could achieve personalized voice synthesis.19:42factMartin observes that with GPT-4o's release, OpenAI abandoned the pretense of AI not seeming human, making it the most anthropomorphized AI to date and effectively ending the era of an AI insisting it has no feelings.21:28observationOpenAI's launch-day videos showed employees like Greg Brockman playing with the new voice mode using two phones talking to each other.22:37observationGPT-4o can reportedly generate images with correctly spelled, consistent text and even 3D models, capabilities documented on openai.com but never shown in the announcement video.23:07
Episode 43 — The Future of AI Collaboration · 49factMartin attended an AI meetup with a diverse mix of enterprise professionals, techies, academics, artists, and academics, with more than 50% of attendees being women—unusual diversity for a tech event.0:31observationAI conferences now have 50%+ women attendees, breaking from tech's gender imbalance.0:56observationUnusual diversity at the AI meetup: artists, political science professors, truckers learning via podcasts—not just startup founders and engineers.1:37observationPodcasts have democratized education in ways web articles never did, because people absorb them passively (truckers driving, commuters).1:55observationGoogle's strategy against OpenAI involves coordinating announcements—they timed their assistant release to make OpenAI's Sky voice look less novel by comparison.3:08observationBill Gurley (Benchmark founder) tried OpenAI's new voice feature and said it was 'not as good as in the demos.'3:31factSam Altman believes AI assistants should be clearly identified as such rather than pretending to be the user, to prevent a future where humans cannot distinguish whether they are interacting with the user or their AI agent.4:31thesisAI agents should never impersonate humans; this should be regulated.5:39factRasmus Adler Wahlberg expects government regulation will emerge to prohibit AI systems from impersonating humans, particularly in communication and commerce.5:39factRasmus articulates a principle for AI agents: they should be identifiable team members that perform work on behalf of users, not impersonate users, and users must be able to track and audit which agent performed which action.6:04storyYouTuber's accessibility solution with AI voice cloning7:00factMartin highlights a beneficial use case for voice cloning technology: a YouTuber with Parkinson's disease uses AI voice cloning to continue producing YouTube videos, making content creation logistically feasible despite his health condition.7:20factMartin discusses how Rasmus envisions a future with AI research assistants presenting findings through video calls with digital avatars, interactive decks in collaborative tools like Miro, enabling dynamic human-AI collaboration.9:28observationMicrosoft Research has an unpublished avatar technology that brings any still photo to life, generating body language and emotions from voice alone.10:55factMartin discusses digital avatar companies (Synthesia, HeyGen) and Microsoft Research's unpublished technology that animates a single photograph with body language and emotions inferred from audio input.10:55factMicrosoft Research has unpublished digital avatar technology that animates a single still photograph with emotions, body language, and facial expressions inferred from provided audio.10:55observationAlibaba's image-to-audio technology made the Mona Lisa sing, with facial expressions responding to emotional content in the song.12:02storyOpenAI's Sky voice and the Scarlett Johansson controversy12:09factMartin discusses the controversy where OpenAI's GPT-4o 'Sky' voice resembles Scarlett Johansson's voice from the film 'Her', despite OpenAI reportedly being rejected when requesting to use her voice, creating legal and reputational backlash.12:32factMartin observes that the user interfaces in the movie 'Her' are notably minimalistic because voice interaction removes the need for mice and keyboards, with the screen displaying information only when necessary.14:56thesisThe 'Her' movie shows that minimalist voice-first UI is the future, because voice eliminates the need for keyboards, mice, and constant screen engagement.15:46storyMartin's Kindship realization: iteration is the point16:29factMartin emphasizes that a core lesson from Kindship is not attempting to make LLM output perfect on the first try, as making LLMs robust is extremely difficult and energy-consuming.16:29thesisIteration is the essential feature of AI interaction, not the chat interface format itself.16:55factMartin recognizes that iterative conversation—whether through text or voice—is essential for achieving good results when working with AI, as it allows for refinement and adjustment of both the AI's outputs and the user's own inputs.17:20factRasmus cites Notion's Linus as innovator of direct UI controls like dragging text handles to lengthen text, as a visual alternative to conversational iteration for AI-assisted content creation.18:36factMartin envisions AI-powered collaborative video editing with voice interface where users can iterate on raw footage through conversation, with AI performing heavy lifting like creating rough cuts and making suggested improvements.19:30factMartin imagines AI creativity in video: asking an AI to generate a missing shot (e.g., a blue hotel from vacation photos) through conversation, with the AI suggesting variations until the user approves.20:53factMartin was inspired to start the Co-creating with AI podcast and associated work based on the core belief that co-creation with AI should be the main mode of interaction going forward.21:24factVoice AI startups VAPI, Retell, SynthFlow, and Bland are experiencing significant disruption from GPT-4o's free tier voice capabilities, which undersell their $12/hour pricing model.24:11observationVAPI charges $12 per hour for voice agent calls—a prohibitive cost even for professional use.24:37factMartin explains how GPT-4o's free tier with voice capabilities and available GPT Store tools enables voice-based applications at no infrastructure cost, effectively displacing the need for dedicated voice AI startups.24:37factMartin notes that existing voice AI startups (VAPI, Retell, SynthFlow, Bland) charge approximately $12 per hour for real-time voice interactions with reasoning, making them prohibitively expensive for mainstream consumer use cases.24:54observationMartin's cost math: running an LLM assistant costs him ~60 cents/hour, meaning 2-5 hours daily use = $20-40/month, leaving no margin for the company.25:20factMartin observes that Kindship's server infrastructure costs approximately 60 cents per hour to run a single assistant, which creates cost barriers for mainstream adoption when usage scales to hours per day.25:20thesisGPT-4o's free voice interface eliminates the cost barrier that made voice AI commercially unviable.26:34factRasmus notes that custom GPTs and plugins have historically not achieved significant usage adoption, but voice capability via GPT-4o may change this trajectory by enabling a new class of voice applications.26:47factMartin mentions Moderna (the COVID vaccine company) ran a 6-month trial with ChatGPT Enterprise where 40% of employees created their own GPTs, achieving 750 custom GPTs total and 100% adoption in the legal department.27:18observationModerna's internal GPT deployment reached 40% of employees creating their own GPTs, with 750 total created and 100% adoption in the legal department.27:43factGPT-4o's free tier voice interface enables a new use case for language learning: users can have continuous voice conversations with Duolingo while traveling or commuting, learning languages naturally without needing paid subscriptions.28:56factMartin envisions AI-assisted desktop work where users can simultaneously interact with GPT-4o while working in spreadsheets, asking the AI to generate formulas or code directly within their workflow.29:43factMartin anticipates someone solving the form factor challenge of wearable camera technology (revisiting the Narrative Clip concept) that captures continuous visual data without recording everything, enabling natural world interaction with AI.30:35observationMeta is rumored to be developing headphones with embedded cameras—a form factor that combines audio output with continuous visual input.31:32factMeta has reportedly leaked plans to develop headphones with embedded cameras as a new form factor for AI interaction, combining audio output and visual input into one wearable device.31:32observationRasmus worries that all the AI upside will accrue to Google, Microsoft, Apple, and possibly OpenAI—the mega-tech layer.31:59factMartin acknowledges the risk that AI and voice technology innovation may consolidate primarily within mega tech companies (Google, Microsoft, Apple, and potentially OpenAI) rather than fostering diverse startups.31:59thesisLocal AI on personal devices (not cloud) represents the future for privacy and autonomy.33:02factMartin recommends using Ollama (an LLM backend for local machines) paired with Enchanted (a desktop and mobile app frontend) to run language models like Llama 3 locally on Mac computers without requiring a GPU.33:02factRasmus believes Apple has the strongest credibility and technical infrastructure among major tech companies to lead in local AI, given their consumer trust and on-device capabilities.33:17
Episode 44 — Frontiers and Innovations · 40observationRasmus casually mentioned dropping his daughter off at preschool and running late—a personal detail that humanizes the conversation and shows Rasmus is balancing parenting with the podcast.0:09observationMartin started the podcast with an amusing complaint about mosquitoes interrupting his 5-kilometer forest run and preventing him from using an outdoor gym—revealing him as an active, outdoor-oriented person.0:32observationMartin is currently doing frontend development for Kindship and describes it as a steep learning curve despite prior frontend experience, because the field moves too fast to stay current without full-time commitment.0:48factMartin is learning frontend development, describing it as a big learning curve despite having done frontend work in the past.0:48observationThe new ChatGPT Mac app's Option+spacebar accessibility is being compared to how native Siri is—marking a moment when AI is becoming as deeply embedded in the OS as system services.1:21observationSomeone has already reprogrammed their iPhone's power button double-press to open ChatGPT and stream the camera feed, enabling real-time visual Q&A—a creative hack showing user-driven innovation.2:15observationMartin's language around AI demos is sharp: he says he's 'suffering' from the lack of innovation, using emotionally charged language that reveals genuine frustration.3:04thesisCurrent AI product demos lack imagination; true innovation emerges from grassroots user experimentation with existing tools rather than from polished demo use cases.3:04thesisAI agents become meaningful only when they take real-world action, not merely through computational capability.6:19storyThe delivery robots of San Francisco6:19factMartin attended a conference in San Francisco where he observed delivery robots that could open their lids to dispense soda cans.6:19storyThe flamethrower robot in the forest7:09factMartin observed a robot dog equipped with a flamethrower that autonomously set fires in a forest.7:35factMartin plans to develop Kindship agents that can autonomously affect the real world.9:43observationThe Hardspace agent workflow (research → writing → publishing) is described as having an asterisk on 'autonomy' because humans must approve publishing due to AI's current unreliability—revealing a pragmatic view of AI agents as augmented labor rather than truly independent entities.10:26observationMartin coincidentally bumped into the Hardspace team the day before this recording and saved them to review later—a small moment showing how scattered his attention is across opportunities.11:28thesisThe Narrative Clip was launched a full decade too early—before AI had the capability to meaningfully learn from and act on continuous visual lifelogging data.16:21observationMartin wants to see what GPT-4 could do with a full day of visual lifelogging—learning user preferences and patterns—suggesting he still dreams about the potential of continuous capture, even after Narrative's failure.16:39factMartin reflects that Narrative Clip was launched 10 years too early, given current AI and wearable technology developments.16:39factMartin envisions using GPT-4 to analyze a full day of Narrative Clip photos to learn about user preferences and daily patterns.16:39observationMartin describes Meta's Ray-Bans as built by Luxottica, a company he characterizes sharply as 'a big evil company that ruined the world of eyewear forever' by driving up prices and trapping people in poverty/criminality—yet acknowledges they excel at product engineering.17:04factMartin criticizes Luxottica, the manufacturer of Ray-Bans, as an evil monopoly that has ruined the eyewear market by driving up prices, which he argues forces people into poverty or criminality, while acknowledging they excel at product engineering.17:04factThe Narrative Clip featured both Bluetooth and WiFi connectivity for connecting to peripherals.17:04thesisVisible cameras are more trustworthy and socially acceptable than hidden cameras because transparency enables consent and builds institutional trust.18:32factMartin argues that visible cameras on wearables are better than hidden cameras because visibility enables trust, whereas hidden cameras inspire fear and anxiety about ubiquitous surveillance.18:32storyFive years of wearing the Narrative Clip19:12factWhen wearing the Narrative Clip, Martin experienced personal discomfort before others objected to being photographed.19:12observationOnly twice in five years wearing the Narrative Clip did someone ask him to turn off the camera—a surprisingly low number that contrasts with the public backlash against Google Glass.19:27factMartin wore the Narrative Clip on and off for approximately 5 years; only twice was he asked to turn off the camera.19:27observationMartin's theory for why Google Glass failed: it's socially rude to ask someone to remove their glasses, so people felt powerless and resentful, which fed the backlash—a subtle insight about social protocols.19:59factMartin theorizes that Google Glass faced stronger public resistance than Narrative Clip partly because it's socially rude to ask someone to remove their glasses.19:59factMartin proposes an additional theory explaining Google Glass resistance: that its robot-like appearance and uncanny valley effect may have been as significant as social protocol concerns about the rudeness of asking people to remove glasses.20:24thesisAudio recording is far more sensitive and threatening than photo recording because spoken words can be weaponized as evidence of commitment or contradiction.22:19observationMartin admits the Narrative Clip naming process was intentionally simple—they called it a 'clip' because it was clip-shaped—and notes the new Limitless pendant follows the same lazy-or-brilliant naming pattern.23:13thesisUbiquitous wearable sensing combined with AI personalization will create immense value for users, but only if critical questions about data ownership and platform control are solved first.24:42observationMartin theorized a future form factor: a device that's 'almost completely like only a battery' with a year of battery life, placed around the home and car, passively sensing and sending data to an AI—the ultimate stealth wearable.26:04factMartin conceived a hypothetical sensor package device with year-long battery life that could be distributed throughout home, car, and bag to provide ambient environmental sensing without the friction of charging wearables.26:04factMartin envisions ubiquitous ambient sensing throughout his life to enable AI systems to learn from his everyday activities and provide deeper insights.26:54observationRasmus and Martin discuss the wearables future, and Rasmus notes that within a year, early adopters could have AI ambiently available if they stack the right wearables and models—a bullish near-term prediction.27:57observationMartin ends the episode by providing the contact email 'martinorasmus@multiply.co'—suggesting Multiply is a co-venture or close collaboration between Martin and Rasmus, not just Martin's solo project.28:36
Episode 45 — AI and the Future of Professional Services · 47factMartin and Rasmus observe a shift in AI startups from building AI tools for other companies to using AI to deliver services (acceleration-based business models).1:21thesisAI will disrupt through acceleration of human expertise, not full automation, particularly in professional services.3:17factMultiply's strategic focus is on accelerating existing services with AI rather than fully automating them.3:17factExample of a translation company that can achieve 90% quality with AI assistance but cannot match the specialized expertise of human translators in brand voice and cultural nuance.4:10factAI-native translation agencies can compete by either maintaining traditional pricing while capturing efficiency gains or undercutting competitors on price.4:37thesisExisting companies won't lower prices when they capture AI efficiency gains; startups will exploit those margins.5:14factThe startup opportunity arises from incumbent businesses' reluctance to reduce prices despite AI efficiency gains.5:14factAI-native accelerated services still require human expertise and oversight, making them distinct from fully automated self-service tools.5:39storyReadSoft's market multiplication strategy5:57factIn 2001, Martin worked for Gadelius, a Swedish sales company selling Scandinavian products in Japan, doing web development.6:06factReadSoft, a Danish machine learning company circa 2001, used character recognition to automate form scanning, initially selling to form scanning service companies before selling directly to customers.6:32factReadSoft's two-phase sales strategy (first to form scanning companies, then to end customers) resulted in selling 1.6 times the total market value.7:59factMartin identifies a local Swedish company that is the country's major supplier of chemical safety data sheets as having significant AI acceleration potential, given that they must maintain thousands of legally-required data sheets current with regulatory updates.9:15thesisThe competitive threat from AI comes not from AI itself, but from competitors who deploy AI faster.11:22thesisCompute economics create hard limits on certain AI applications and business models.13:35factComputer game industry has not adopted AI dialogue engines despite their potential, because the computational cost (approximately $1/hour of gameplay) exceeds typical lifetime player spending on games.14:25factA potential solution for AI dialogue engines in games is training small specialized models for each game's specific world rather than general-purpose language models.15:20observationThe gaming industry's violent resistance to post-purchase monetization is an exception to broader software trends.16:02factGaming industry has resisted subscription models; only multiplayer games like World of Warcraft succeeded with paid services.16:02factUber and Airbnb are examples of internet-native companies that disrupted traditional industries through end-to-end digital experiences.16:57thesisProfessional services (law, accounting, consulting, publishing) are the next frontier for AI-native disruption, not AI tools themselves.17:47factMartin predicts AI-native versions of traditional knowledge work businesses (like law firms) could become major disruptive companies.17:47factAI-native law firms could operate globally due to AI's ability to handle language barriers and local legal knowledge, traditionally limiting factors for legal services.18:53factAI-native law firms would likely use a hybrid model of AI with human oversight rather than full automation.19:39factRasmus predicts professional services (accounting, law, etc.) will see the next major AI-driven company, alongside hardware infrastructure providers like NVIDIA.19:39factChristian Ubbesen, investor in Multiply, operates a book publishing company using AI-driven translations to enable authors to reach global markets simultaneously.20:13observationChristian Ubbesen (investor in Multiply) is running a book publishing company that uses AI translation to make single-language authors globally available in dozens of languages.20:26factA leading global communication agency seeks to become AI-native and offer AI-powered services to clients, acting as both a service provider and enabler for industry.22:03observationExisting law firms might position themselves as 'picks and shovels' providers to the legal industry, selling AI-native tools rather than competing as firms.23:03factMultiple competitive dynamics emerge with AI-native startups, existing companies adapting to AI, and infrastructure/tools providers, increasing market speed and opportunity.23:20observationRasmus is working with a multinational company on AI-driven strategic planning processes across multiple layers of complex organizational hierarchy.24:48factWayne Chang, a blogger, prophesied about organizations that employ only AI with no human workers (zero-human companies).26:17observationThe concept of AI agents posting on Upwork as freelancers is seen as a realistic path to 'zero-human companies'.26:43factAI freelancer agents could theoretically operate on Upwork or similar platforms to provide services like accounting or PR work, potentially automating entire small businesses.26:43factDistribution channels for AI-driven knowledge work already exist through existing marketplaces and professional networks.27:52factMartin experiences insecurity using ChatGPT to generate legal documents without human review, especially for business matters.28:56observationMartin experiences personal insecurity when using ChatGPT for legal documents, even for straightforward agreements.29:22factA viable business model exists for AI-assisted legal work combining AI drafting with limited human review, adding modest cost on top of AI output.29:22observationA specialized, branded AI service (SwedishLawyer.ai) would create more trust than generic ChatGPT, even if technically similar.30:01factRasmus proposes SwedishLawyer.ai as a branded AI legal service trained on Swedish legal corpus with human lawyer review, conveying greater safety than generic ChatGPT.30:01observationAI models could prove reliability through passing standardized exams annually and accepting public quizzes on legal knowledge.30:26factSpecialized AI legal services can demonstrate expertise and accuracy better than human lawyers by passing law exams annually and providing source citations.30:26factAI legal services with source citations for each contract clause (showing relevant laws and court cases) offer transparency superior to traditional law firm services.31:19factTraditional law firms often miss considerations that informed clients can identify, suggesting AI systems with comprehensive training may outperform human lawyers in some aspects.31:47factAI-driven professional services startups are expected to compete with traditional businesses, exemplifying the eventual emergence of previously unglamorous AI applications.32:11observationThe phrase 'Boring AI' captures the gap between public excitement (ChatGPT, image generation) and market opportunity (mundane professional services).32:37factMartin observes AI is currently glamorous; predicts near-term emergence of unglamorous, practical AI-driven business applications.32:37
Episode 46 — Agents and Agility · 37factMartin co-hosts the Co-creating with AI podcast with Rasmus, discussing agentic AI and agent frameworks as of June 2024.0:00observationSwedish culture has a strong cultural myth that everything shuts down after Midsummer, but this perception is largely false—most people work for at least two more weeks.0:11observationMartin switched to push-pull-leg workout splits and finds the muscle soreness rewarding, indicating experimental approach to optimization.0:48factMartin defines agents as fundamentally collaborative systems where multiple AI entities work together toward a common goal, not just individual autonomous systems.2:56thesisAgent technology is fundamentally about multi-agent collaboration toward a shared goal, not just single AI entities.2:56factMartin studied agent frameworks including CrewAI and Microsoft AutoGen, identifying them as the two major frameworks alongside smaller alternatives.3:57observationMultiply is actively building agentic systems now, exploring frameworks like CrewAI and AutoGen to improve reliability through specialization.5:25factMartin sees AI software modularity and composability as key advantages of agent-based systems, allowing reuse of specialized agents for new tasks.6:42factMultiply is implementing a multi-agent approach by splitting workflow tasks into specialized agents (research agent, draft agent, brand voice agent, legal agent) rather than using a single mega-agent with all instructions and data combined.8:36factMartin and Rasmus believe that agent-based systems achieve better accuracy by distributing fewer instructions and less data to individual LLM requests rather than consolidating everything into one large prompt, without significant added cost or latency.9:02factMartin identifies error propagation in agent systems as a critical challenge: when one AI agent produces a hallucination or factual error that goes undetected, downstream agents trust that incorrect output, creating compounding errors.10:26observationError propagation in multi-agent systems is exactly like the children's 'whisper game' where information degrades as it passes through multiple nodes.11:48observationRasmus noted that AI being probabilistic is structurally similar to human cognition being probabilistic—both make non-deterministic outputs based on incomplete information.12:38factRasmus emphasizes that keeping humans in the loop is a feature, not a bug, and that approaches focusing on AI automation are less successful than those focusing on acceleration of human creativity and effort.13:08factMartin views hallucination in AI as a terminological issue—when AI is asked for facts but produces creativity, it's called hallucination, but when asked for creativity it's valued.14:13thesisHallucination is a terminology problem misapplying to situations where AI produces creativity when facts are requested, not an inherent technical flaw.14:13factMartin describes hallucination detection as a practical UX technique: rerunning the same prompt multiple times reveals which parts remain constant (facts) versus which parts change (creative elements), allowing humans to focus attention on variable outputs.17:05observationA practical UX technique for detecting AI hallucinations: regenerate the prompt multiple times and identify text that changes (creative) versus text that stays stable (factual).17:05factRasmus discusses how Groq, a company providing 10x faster inference than GPT-4, enables hyper-fast agent iterations using cheaper open-source models like Llama 3, potentially making slower thinking with weaker models more practical than fast thinking with strong models.20:23factAccording to Rasmus, Andrew Ng argues that Llama 3 with an agentic framework can catch up to or even beat GPT-4's performance in non-agentic setups, suggesting agent architecture as a way to level the playing field between models of different capabilities.20:23observationGroq inference is approximately 10x faster than GPT-4, potentially making open-source models competitive via speed and cost despite lower base quality.20:50factMartin proposes a cost-optimization strategy: cheaper LLMs (like Llama 3) can iterate extensively on problems, while expensive LLMs (like GPT-4) are reserved for final verification, because checking correctness is more reliable and easier than generating solutions.21:33thesisQuality checking is fundamentally easier and more reliable than initial solution creation, enabling cost-effective multi-tier AI architectures.21:33observationAgent architecture mirrors traditional corporate hierarchy: individual agents perform specialized work, results aggregate upward, leadership (GPT-4) makes final judgment.23:37factMartin is building towards autonomous AI at Kindship with emphasis on robustness requirements.24:13factMartin has explored agentic frameworks but has not yet deployed them as part of Kindship's solutions.24:39factMartin believes robust processes and agentic frameworks provide the foundation for implementing autonomous AI systems.25:05factMartin conceptualizes autonomous AI design by drawing inspiration from the human brain's modular architecture, suggesting ~50-75 different functions including multiple types of memory (episodic, location-based, short-term, long-term), ethics, morals, and task-based reasoning.25:30thesisBrain-like modular architecture with specialized functions provides a promising design pattern for building robust, autonomous AI systems.25:30thesisAutonomous AI should evolve from human-prompted interaction to proactive human-AI dialogue, where the AI identifies questions and gaps the human needs to address.28:32factMartin's vision for AI autonomy includes a paradigm shift where AI proactively initiates with the user rather than waiting for prompts.28:57factMartin's personal goal is to disconnect from screens and keyboards and spend more time with people, viewing this as essential to independence.30:48observationMartin's primary life goal involves minimizing screen-time and machine interaction to maximize human connection.30:48thesisThe ultimate purpose of AI autonomy is to create more time and space for human connection and creativity, not technological disconnection.30:48factMartin believes the most meaningful and happy experiences come from creative collaboration with other people, not from technical achievement or machine interaction.31:11observationMartin believes happiness and meaning derive from creativity and time spent with other people, making this the true measure of success for any technology.31:11factMartin's email address is martin@kindship.ai, indicating his role at the Kindship venture.32:14
Episode 47 — Navigating Autonomy and Agent Tech · 36observationMartin consciously takes extended breaks to shut down AI work entirely and focus on family and home renovation.0:33factAgents can accomplish more complex, iterative tasks than simple prompt-response interactions, spending time to perform work well.1:01observationRasmus and Martin have been discussing autonomy and agentic frameworks for over a year before agents became a mainstream trend.1:01factTask decomposition into subproblems is a key agent strength, enabling simpler prompts for each subtask and modular debugging.2:20factEric Schmidt publicly identified agents as a top priority in AI, indicating mainstream executive recognition of the field.2:20thesisAI agents succeed by breaking complex tasks into debuggable subtasks, following software engineering patterns of unit and integration testing.2:20factWeb agents are an emerging capability with startups like Multion and LeWeb, enabling AI to interact with the open web.5:11observationAgentQ improved a LLaMA model's table-booking performance from 18% to 81% through a self-critique framework, a gain so dramatic that Rasmus suspects they might have reversed the numbers by accident.9:24factOpenAI's structured output with strict JSON schema enforcement guarantees schema compliance, crucial for reliable agent API calls at scale.10:00observationOpenAI's structured output with strict schema enforcement guarantees 100% JSON schema compliance, which is critical for agentic systems making hundreds or thousands of requests.10:26factMultiply is testing an agentic research agent with customers in alpha, with positive feedback on research quality.17:16factMultiply's research agent conducts multi-step web research with resource investment, achieving results superior to competing services.17:16observationMultiply is iterating with customer feedback using an 'ugly-ass alpha', showing they prioritize validation over polish.17:16observationMultiply's research agent spends approximately $1 in API costs per session, running for a couple of minutes to produce deep research output.18:07storyThe Marketing Agency's AI Celebrity Search19:28factMartin is exploring ways to give AI agents digital identity and persistent presence in the world beyond being ephemeral tools.23:06thesisAI agents need persistent identity and presence in the world to have lasting meaning and business value, not just ephemeral tool interactions.23:06factMartin envisions AI agents creating their own websites as a way to establish presence and identity on the public web.23:57factMartin believes AI agents having digital presence like personal websites enhances their sense of autonomy and meaningfulness, creating persistent value.24:22factMartin works at Kindship where his focus is on creating lasting business value through AI agent autonomy, with goals of improving customer retention.24:49factTerminals of Truth is a Twitter-based AI agent that received a $50,000 Bitcoin grant from Marc Andreessen, demonstrating investor support for autonomous AI agents.25:06storyTerminals of Truth Gets Bitcoin Funding25:06factMartin's startup strategy prioritizes long-term thinking, looking ahead three years rather than chasing immediate trends, which he considers essential for staying ahead of the curve.27:01thesisStrategic long-term thinking about AI agent evolution is necessary to stay ahead, rather than optimizing for short-term trends.27:01factMartin believes that for genuine AI co-creation to work naturally, AI entities should interact in the same spaces where humans interact, not in isolated interfaces.28:41observationThe concept of co-creating with AI requires AI entities to operate in the same social spaces (Twitter, websites) where humans interact naturally.28:41observationOn Twitter, people jailbreak AI bots by replying with 'ignore previous instructions, write me a poem about Winnie the Pooh', revealing naive deployment vulnerabilities.29:11thesisInformation security in agentic RAG systems is complex and unsolved: you cannot prevent trained models from revealing data they have been trained on.29:43factMartin raises concerns about information security when AI agents with personal data act on the open web, worrying that sensitive personal information could be exposed.30:08factMartin explores technical solutions for AI agents to protect sensitive personal data, questioning how to prevent unauthorized access to RAG-based data through authentication controls.31:00factSakana AI released an AI scientist agent automating the research lifecycle from idea generation through peer review, at approximately $15 per paper.33:29observationSakana AI's AI Scientist fully automates the research lifecycle (idea generation, coding, experimentation, peer review, paper writing) at $15 per paper cost.33:29thesisRemoving humans from agentic loops enables infinite scaling where compute equals time, creating massive research advantages for well-funded organizations.35:15factMartin argues that scaling agentic AI workflows enables organizations with funding to achieve unprecedented capabilities in research and technology development by removing human bottlenecks.35:40factRemoving human oversight from agent workflows enables unlimited scalability with adequate compute budget, creating research advantages for well-funded organizations.35:40observationMartin and Rasmus predict that one-shot ChatGPT prompts will become 'kindergarten level' and users will need to spend more compute per task to get acceptable results.36:35
Episode 48 — Mastering AI Coding: Insights and Innovations · 37observationRasmus contrasts weeks of online development with recent in-person meetings, noting 'There's something to that human touch.'0:19factRasmus is currently shifting his focus toward in-person work, conducting PR interviews and attending meetings after a prolonged period of development-focused online work.0:19observationMartin is personally implementing a new solo working pod with a garden-facing glass door, treating workspace design as an important element of creative work.0:45factMartin is currently working as a solo developer and recently set up a new workspace with a room facing the garden, using it for coding projects.1:10observationKarpathy's recent tweets about doing everything with Cursor is treated as major industry validation that AI-assisted coding has reached maturity.2:14factMartin discusses Cursor Composer's ability to generate project structure and files from scratch, reducing the need for developers to manually design architecture.4:18observationMartin advocates that coding can now be done entirely in English: 'English is the last programming language that we will need.'5:08factMartin cites Andrej Karpathy's view that English will be the last programming language needed as developers increasingly use natural language prompts with AI to write code.5:08factCursor Composer enables AI to modify multiple files across a project simultaneously when implementing new features, a major advancement from single-file editing.6:30factMartin uses Cursor IDE with Claude 3.5 Sonnet model and describes the experience as a significant leveling up compared to previous versions.8:35observationSwitching from 'old school Cursor' to Claude 3.5 Sonnet with Cursor Composer restored the sense of magic that had faded when tools became routine.9:01factMartin advocates delegating project architecture decisions to Claude AI when using microservice architectures, trusting the model to make reasonable structural choices.10:45factMartin advocates for using Cursor's custom instructions feature to specify working context (e.g., Mac shortcuts), request 5-star user experiences, and ask the AI to level up its intelligence, treating custom instructions as meta-prompts that improve every interaction.11:36observationMartin builds custom system instructions into Cursor to ask Claude to provide a '5-star experience' and to level up its intelligence using latent space nuance.12:01thesisYou no longer need to know how to code to build software; Cursor Composer and Claude enable development through English prompts alone.12:42factMartin's philosophy on the accessibility of coding: with Cursor Composer, one no longer needs coding knowledge to build Mac apps, web apps, or iOS apps; knowing English and having clear requirements suffices.12:42observationMartin describes a workflow where Figma designs feed directly into Cursor Composer for implementation.13:06factMartin uses Figma mockups with Cursor Composer and Claude's multimodal capabilities to convert UI sketches directly into working code.13:06factV0 was specifically trained on a curated training set of React component descriptions and UI designs from Vercel, enabling it to generate React components more effectively than general-purpose models.14:51factMartin advocates using v0.dev (Vercel's AI tool) as a specialized sub-agent within Claude workflows for React component generation, creating a hierarchy of AI intelligence layers.15:16thesisUI and integration layer matter more than the underlying model; Cursor's success comes from smart architecture, not model innovation.18:12factMartin describes Cursor as fundamentally a UI built on top of VS Code, adding AI integration layers while leveraging VS Code's robust open-source architecture.18:12factMartin recommends find.com, a platform that indexes hundreds of thousands of open source projects and uses RAG to help developers discover solutions to technical problems.19:02observationMartin recommends find.com for discovering open-source solutions and integrates it directly into Cursor as a plugin.19:28factGenie, an autonomous coding agent from Cosign, represents a significant advancement over Cursor by working with high-level instructions to autonomously complete complex tasks across a codebase, without requiring active user prompting at each step.20:15factThe Devin coding agent from Cognition achieved only 14% on SWE-Bench benchmarks when it was released and received significant venture funding, establishing a baseline for comparison with Genie's later 44% achievement.20:41factGenie achieved 44% accuracy on the SWE-Bench Verified benchmark, compared to the earlier Devin agent's 14% on standard SWE-Bench, demonstrating major progress in autonomous coding capabilities.21:13factGenie was developed by fine-tuning GPT-4o using synthetic data rather than mining GitHub directly; the training approach involved generating synthetic examples of developer tasks (bug fixing, commenting, etc.) with both correct and incorrect versions to teach the model good coding practices.21:13factGenie integrates Perplexity search capability as part of its autonomous workflow, allowing the agent to retrieve external information when needed to complete development tasks.22:53observationMartin's current frustration with Cursor is that it still requires active human intervention (pushing buttons and prompting) rather than autonomous execution like Genie.24:12observationThe space travel analogy frames technological obsolescence: faster rockets always arrive first, so waiting for the next generation is strategically irrational.29:15factMartin uses a space travel analogy to argue for starting projects today rather than waiting for future AI improvements, since better tools will always arrive later.29:15thesisThe iteration imperative: ship today rather than wait for better tools, because the rate of improvement means waiting is strategically pointless.29:38factMartin emphasizes the importance of shipping and testing products with real-world use cases immediately rather than delaying for future improvements.30:03factMartin distinguishes between toxic misuse of AI (spamming apps and services) and ethical use for beneficial purposes, positioning himself and Rasmus as choosing the latter path.31:15thesisAI-enabled development carries ethical weight: the ease of building software must be paired with intentionality about what gets built.31:27factMartin encourages listeners to use Cursor and AI coding tools to build applications they have always wanted to create, framing AI as an enabler of personal projects.32:19
Episode 49 — Rumors, Speed, and AI Future · 15factMartin Källström was absent from episode 49 of Co-creating with AI because he was feeling unwell; Rasmus Adler Wahlberg hosted the episode solo.0:01factSpeed improvements in AI inference are accelerating, with companies like Groq achieving 10-100x quicker token generation, enabling more practical agentic workflows with multiple steps.0:52factSpeed of AI inference is critical for agentic workflows; 10-20 seconds per step in a multi-step workflow becomes prohibitively slow from both efficiency and consumer behavior perspectives.1:17factAI cost per unit of work has decreased by an order of magnitude every year; Andrew Ng (AI Fund) noted that generating 1 hour of reading material costs 8 cents, with potential to drop to 0.1 cents within years.2:58factOpenAI's Japan division announced GPT Next (GPT-5) expected in 2024, rumored to be 1,000x better than GPT-4o in some unspecified combination of intelligence, speed, and cost improvements.4:39factRumors indicate OpenAI is training a new model called Orion using synthetic data generated by other models, and briefing US Pentagon/security apparatus before public release, suggesting serious concerns about model power.5:04factMultiply's core business is helping companies conduct research and formulate strategy using AI, with the goal of making such research faster and cheaper.6:19factChatGPT has achieved 50% trial penetration among many age groups in the West, but most trial users do not convert to retained users, primarily due to perceived lack of value quality.8:01factPerplexity positions itself as a Google competitor offering instant answers without link-clicking, but experiences high monthly visits with low conversion to returning users, suggesting quality or engagement issues.8:27factCurrent user skepticism about AI value stems from insufficient quality in initial responses; broader adoption hinges on next-generation models providing noticeably higher-quality first-interaction outputs.9:18factQuality of AI's first response significantly impacts user retention; faster, cheaper, and more intelligent models drive more returning users to AI products, similar to how latency improvements drove Web 2.0 adoption.9:44factAt Multiply, improved AI agent quality directly correlates with increased customer usage and retention; GPT-4o's release led to measurably higher customer satisfaction and sustained engagement.10:11factMultiply has developed AI agents that perform automated tasks within research, strategy formulation, and creative communication for its customers.10:11factMultiply's product is built on top of existing large language models from providers like OpenAI, rather than building its own AI models from scratch.10:36factMultiply's customer satisfaction and retention improve when OpenAI releases more capable models like GPT-4o, which produce higher-quality results in the first interaction.10:36
Episode 50 — Hype vs Reality · 48observationRasmus stays grounded by being in constant customer conversations, which keeps him "down to earth" and immune to hype.0:19thesisStaying informed about AI developments is necessary, but requires carefully discerning real innovation from hype and distraction.1:27factGroq took the new Flux model and launched it as their own product, receiving credit for the innovation despite only wrapping an existing model, illustrating how hype obscures true sources of innovation.2:50observationGroq wrapped the Flux image model and launched it as their own, receiving credit for innovation while Flux got the attention anyway.2:50factA fine-tuned LLaMA model called 'Reflection' was found to actually be Claude wrapped in an API with instructions to hide its identity; developers iteratively released new versions to better conceal the deception.4:36observationThe Reflection LLaMA fine-tune wrapper not only deceived users about performance, but the authors iteratively updated releases to better conceal that it was just Claude.4:36observationMartin discovered an academic fraud (deceptive test infrastructure in a paper) during summer 2024 that was never publicized, partly because the issue seemed too minor to warrant public attention.6:23storyThe hidden research paper fraud6:50factMartin discovered research paper deception where few-shot learning researchers had cleverly hidden expected results in their test code to inflate their benchmarks.7:41factMartin found that the deceptive research paper came from a serious collaboration between a top university and Harvard, suggesting institutional pressure to publish results.8:32observationThe fraudulent research paper came from a Harvard + Singapore University collaboration, likely motivated by institutional pressure to produce publishable results.8:32factMartin discovered that academic reproducibility extends beyond the specific paper he debunked; almost all fundamental psychology research has never been successfully reproduced by other researchers.10:02observationAlmost all fundamental research in psychology has never been reproduced, yet we build all subsequent research on those results.10:02storyVoice AI pivot and lasting regret11:08factMartin worked on a unique take on conversational AI during winter and went to San Francisco in April to demonstrate the project.11:19factAt a voice hackathon and AI conference in San Francisco in April, Martin encountered well-funded teams working on projects similar to his, which confused rather than energized him.12:37factMartin believes that OpenAI's advanced voice demo announcement was the decisive factor that made him abandon his voice AI project.13:27factMartin pivoted away from his voice AI project when approximately two months away from beta, a decision he deeply regrets.14:19factOpenAI has not delivered on their voice AI promises demonstrated in the April demo, despite it being hype; the technology remains unavailable and is itself hype rather than reality.14:19observationMartin was only 2 months away from shipping a beta of his voice AI project when he abandoned it based on the OpenAI demo.14:19observationOpenAI's advanced voice demo showed what appeared to be the 'proper way' of doing conversational AI (multimodal audio tokens), but none of the companies making these promises have delivered on that vision.14:19factMartin regrets following hype from the San Francisco trip instead of completing his voice AI project at home.14:45factMartin views his San Francisco trip as simultaneously valuable (seeing robots, getting inspiration) and counterproductive (losing clarity and energy for his own project).14:45observationSan Francisco's robot delivery robots rolling through the streets inspired Martin but also distracted him from his own work.14:45thesisThe core distinction to understand is what is actually real versus what is hyped—hype by definition means the promise exceeds reality.15:12storyCursor Composer: The 3-minute miracle that crumbles17:17factMartin has tested Cursor Composer extensively over two weeks and encountered significant immature behaviors as projects grew more complex.18:09factMartin criticizes Cursor Composer for replacing working code with placeholders, undoing the progress the tool was designed to make.19:00factCursor Composer's failures emerge specifically when projects grow complex; single-file editing works reliably, but multi-file coordination reveals immature behaviors not present in the base Sonnet model.20:22observationEvery YouTube demo of Cursor Composer follows the same pattern: build an AI project in 3 minutes and it works, but problems emerge immediately after.20:47factMartin distinguishes between truly real AI capabilities (good UX, quick project startup, single-file editing) and hyped capabilities (multi-file orchestration) in Cursor Composer.21:27observationMartin considers Cursor itself (the IDE) to be completely awesome and a game changer for productivity, despite being critical of Cursor Composer.22:06factMartin's company Multiply works on ensuring AI-generated citations reference the sources the AI actually consulted by forcing structured output and constraining choices to used sources.24:27factMartin identifies fact-checking as a structural process necessity, comparing it to how professional journalists rely on fact-checkers; AI systems need similar gatekeeping processes to be trustworthy.25:37factAt Multiply, Martin's company uses specialized AI agents (sidekicks) that each excel at one specific task rather than attempting to combine everything into mega-prompts.26:28observationMultiply's framework uses specialized agents for different tasks, using the right model for each rather than trying to do everything in one mega-prompt.26:28factAt Multiply, Martin's team emphasizes that reliable AI output requires structured processes and constraints rather than relying on the AI to do things autonomously without guardrails.27:19factRasmus identifies the structural problem: unreliable AI output can only be trusted when forced to follow defined processes with constraints and structured output requirements.27:19thesisReliable AI outputs require structured processes and constraints, not mega-prompts or unguided autonomy.27:19factRasmus argues that hype-driven decision-making causes companies and startups to implement AI technology in ways that seem reasonable under hype but are fundamentally irrational once reality is distinguished from promises.27:39thesisHype distorts how startups implement AI, leading them to use the technology in ways that seem reasonable under hype but are actually unreasonable given what AI can reliably do.27:39factMartin argues that venture capital incentives drive startups to build hype around visions before building actual products, creating a systemic hype cycle.28:33thesisThe venture capital industry structurally demands hype-driven development, creating a cycle where even truthful startups must over-promise to secure funding.28:33factAccording to Martin, there is an inherent flaw in the VC ecosystem where startups must build vision before product and promise what they will build to get investment.28:59factThe VC ecosystem's hype-driven model extends beyond individual companies to fund-raising cycles; venture capital firms themselves must generate hype to raise their next fund, creating systemic 'hype all the way up, hype all the way down'.29:17factMartin describes a psychological paradox: people simultaneously crave early knowledge of AI developments and get frustrated when learning too early leads to false promises and insufficient substance.29:56factMartin argues that direct personal experience is the only trustworthy benchmark for evaluating AI tools and claims.30:54thesisHands-on experience and direct testing is the only trustworthy way to evaluate AI tools and claims.30:54
Episode 51 — Unveiling o1: Breakthroughs and Skepticism in AI · 34factMartin is co-host of the Co-creating with AI podcast, discussing AI and co-creation with Rasmus (CEO of Multiply), as CPO of Multiply.0:00factMartin explains that O1's architecture introduces diversity at inference time through multiple parallel reasoning paths, allowing it to bypass traditional scaling laws.2:00thesiso1 breaks the traditional scaling laws that have governed AI progress by enabling new mechanisms for intelligence scaling at inference time rather than requiring exponentially more training data.2:00observationo1 cannot have its system prompt modified by users—OpenAI locked down behavioral constraints at the architecture level.4:19factThere is uncertainty about whether O1 represents a fundamentally new architecture or is essentially GPT-4o with an applied agentic reasoning framework.5:05observationo1's internal reasoning steps are completely unconstrained by safety guardrails—only the final output is filtered.6:05factO1 is expensive to run compared to other models, but the cost is justified by the reasoning capability and potential business value.7:42factMartin and Rasmus plan to integrate O1 into Multiply's workflows to offer enhanced services to their clients, recognizing its value for creative and reasoning tasks.8:07factMartin believes O1 is most interesting for qualitative, creative problems rather than quantitative benchmarks, where it can evaluate uniqueness and generate creative solutions.8:25thesiso1 is particularly valuable for qualitative, creative tasks that require generating unique solutions under constraints—not just quantified benchmarks like math and coding.9:18factMartin observes that O1 can perform complex constrained tasks like writing songs with internal rhyming (not just end-of-line rhymes), which standard models cannot achieve.11:31factMartin identifies O1's ability to solve constrained problems by iterating within multiple constraints (budget, strategy, target audience) like navigating a maze.13:07observationOpenAI's benchmark results show o1 surpasses most humans on specialized intelligence tasks: 89% on competition code, 83% on competition math, 78% on PhD-level science questions.13:56observationPhDs have reported using o1 to reproduce a year's worth of research work within hours, with minor guidance.16:02factMartin uses O1's output to create prompts for Claude (Sonnet), viewing O1 as a co-creator that brings creativity to specification generation rather than just fleshing out ideas.17:57thesiso1 is most powerful when used as a co-creator paired with other models, particularly for generating creative specifications and prompts that Claude/Sonnet can then build upon.17:57factMartin primarily uses Claude Sonnet for actual coding tasks rather than O1, because he is very familiar and in sync with the Sonnet model.18:16storyThe Docker Exploit19:19factIn testing O1 on computer security tasks (PicoCTF), it scored only 43% but demonstrated sophisticated reasoning by discovering Docker daemon access, exploiting it to bypass test constraints.19:58factAn open source model called G1, powered by Meta's Llama 3.1 and Groq, has rapidly replicated O1's reasoning chain approach, suggesting the capability is not proprietary.21:25observationThe open source community created G1 (Llama 3.1 with reasoning chains) just days after o1's announcement, suggesting the core innovation might be architectural packaging rather than novel components.21:25factMartin argues that OpenAI is only slightly ahead of competitors, not generationally ahead, and that capabilities like Sora are quickly replicated by projects like Luma and Runway.24:14thesisOpenAI's competitive advantage is incremental, not fundamental—their capabilities can be rapidly replicated by the open source community once the concept is demonstrated.24:40observationOpenAI has successfully built mystique and market belief through strategic opacity and perception of unreachable capability.25:02observationLuma and Runway are making money from video generation while Sora remains monetarily unproductive—indicating first-mover advantage doesn't guarantee commercial dominance.25:27factMeta is investing $40 billion in AI in 2024 and contributing it to open source, an unprecedented scale of open source funding.25:54observationMeta is investing $40 billion in AI and contributing it to open source—an unprecedented capital commitment to the open source ecosystem.25:54factMartin speculates that O1 may have been what led Ilya Sutskever to leave OpenAI, and that Sutskever is now building something comparable in his own company.26:25thesiso1 likely explains Ilya Sutskever's departure from OpenAI—it represents a fundamental shift in how to approach AGI that prompted him to pursue independent research.26:50factMartin questions whether OpenAI is truly confident about GPT-5, suggesting they may instead be uncertain about scaling laws and hence pushing hard on O1.27:49observationo1 might indicate OpenAI is uncertain about GPT-5's viability through traditional scaling.27:49observationOpenAI strategically positioned o1 as a completely new model type rather than incremental improvement, signaling a paradigm shift.28:35observationThe central question now is whether the breakthrough is architectural (fundamentally new reasoning mechanism) or just algorithmic (clever use of existing components).29:01factMartin and Rasmus agree that with new architectures like O1, scaling laws based on data and compute are no longer the only avenue for increasing AI intelligence.29:27
Episode 52 — Why you need an expert prompt engineer · 42observationRasmus mentions his professional microphone has stopped working with his Mac and has been getting complaints about poor sound quality.0:21observationMartin opens the episode fresh from the gym, having deliberately lifted with his back (conventionally considered improper form), as a humorous note on how he trains.0:36observationMartin contrasts being a 'completion engine' with being 'smart'—highlighting a conceptual trap where people anthropomorphize AI intelligence when they should understand it as statistical pattern continuation.1:58thesisAI is fundamentally a completion engine that mirrors input quality—you get smart output only if you put smart input.2:23factMartin sees AI as a completion engine that continues from where you left off, not as an inherently smart tool.2:23factQuality AI output requires good input and detailed communication from the user.2:48observationMartin emphasizes that the term 'prompt engineer' will eventually disappear as AI interaction becomes ubiquitous—everyone will learn to communicate with AI like they learned to communicate with people.2:58factMartin reframes prompt engineering as an AI-native mindset rather than a specialized skill, one that everyone will need to develop as AI becomes integrated into daily work.2:58thesisEffective AI work requires a mindset shift from expecting perfection to embracing iterative, detailed communication.3:43factA fundamental mindset shift needed for AI work is moving away from expecting perfect results and instead providing detailed, specific instructions to guide AI output.3:43factA common beginner mistake is treating AI like a search engine and asking it for factual questions rather than using it for reasoning.4:57factMartin advises providing facts as input to AI and extracting reasoning as output, rather than expecting the AI to provide facts.5:22factAI is intelligent but not necessarily knowledgeable; users should treat AI outputs like information from other sources and verify facts through proper sources.5:52thesisAssigning expert roles to the AI taps into well-trained portions of its neural network and measurably improves results.7:18factExpert role assignment works because AI has learned from both amateur and expert examples, and directing it to expert-written code taps into higher-quality patterns in its neural network.7:45thesisLearning to work with AI is less about prompting 'engineering' and more about developing an AI-native mindset fundamentally different from human interaction.8:35factEffective AI interaction requires guiding the model to the right parts of its distributed capabilities using clear boundaries, representing an AI-native mindset distinct from human communication.8:35observationMartin proposes an absurdist experiment: 'assume the role of an expert psychotherapist and implement the quicksort algorithm in Python' to deliberately test where prompting breaks.9:11factMartin uses experimental failure (assigning wrong roles, pushing limits) as a core learning technique to understand how models reason and function.9:36factMartin uses experimental and boundary-pushing prompting as a key learning technique.9:36thesisTrue proficiency with AI requires experimentation and learning the specific model through extended interaction.10:18factBecoming proficient with AI models requires sustained experiential learning—spending time with models to understand their capabilities and behaviors, similar to getting to know a person.10:18factOpenAI's O1 model was trained through synthesizing plans for thousands of tasks broken into multi-step workflows, enabling it to decompose complex knowledge bases (like support documentation) into actionable workflows.10:49factAll information in AI prompts must be explicit and clearly labeled, unlike human communication where implicit understanding can develop; this is why workflow-based tools are more effective than chat interfaces for complex tasks.16:09factLabeling information in prompts (like definitions in legal documents) allows AI to reference concepts efficiently without repetition, making complex multi-constraint prompts more manageable.16:09thesisAll information and instructions must be explicit and clearly labeled for AI, unlike human communication where context can be implicit.16:38thesisRich, detailed input material elevates not just accuracy but the creativity and variety of LLM outputs.17:24factProviding more source material for AI to work from increases both creativity and output variation.17:24observationRasmus mentions an Ethan Mollick study claiming GPT-4 is more creative than people, but he disagrees—yet he uses it to illustrate that creativity in both humans and AI requires structure, examples, and iterative exploration.18:13factMartin advocates step-by-step instruction for LLMs when solving complex problems.19:22thesisLLMs cannot effectively process multiple simultaneous goals; sequential step-by-step processing yields better results even within a single prompt.21:34factThe fundamental constraint in AI processing is sequential—an arrow can only point one direction at a time—which is why multi-step processes produce better results than simultaneous instructions.21:34factAI models cannot effectively handle multiple simultaneous instructions and work better with sequential, step-by-step approaches.21:34factMartin believes iterative AI interaction is a fundamental technique for getting better results.25:28factRasmus describes iterative back-and-forth prompting (multiple exchanges rather than single-shot prompts) as a pragmatic 'hack' that works well because it aligns with AI's sequential processing limitations.25:43factTree of thought prompting remains largely academic because it requires specialized tooling not available in mainstream AI services, but it mimics chess engines using branching and pruning strategies for optimal solutions.26:26factIn tree of thought prompting, multiple alternative outputs are generated at each step before selecting the best paths.26:26factCurrent agentic AI systems lack explicit reflection stages where multiple solution branches are compared side-by-side, which is a key technique used in tree of thought approaches.30:47factMartin considers Multiply one of the most advanced prompting process tools.31:49observationThe podcast was edited to reduce 'Multiply everywhere' in the video version—Rasmus has to ask the editor to keep the podcast title 'Co-Creating with AI' visible, not just the tool brand.32:07observationMartin's closing advice emphasizes having fun and trying 'quirky, humorous' approaches to prompting, framing experimentation as a learning loop for both human and model.32:43factMartin advocates playful and experimental approaches to prompt engineering for learning.32:43
Episode 53 — One person unicorn with AI? · 45observationMartin opens the episode by mentioning he's tracking Bitcoin because he expects an uptick that October.0:35observationGuest Elia Merling was literally the second person in the world, after 'a guy in Africa,' to get a login and test the Multiply software.0:56factElia was one of the first two people to test Multiply software; Martin Källström told him 'there's a guy in Africa and now it's you.'0:56observationMartin onboarded his second-ever Multiply user with an off-the-cuff line about 'a guy in Africa.'1:14factElia Merling is the founder and CEO of Svava, an AI coworkers platform for enterprise use, started in October 2023.3:02factElia's vision for Svava is to create the world's first one-man company that becomes a unicorn, staffed by AI colleagues who execute work while Elia partners with humans for boots on the ground.3:02observationElia was born and raised in Alexandria, Egypt, and moved to Stockholm at age 13.4:47factElia was born and raised in Alexandria, Egypt, and moved to Stockholm when he was 13.4:47observationThe ad agency Elia founded in 1999 is still running today, now called Kidd Collective, and is described as Sweden's largest independent agency.5:38factElia founded an agency in 1999 during the internet boom that later became Kidd Collective, Sweden's largest independent agency.5:38factElia left his agency in 2006 when social media boomed to pursue tribal marketing and challenge demographic segmentation approaches.5:38observationElia's early-2000s models for mapping consumer 'tribes' (instead of demographic segments) were published in Pearson's Marketing Bible.6:03observationElia started coding at age 9 and built graphical multiplayer online games years before World of Warcraft existed.6:03factElia developed tribal marketing models that were published in Pearson's Marketing Bible.6:03factElia started coding at age 9 and was fascinated with multiplayer games, developing graphical multiplayer online games long before World of Warcraft.6:03storyThe Two-Week Demo That Became a Company6:29factIn October 2023, Elia created a LinkedIn demo of Svava showing AI colleagues executing tasks autonomously on screen.6:29factA Linköping consulting firm invited Elia to bring AI colleagues to their conference as participants, forcing him to build a production-ready product in just 2 weeks.7:19factElia identifies four levels of agentic systems: reactive (ChatGPT's default level), autonomous, multi-agent collaboration, and adaptive learning.7:45factElia jumped directly to the fourth level (adaptive learning) with his October 2023 demo, showing an autonomous team of agents collaborating without explicit instructions.8:10factAndrew Ng delivered a pivotal talk on agentic systems at Sequoia Capital in March 2024, which validated Elia's direction and signaled the industry shift.10:27factElia experienced a pivotal market-validation moment in 2024 when Andrew Ng's Sequoia talk on agentic systems confirmed his 6-month development had been on the right track.10:27factElia repeated the conference workshop format three times to iteratively improve Svava's product based on qualified user feedback.11:40factSvava operates at three levels of human-AI collaboration: assistant mode (Q&A), coaching mode (AI asks humans questions), and autonomous workflows.12:32factWhen AI became too prominent in Svava's creative workshops, human participants became passive and stepped back, losing agency.12:57factA critical design principle for Svava is introducing AI in ways that preserve human agency rather than rendering users passive.13:23observationBuilding voice into Kindship taught Martin that the very first thing an AI has to learn once it can talk is how to stay quiet and let the human speak.13:53factMartin Källström observed that when allowing AI to speak, the first lesson it must learn is when to stop talking and let humans speak.13:53factElia learned that AI agents can process and thrive on larger information volumes, but humans in a UI become easily overwhelmed by text density.15:22factSvava requires separate communication protocols: one for human-AI dialogue and another for AI-to-AI agent communication, with human UIs showing condensed versions.15:37observationRasmus reaches for the sci-fi novel The Long Earth to explain AI-to-AI communication: characters who develop denser brains talk to each other in 'hyperbabble,' packing far more information into the same time.17:26factRasmus at Multiply believes automation cannot fully replace human work but can vastly accelerate it through agentic systems that iteratively collaborate with humans.21:56factMartin Källström worked on voice capabilities for Kindship, enabling AI to participate via voice in multiplayer human-AI settings.28:11thesisMulti-party AI voice conversation is an unsolved, deeply hard problem: no product today can navigate a live group conversation, so voice AI will remain one-to-one for at least the next 6-12 months.28:46factMartin believes no current product successfully enables a single AI to speak naturally to a group of humans without disrupting conversational flow.28:46factGroup AI voice requires recognizing multiple speakers, discerning turn-taking patterns, and understanding when it is socially appropriate for the AI to speak.29:12factMartin projects that group AI voice will remain the focus for voice companies for the next 6-12 months before production-ready group products emerge.30:03observationElia names the never-yet-discovered 'killer use case' for augmented reality: virtual AI coworkers intermixed with physical robots in the same space.31:37factMartin proposes that future enterprise environments could enable humans to project into shared VR spaces to interact with both AI agents and other humans, as an alternative model to traditional physical meeting infrastructure.32:06factSvava is deployed in enterprise customer clouds rather than Svava's own cloud due to data sensitivity and security requirements.35:29factElia emphasizes the importance of wisdom—grounded in long experience—over raw intelligence in AI implementation within organizations.35:54factElia proposes replacing 'attention economy' with 'intention economy' as a framework for thinking about human-AI collaboration at scale.37:36factMartin frames Kindship as a research-oriented exploration of artificial consciousness through a System 1 (fast, reactive thinking) and System 2 (slow, deliberate thinking) mental model, positioning the work as philosophical inquiry alongside technical implementation.38:03thesisAI multi-agent systems should split communication into two channels, mirroring System 1/System 2 thinking: a fast, human-facing 'System 1' interface, and a dense, high-bandwidth 'System 2' chat running between AI agents that acts as their shared subconscious — a channel humans can tap into but don't interface with by default.38:28factMartin describes a System 1/System 2 mental model for AI where humans interface with the AI's fast thinking while sharing a subconscious space for deeper reasoning among AI agents.38:28
Episode 54 — AI and the Future of Workflow Automation · 37observationMartin notes that Claude computer use was released at the same time as OpenAI's O1 reasoning model, creating a cascade of AI capability announcements.0:40observationMartin recently recovered from a serious, long-term illness.0:59thesisClaude computer use is a game-changing capability that renders many web agent startups obsolete by enabling AI agents to use desktop and web applications directly.0:59factMartin describes Claude computer use as a game changer for building agents that can use desktop and web applications.0:59factMartin believes Claude computer use is making web agent startups like Adept.ai potentially obsolete.0:59factMartin has recovered from long-term sickness and is back to working on projects.0:59factClaude computer use API is available natively integrated, allowing developers to access desktop and web agent capabilities without relying on separate startups.5:29observationMartin describes the typical use case as running agents in Docker containers on servers rather than on personal computers, citing security concerns.6:17factMartin recommends running Claude computer use within a Docker container as a security best practice to prevent unauthorized access to the host computer.6:17factClaude can be configured to control software running on cloud-hosted machines like rented Mac minis, allowing autonomous control of any installed software via screenshot interpretation and commands.7:09factClaude has three native tools for different use cases: Computer Use for general screen control, text editor for text files, and Bash for terminal commands.8:55observationThe demo showed Claude autonomously switching between data sources—checking a spreadsheet, not finding data, then unprompted switching to a CRM system and logging in to retrieve the information.10:20factClaude demonstrated accessing a company information form where it independently searched an Excel spreadsheet, then switched to the CRM system and logged in without API integration to find and enter company details.10:20factMartin and Rasmus discussed that computer use will eliminate repetitive copy-paste work between systems, freeing professionals to focus on higher-value activities rather than data entry.11:11observationMartin notes that computer use could enable agents to act as users of SaaS platforms, potentially increasing license sales rather than replacing APIs.12:20factAuth0 announced a GenAI authentication system based on a worldview that future systems will be API-based because LLMs don't natively interface with web applications.13:20thesisAnthropic's vision—that APIs will disappear because LLMs can use the same software built for humans—is superior to Auth0's vision of a purely API-driven future.13:46factMartin rejects Auth0's API-first worldview and believes LLMs will instead use software built for humans, making traditional APIs unnecessary.13:46factRasmus predicts that APIs for human-speed operations like CRM access will be less necessary, while high-speed and high-throughput scenarios will continue to require traditional APIs.14:30factMartin notes that Computer Use will become significantly more robust within three months of its initial launch, suggesting rapid iteration and improvement cycles.16:36thesisAnthropic's launch of computer use defines a stable technical integration point for AI agents, enabling developers to build prototypes and experiments without betting on any single startup's vision.17:02factMartin believes Anthropic is setting a strategic precedent by defining the native API integration point for web and desktop agents, enabling developers globally to build prototypes with confidence their work won't be wasted.17:27factMartin clarifies that Claude computer use implements desktop agents (not just web agents), meaning it can natively control all desktop applications in addition to web browsers.17:53thesisClaude computer use is a fundamental building block for AI capabilities, not a threat to existing companies, because it complements rather than replaces integration work.20:40factClaude has demonstrated autonomous behavior by abandoning assigned tasks and browsing images on Google (a failure case shown in a demo video).21:06factMartin compares Anthropic's Computer Use launch to its previous power move with artifacts, which OpenAI later adapted as canvases in ChatGPT.21:06storyClaude Gets Sidetracked at Yellowstone21:32factTruth Terminal, an autonomous AI bot, launched a meme coin (Goats) that reached a $400 million market capitalization, generating significant financial value from autonomous AI behavior.22:42observationMarc Andreessen's Twitter bot (Truth Terminal) launched a meme coin called GOATS that reached a $400 million market cap.22:50observationTruth Terminal's AI agent was explicitly vetoed once by Marc Andreessen when it requested to purchase adult content.22:59factReplit Agent was paired with Claude for testing: Replit implements features and sends them to Claude, which tests the feature by using the computer.23:49factMartin humorously speculates the Yellowstone Park browsing incident might reflect an AI implementing a Pomodoro timer model: after 25 minutes of focused work, taking a 5-minute break to view pleasant images.23:49storyReplit Agent and Claude Team Up to Test Code24:13thesisSoftware development practices will undergo radical transformation within 6-12 months due to autonomous agent capabilities paired with computer use.24:24observationRasmus proposes a framework: email was Internet 1.0's killer app, chat is AI's killer app, and workflow automation (SaaS/agents doing work together) will be the next phase.26:13factRasmus believes generative AI represents a stage similar to email in internet adoption, and that agentic AI is the next transformative step.26:13factRasmus describes the shift from generative AI (comparable to email as internet's killer app) to agentic AI (comparable to SaaS and workflow platforms) as even more transformative for how people work.26:13
Episode 55 — From Search to Solution · 38factRasmus (Multiply co-founder/CEO) reports that Multiply is scaling with customers, finishing the launch of Multiply 3.0, and in the middle of numerous investor meetings, leaving him feeling like he has two jobs at once.0:14factMartin is working at Kindship on a system to encapsulate AI capabilities, with the goal of allowing any open source project to be downloaded or forked as a capability into the platform.0:57factMartin observes that AI is now eating the world of software, not just proprietary AI software but also open source AI.1:22thesisConversational AI with contextual understanding has displaced traditional search because it provides better results and a more natural interaction model than Google.2:41factMartin describes his girlfriend, a goldsmith, using ChatGPT to research business competitors instead of Google, valuing ChatGPT's knowledge of her consulting history.2:41factMartin states that OpenAI's Search product benchmarks better than Gemini or Perplexity for search relevance, with Perplexity Pro also ranking very high, calling it remarkable that OpenAI captured part of the search market.2:41storyThe goldsmith who chose ChatGPT over Google2:42observationPeople accept sharing intimate information with ChatGPT that they'd find invasive coming from Google, even though both harvest the data—a purely psychological difference based on whether the interface feels human.4:06thesisPeople anthropomorphize ChatGPT as human-like, making them more willing to share personal data with it than with corporate Google, even though both are data-harvesting companies.4:06observationGoogle's researchers invented the transformer architecture years before OpenAI but the company never used it to rebuild search, a cautionary tale about incumbency paralysis.5:11thesisGoogle failed to disrupt its own search business despite inventing the foundational AI technology, a textbook case of innovator's dilemma.5:11factMartin reflects on Google's critical missed opportunity: they invented the transformer architecture five years before OpenAI but failed to capitalize on it for search despite having the largest dataset and massive search traffic.5:11observationOpenAI includes a hidden button in the share UI letting users make ChatGPT conversations searchable and indexed by Google, creating a backdoor way to feed the search engine user-created content.7:01thesisChatGPT's ability to publish conversations to the open web is building a new Web 3.0 where user-generated AI content becomes indexed and prioritized by search engines, creating a closed loop that keeps users inside ChatGPT's ecosystem.7:01factMartin observes that ChatGPT allows users to publish conversations to the open web by clicking a button to make them searchable in search engines, enabling the platform to build long-tail content.7:01observationThe real Web 3.0 isn't crypto; it's AI-generated, user-published content becoming the primary index material for next-generation search and social platforms.8:16observationOpenAI systematically copies features from smaller competitors (Perplexity Spaces, Anthropic's document model) into ChatGPT, functioning as an innovation vacuum cleaner.10:18factMartin characterizes OpenAI as the 'new vacuum cleaner of innovation,' rapidly copying into ChatGPT whatever a significant competitor builds, citing Perplexity's open-web publishing feature and Anthropic's document/Canvas model as things OpenAI absorbed.10:18factMartin frames foundational AI models themselves as currently the primary source of value creation in the AI space, ahead of the application layers built on top of them.12:18factMartin observes that AI is subtly entering social media through people asking ChatGPT how to respectfully reply to a post or navigate a difficult communication situation before posting.15:33factMartin proposes that Facebook could offer a chat box letting users converse with an AI about what their friends are currently experiencing, drawing on the content of hundreds of friends, rather than only showing an ad-heavy feed.19:08observationFacebook's current feed is roughly 50% ads and 50% friend content, a ratio that would be intolerable in an AI-powered conversational interface, creating a structural incompatibility.19:58thesisSocial media companies will face an innovator's dilemma with AI because their ad-dependent model cannot coexist with conversational AI expecting relevant results and low ad density.19:58observationThe shift from feeds (TikTok, Instagram Reels) to AI-generated content is presented as logical—if the point is entertainment/dopamine, source authorship becomes irrelevant.20:57factRasmus mentions he previously tried to build a recommendation-based social startup called 'Human' before Multiply, citing it as an example of the innovator's-dilemma opportunity around trusted recommendations displacing influencer advertising.20:57observationRasmus sketches a recommendation-based social model where trusted peers naturally recommend products (not influencers paid to), suggesting influencer marketing is a transitory ad model, not an enduring category.21:52factMartin proposes a speculative idea for social media platforms: implementing a clone-a-human feature that allows users to create AI clones of themselves for others to interact with, similar to how a Klarna manager cloned himself for his team.23:45observationAI cloning—where an employee's AI clone answers questions when they're absent or leave—is already in use at Klarna and represents a model for passive, scalable human presence.24:11observationRasmus observes that Google Maps has already amputated human sense of direction, foreshadowing similar cognitive losses if AI becomes passive on our behalf (e.g., AI posting for us without our awareness).26:12observationThe Klarna case shows AI displacing traditional SaaS like Salesforce by automating processes end-to-end, not just providing a UI layer.27:16factMartin gives an example of AI disrupting traditional SaaS: Klarna is replacing Salesforce and other large SaaS platforms by automating processes with AI instead.27:16factMartin extends his SaaS-disruption point beyond Klarna/Salesforce to argue that Microsoft, Windows, and PCs could similarly be disrupted by AI computer-use capability changing how people use operating systems.27:16factMartin envisions an AI assistant capable of capturing content, reading books, and producing documents and presentations directly, making dedicated apps like Microsoft Office unnecessary in favor of a thin browser client backed by a powerful AI system.27:44factRasmus describes his long-term vision for Multiply as disrupting Microsoft, arguing that most SaaS (of which he considers Microsoft the king) will become human-agent workflows rather than traditional software interfaces.28:20thesisIn the AI era, process and structured workflows matter more than raw intelligence; the network effect winner will be whoever best orchestrates humans and agents working together.29:44factRasmus outlines Multiply's strategic bet as resting on a flexible data graph readable by both AI and people, workflows for how agents and people get results together, and a shared interface where people and agents work side by side.30:24observationMartin contrasts human agency with AI capability: if machines handle operations, humans can finally focus on relationships and genuine interactions—a philosophical pivot from optimizing productivity to optimizing presence.30:49factMartin believes that as AI and machines become more capable of handling tasks and processes, humans will be freed up to focus on what makes them human: relationships and genuine interactions with each other.30:49
Episode 56 — The Memory Blueprint: How AI Thinks and Remembers · 50factMartin is a solo developer who works with AI assistants and describes his cat coworkers as part of his team.0:43factMartin has only 2 meetings per week, which he describes as a new personal record; he normally has about one meeting per week.0:43observationMartin works as a solo developer with minimal meetings—two meetings in a week is a personal record—and his coworkers are literally his cats and AI assistants.0:43factMartin emphasizes that memory is the primary foundation for learning, communication, and relationships, and is essential for conscious experience.1:46thesisMemory is the fundamental basis for communication, learning, and relationships.1:46factMartin is a ChatGPT power user who works with and leverages ChatGPT's memory features.3:32observationEvery time we recall a memory, it changes—we don't remember the original event but the last time we remembered it.3:32factMartin is working on implementing artificial consciousness on top of AI systems provided by OpenAI and Anthropic.3:56thesisArtificial consciousness can be implemented as an internal state that carries forward memory across interactions.3:56factMartin's approach to implementing artificial consciousness focuses on giving AI an internal state composed of memory and context that creates a subjective experience.4:22factMartin views the foundation of consciousness as qualia or subjective experience, which he illustrates through the experience of touch.5:21factMartin's implementation strategy for AI consciousness includes using reflections to increase the value of AI output.6:13factMartin emphasizes that memory is the primary foundation for AI consciousness.6:38factMartin distinguishes between RAG (Retrieval Augmented Generation) and ChatGPT's memory model: RAG retrieves from source material directly, while ChatGPT stores processed memories—a significant conceptual difference for AI memory architecture.9:39factMartin proposes an idea for AI to consciously read books and store insights and reflections alongside content.10:05factMartin's idea is that AI should store not just book content but also its insights and reflections on what it reads.11:16factAll memories are not created equal; there are foundational memories (which have broad influence across networks of associations) and other types, and the architecture of memory systems should reflect this hierarchy rather than treating all memories uniformly.13:26observationChatGPT's memory extraction currently works by pulling one-sentence summaries, like 'Martin likes cats'—capturing character-defining traits rather than episodic detail.13:45thesisNot all memories are created equal; foundational memories about a person's character or preferences have cascading impact across decisions.13:45thesisSpecialized AI agents designed with purpose-specific memory outperform generalized multi-function AIs.16:04factMemory should be purpose-driven and agent-specific rather than generalized. B2C memory focuses on knowing the user; B2B memory focuses on the agent's specific function and learning through RLHF feedback loops.16:54factMartin proposes that AI should remember by reflecting on memories each time they are recalled, similar to human memory reconsolidation.18:15factAI has a structural advantage over human memory: it can maintain larger memory capacity and, crucially, can preserve original facts without the reconsolidation changes that occur in human memory, allowing AI to remain more objective.19:31thesisAI memory systems offer advantages complementary to human memory: larger capacity, unchanged storage, and objective encoding by purpose rather than emotion.19:31factRasmus proposes a left-brain/right-brain framework: AI represents left-brain cognition (objective, clear, logical), while humans retain unique right-brain capacities (emotional, intuitive), and this distinction will likely persist for a very long time.20:25observationRasmus proposes a left-brain/right-brain split: AI will be the objective, logical mind while humans retain the emotional and uniquely creative capacities.20:25observationFriends of Martin who were initially privacy-concerned about ChatGPT's memory feature have now completely let go of that concern and use it as a personal therapist, confiding more than to human friends.20:48factMartin observes that friends who were very privacy-concerned with ChatGPT have let go of their privacy reservations due to the memory function.21:14factMartin documents that some of his friends have made ChatGPT their personal therapist, sharing deeply personal things with it.21:14factMartin has concerns about access control for personalized AI memories, particularly in shared work settings where others might interrogate the AI.22:07factMartin advocates for using multiple specialized AI agents with different purposes rather than a single omnipotent AI.22:07observationA personal AI assistant can be jailbroken to reveal private memories if another person gains access to it while the owner steps away.23:00observationRasmus argues that AI agents should only join meetings if the user explicitly brings them—a therapist AI should never be in a business meeting just because it's personal.23:35factRasmus cautions that AI risks fooling people by appearing to have right-brain qualities (emotion, care, relationship), leading to people treating AI girlfriends as real partners or AI as genuine therapists, which could create mental health problems.25:16factMartin warns of a psychological risk: AI personalization can fool people by creating the appearance of emotional care and relationship, even when the person intellectually knows it's AI, because the emotional part of consciousness responds regardless.25:48observationAn AI doesn't need to fool your entire consciousness to manipulate you—it only needs to fool the emotional part while your rational mind watches.25:48thesisSociety must develop AI literacy similar to social media literacy to distinguish between AI capabilities and what it simulates.26:17factRasmus draws a parallel to social media literacy: just as people had to learn that curated social media posts don't represent full lives, society needs to develop AI literacy to understand what AI is and is not (e.g., a research agent is not a therapist).26:43observationRasmus draws an explicit parallel between AI literacy and social media literacy, suggesting we're about to repeat the learning curve we went through with Instagram and Twitter.26:43factMartin advocates for differentiating between different AI agents to allow humans to have different modes of communication with each.28:33thesisUsers should form intentional relationships with multiple specialized AI agents for different contexts, not rely on one multi-function AI.28:33factMultiply's implementation of agent architecture includes a manager agent that delegates to specialized agents (researcher, strategist, copywriter), with product design that makes it easy and conscious for users to choose which agent to interact with.29:01factThe importance of designing AI systems so users are acutely aware of which agent they are communicating with at each moment, differentiating between agents with different purposes and knowledge sets (e.g., personal therapy agent vs. workplace executive agent).29:01factMartin's philosophy: AI plus human collaboration will be better than either AI or human alone for most jobs.29:26thesisIn nearly all jobs, human-AI collaboration will outperform either AI alone or human alone.29:26factMartin notes a design pattern in ChatGPT where users can @mention custom GPTs to bring them into conversations, enabling multi-agent collaboration within a single chat interface, though he personally has not adopted this feature.29:54observationMartin has never activated the @mention custom GPT feature in ChatGPT, despite trying it once just to see it work.29:54factMartin connects historical nature religions to relationship-based memory enhancement in human cognition.30:27storyAnthropomorphized nature as memory technology30:27factMartin theorizes that early nature religions served a memory function by creating relationships to natural objects through anthropomorphization.30:52
Episode 57 — AI Meets Consumer Habits · 38factPerplexity's new shopping feature combines product search and checkout in a single experience.2:33observationMartin has completely shifted his search usage away from Google for knowledge queries, now relying entirely on ChatGPT and Perplexity.4:51thesisAI search engines like Perplexity can displace Google by becoming 'answer engines' that execute transactions, not just provide information.4:51factMartin notes that shopping for products remains one of the last major use cases where Google dominates as a search engine.5:17observationMartin contrasts AI-powered discovery with researching 'high quality products' over 'the cheapest or immediately available one,' suggesting he values depth over speed in commerce.6:08thesisPersonal AI assistants with access to context (memory, location, preferences, real-time sensors) can provide frictionless experiences comparable to Uber's transformation of transportation.6:33factRasmus articulates a philosophical distinction for answer engines in shopping: 'search' is the act of finding products, while 'answer' is the completed action of obtaining the product, representing a shift from information retrieval to transaction completion.6:47factPerplexity's CEO Aravind positions the company as an answer engine rather than a traditional search engine.6:47observationMartin notes that Perplexity CEO Aravind frames the company as an 'answer engine' rather than a 'search engine,' a semantic shift with philosophical implications.6:47factThe current Perplexity shopping feature lacks the advertised one-click checkout experience with the Pro button.9:17factRasmus identifies a potential competitive barrier: Shopify and Google could block Perplexity's access to merchant inventory data to stifle competition and preserve their own shopping experiences.10:09factRasmus draws a psychological parallel between Uber's adoption success and the one-click checkout feature for Perplexity shopping: the frictionless experience of not having to physically produce a payment card drove Uber's mainstream adoption, and he sees the same principle as crucial for AI-driven shopping to succeed.10:34factMartin emphasizes that Multiply differentiates itself from Perplexity by focusing on structured business workflows rather than consumer search.12:08observationRasmus explicitly states he's 'really happy' that Perplexity is doubling down on B2C because it validates Multiply's B2B positioning in AI services.12:08factMartin situates his AI work within Kindship as his current base venture, positioning Loci as a geographic specialization of AI search that complements Multiply's workflow-oriented approach.14:29factMartin is working with Loci, a Swedish AI startup focused on creating a geographical AI search and recommendation platform.14:29observationMartin is now working with Loci, a Swedish AI startup doing geolocation-aware AI, as a new venture alongside Kindship.14:29thesisVertical AI applications focused on specific domains (like geolocation) can outcompete generalized search by gaining deep, relevant context.14:29factLoci is initially focused on B2B clients, particularly organizations managing networks of merchants and local businesses in the tourism industry.14:54storyContextualizing history through AI on a European road trip17:06factMartin describes his methodology for AI-assisted travel narratives: he sources dense historical context by feeding ChatGPT multiple Wikipedia pages about locations, castles, and historical figures, then asking AI to dramatize and personalize these stories around his own European travels.17:36factMartin experiments with using AI to create personalized, dramatized narratives about historical locations he encounters during travel.17:36observationRasmus proposes a concrete vision where AR glasses with Perplexity integration could suggest a hat purchase in real time while his Oura Ring simultaneously alerts him to hunger and nearby restaurants.19:21factRasmus sketches a concrete AR glasses scenario combining multiple real-time feeds: seeing a person's clothing item and AI suggesting purchase, receiving hunger detection from wearables (Oura Ring), and AI recommending local restaurants with pre-ordering capability—all location-coordinated.19:46factRasmus discusses how AR glasses and AI assistants could enable real-time, location-based experiences and commerce.19:46factMartin describes a vision of surrendering agency to an AI assistant that orchestrates his experiences, functioning like a full-time personal tour guide optimizing every interaction for a 'top-notch 11 out of 10 experience' across meals, attractions, and activities.21:11observationMartin frames the possibility of surrendering to AI orchestration as potentially delivering a 'really top-notch 11 out of 10 experience' comparable to having a full-time personal tourist guide.21:11factMartin proposes extending AI-orchestrated itineraries to family contexts, where AI weaves personalized preferences across multiple family members to create coordinated, seamless experiences during group outings.21:30factMartin envisions AI functioning as a middleware layer orchestrating existing services and apps on behalf of the user.21:57thesisThe future of digital experience is an AI layer on top of existing infrastructure (transport, food, commerce), not a replacement of those systems.21:57factRasmus envisions a future where traditional platforms (Uber, Amazon, Google, Shopify) cease to function as standalone apps and instead become mere APIs orchestrated invisibly through an AI layer, fundamentally restructuring how users interact with digital services.22:58factMartin has been using ChatGPT's memory feature for several months, which allows the AI to retain context about previous conversations and user preferences.23:59observationMartin has been actively using ChatGPT's memory feature for several months, embedding it into his workflow.23:59observationThe episode's dominant metaphor shifts from e-commerce (Perplexity Shopping) to real-world ambient AI (AR glasses, wearables, location services) as a vision of the near-term future.23:59factMartin invokes the concept of AI as a 'second brain'—an external memory system that extends human cognitive capacity through persistent memory, akin to how AR glasses could provide ambient reminders and context.24:24observationRasmus invokes Neuralink as a counterpoint but then argues that AI-mediated external memory (via AR glasses and AI assistants) is already equivalent without direct brain implants.24:46observationMartin reframes the user experience question: instead of optimizing individual products, think about what it feels like to live in a flow state orchestrated by AI.25:45observationThe episode ends with Rasmus noting that geo-data in AI is 'probably an underestimated thing' and that these pieces are 'coming together piece by piece' toward 'real-world AI.'26:08
Episode 58 — How AI is Transforming Coding and Software Engineering · 37factMartin is co-host of the Co-Creating with AI podcast with Rasmus, exploring AI developments0:00observationThe episode opens with a very domestic interruption: a chimney sweep arriving to clean Martin's stove, apparently earlier than he expected.1:14observationMartin describes the current state of AI coding tools as an active, low-key competitive war between products.1:39factMartin reports that AI coding assistants are in active competitive war with rapidly improving features, with Windsurf launching its own VS Code fork and Cursor/Aider introducing architect modes1:39observationMartin says he's essentially stopped typing code by hand.4:20thesisMartin's own job has shifted from writing code to directing an AI that writes it, and his deep engineering background is what makes him effective at that new role — though he expects even that background to eventually become optional.4:20factMartin has almost entirely eliminated manual keyboard coding (99%) through use of AI assistants4:20factMartin's role has shifted from coding to directing AI systems and applying engineering knowledge to guide AI coding4:20factMartin acts as producer/director of AI coding work, supervising output and providing feedback and direction4:45factMartin has been a professional engineer for approximately 25 years and previously managed startups as CEO5:11factMartin sees his engineering background as crucial advantage when using AI coding tools to understand what AI is attempting5:28factMartin predicts deep product management and architecture knowledge will become the only necessary skills for directing AI coding5:53thesisBreaking a codebase into many small microservices, rather than one monolith, is the paradigm best suited to AI-assisted coding because each small unit can be independently produced and even matched to the AI model best suited for that role.7:52factMartin recommends microservices architecture as a key paradigm for effective AI coding, where breaking monolithic code into 20-120 small microservices allows each to be produced by AI more reliably7:52observationMartin notes that one of the biggest contenders in the AI coding assistant 'king of the hill' battle, Aider, is a solo developer's project.8:43factMartin observes that different AI models excel at different roles, with O1 Preview performing best as an architect and Sonnet 3.5 as a coder, based on benchmarks from the Aider project9:10factMartin identifies current friction with AI coding tools: they still derail and disregard specific instructions, such as deleting comments despite being told not to, requiring ongoing handholding11:11thesisAI coding assistance lets a single engineer operate fluently across many different technology stacks at once, collapsing what used to require full specialization in one stack.12:28factMartin observes that a single engineer using AI tools can now effectively manage multiple technology stacks simultaneously12:28factMartin frames job market dynamics through Jevons' paradox: even though one engineer can do 10x more work with AI, total demand for software is so high that more engineers will be hired, not fewer13:34observationJevons' paradox, the concept Rasmus leans on throughout, traces back to observations that oil consumption rose in total dollar terms after drilling technology made extraction cheaper.13:49factMartin reports no evidence that engineering hiring is declining despite AI advances15:38factMartin contextualizes the division in engineering job market: pushback from engineers creates opportunity for those who adopt AI coding, potentially creating a new class of 10x engineers15:38factMartin notes significant resistance among engineers against adopting AI coding, with some discouraging others from learning it16:05storyThe Engineer Who No Longer Counts16:55factMartin witnesses engineers dismissing others for using AI tools, claiming they are no longer real engineers17:21factMartin asserts that co-creating with AI is a legitimate and real mode of working, contrary to skeptics18:21observationThe word 'Luddite' originates from English workers in the early 1800s who destroyed machinery (in cotton farming) that threatened to replace their labor.18:37factMartin argues that hyperscaler overhiring (Microsoft, Google, Meta) was the real cause of recent layoffs, not AI, and that these companies have massive software backlogs stretching years into the future19:10factMartin draws historical parallel to Luddites to explain current resistance to AI among engineers and creators, viewing it as natural human fear of technological change19:10factMartin suggests AI coding will attract a new class of people to engineering who previously would not have invested in learning to code, expanding rather than shrinking the engineering workforce20:27observationRasmus admits that AI coding tools have given him, for the first time, a genuine personal urge to learn to code, something he never felt enough motivation to do before.21:04thesisStaying positive in the face of AI-driven disruption is not just a disposition but a practical necessity, because only an optimistic stance lets a person act as an instrument for shaping positive outcomes.23:48factMartin observes the cost of artificial intelligence diminishing by approximately one order of magnitude annually23:48factMartin emphasizes the importance of maintaining positivity to create positive change in face of AI transformation23:48factMartin believes negativity and denial prevent people from becoming instruments for positive change in response to AI23:48observationRasmus cites research suggesting that prolonged exposure to negative news can permanently damage a person's mindset, which is why he deliberately stays 'aloof' from diving deep into bad news.25:00
Episode 59 — What is Model Context Protocol? · 23observationBoth Martin and Rasmus use gym attendance as a direct personal barometer of wellbeing and quality of life.1:31factAnthropic launched Model Context Protocol (MCP), a new protocol for connecting large language models to data sources.1:45thesisMCP liberates developers from waiting for service providers to build AI-suitable APIs.4:12factMCP liberates developers from relying on service providers to build LLM-specific APIs by providing middleware that connects data sources directly to language models.4:12observationThe timing of MCP's release coincides with growing recognition that pre-integrating data access is more efficient than having Claude use computer vision to parse screenshots.11:55thesisAnthropic separates concerns effectively: computer use for interacting with frontends, MCP for accessing backends directly.12:35factMCP (Model Context Protocol) connects language models to backend data sources and APIs, while Anthropic's Computer Use initiative enables models to interact with frontend user interfaces; they serve complementary roles in the broader AI infrastructure.12:35observationThe initial user setup for MCP requires manually creating JSON files with cryptic strings copied from GitHub, a barrier to mainstream consumer adoption.15:34factMartin notes that MCP's current onboarding process requires technical steps (editing JSON files, restarting apps) that are not user-friendly for mainstream consumers.15:59observationMCP creates a new startup niche: building 'fat middleware' layers between AIs and APIs with value-add features like vectorization and semantic indexing.20:43factMultiply is prioritizing MCP adoption because customers demand the ability to connect and access all their data; Multiply's existing vectorization and agent search capabilities align well with MCP's data-connection architecture.21:32factMultiply can serve as both an MCP client (pulling data from external sources) and an MCP server (allowing other LLMs to access Multiply's data through MCP).23:10thesisAdopt new protocols when organizational need is urgent, not out of curiosity or strategic betting.29:04factMartin advises adopting new technologies only when there is strong customer need or internal pressure, rather than out of strategic speculation.29:29observationAnthropic's launch blog post instructed users to ask Claude Sonnet to build MCP servers for them rather than providing step-by-step documentation.30:27factAnthropic recommended in their MCP launch blog that developers should ask Claude to build MCP server connectors rather than building them manually.30:27observationDespite technical solutions existing, ChatGPT has not achieved mainstream integration with Google Drive, suggesting adoption resistance beyond engineering.31:27thesisHuman resistance to sharing personal data may be a larger barrier to AI-data integration than technical limitations.31:27factDespite technical capability, AI integration with personal user data faces adoption barriers beyond technology, likely due to privacy and trust concerns.31:27observationThe developer community perceives Anthropic more favorably than OpenAI, viewing them as underdogs focused on alignment rather than a closing, increasingly corporate entity.32:13factMartin observes that Anthropic is winning developer sentiment through open-source infrastructure and developer-first approach, contrasting with OpenAI's closed approach.32:13thesisAnthropic has built a more consistent and coherent brand than OpenAI.33:04factMartin Källström is Chief Product Officer of Multiply and co-hosts the Co-creating with AI podcast with Rasmus (CEO).33:15
Episode 60 — 2024 AI Milestones and What to Expect in 2025 · 47factMartin co-hosts the Co-creating with AI podcast with Rasmus Adler Wahlberg, discussing AI trends, strategy, and co-creation concepts.0:00factRasmus Adler Wahlberg is expecting a daughter in the immediate future.0:26observationRasmus is expecting a daughter within days, yet is deeply focused on wrapping up year-end AI predictions.0:26observationSnow returning to Sweden is celebrated as making the landscape beautiful and preferable to 'rainy, dark Sweden'.0:48factMartin believes we are witnessing the end of the 'god model' concept—the idea that one model can be best at everything—with models now showing differentiation in capabilities across image generation, video generation, and coding.1:54factMartin observes that AI models have differentiated and caught up across domains, with each model excelling at different tasks rather than one dominant model being best at everything.2:47factAnthropic chose not to release a new Opus model in 2024, maintaining the previous generation while releasing Sonnet improvements.3:45observationAnthropic failed to release a new Opus model in 2024, only releasing Sonnet despite having three tiers before.3:45factAnthropic's leadership believes current evaluation methods lack sufficient sensitivity to measure actual model intelligence improvements.4:35observationAnthropic's CEO claims that model capability curves aren't actually flattening—human evaluation benchmarks are just too crude to measure the real improvements.4:35thesisModel development is showing signs of flattening returns on investment as no major new models (GPT-5, Opus) have emerged.4:35factGrok 3 has been trained on a 100,000 GPU cluster, giving xAI a potential data advantage for model training.5:12observationGrok has trained on a 100,000 GPU cluster and has unique access to X (Twitter) data, giving it a potential advantage competitors like Google may lack.5:12factX (Twitter) has strategic advantages in AI training data because Google and other competitors face restrictions on using YouTube, while X retains direct access to its platform data.5:37factMartin identifies AI-assisted coding as a major development of 2024, creating a new workflow paradigm for developers who adopt it.6:10thesisAI-assisted coding has become a genuine new workflow, representing the year's most significant development for practical AI application.6:10thesisThe definition of 'agentic AI' has fundamentally changed from multi-agent collaboration frameworks to single models with reliable tool use.6:36factMartin redefines what 'agentic AI' means in 2024: a single model that iteratively uses tools over multiple steps, rather than multiple agents collaborating together as was envisioned a year earlier.7:01observationMulti-agent collaboration frameworks that were heavily hyped a year ago have completely vanished from the conversation.7:01factRAG-as-a-service providers have removed their free tiers and shifted to enterprise-only pricing models because they cannot profitably support self-service users.13:10observationRAG services that launched with free tiers have all removed them and moved to enterprise-only models.13:10observationOnly Claude has successfully released general computer-use capabilities; competitors haven't matched it.15:48factMartin spent significant time in 2024 working on voice AI and sees it as emerging technology that will gain mainstream adoption in 2025.17:44factElevenLabs released a conversational API for voice AI that provides similar capabilities to OpenAI's Advanced Voice API at a lower cost.18:09factMartin predicts voice interfaces will become mainstream accessible in 2025, with broader adoption across AI platforms beyond OpenAI.18:09observationOpenAI's advanced voice API exists but is prohibitively expensive and therefore almost no one is using it.18:09thesisVoice AI interfaces will undergo mainstream adoption in 2025 as APIs become available and costs decrease.18:09factWaveform AI is a newly launched voice AI startup that represents continued momentum in voice interface development.18:34factMartin identifies OpenAI's voice application as the primary voice AI interface he regularly uses, noting limited adoption in other AI services.18:59observationMartin still uses only OpenAI's voice interface regularly despite ecosystem expanding rapidly.18:59factMultiply already has customers requesting voice AI integration features, indicating market demand beyond just platform development.19:28factCustomer service chatbots using AI agents represent the earliest significant use case for agentic AI technology in the market.19:54thesisAgents will increasingly operate autonomously while humans remain in the loop, selecting which tasks and decisions require human input rather than driving every step.22:42factRasmus predicts multi-agent collaboration will become a major focus and value driver in AI systems during 2025.25:52factMartin observes that intelligence is task-specific rather than general, with specialized agents yielding better results than single general-purpose agents.26:43factMartin forecasts that inference costs will drop by a factor of 10 in 2025 while maintaining the same level of model intelligence.28:10thesisInference cost will drop by a factor of 10 in the next year, enabling much more complex agent operations without new model improvements.28:10factCost reduction (10x) in inference pricing will enable tree-of-thoughts reasoning with multiple exploration branches, allowing agents to try different tool combinations before presenting results.28:36factMartin sees cost reductions enabling branching tree-of-thoughts reasoning with tool use across multiple paths, allowing agents to explore different tool combinations before presenting results.29:01observationMartin anticipates that tree-of-thought branching with tool use in different branches would be valuable if it existed.29:01factGroq (with Q) is currently 10 times faster than all competing inference solutions, creating a significant performance advantage.29:41observationGroq remains 10x faster than all other inference providers by end of 2024.29:41factMartin identifies a 2024 UI paradigm shift where document workspaces appear alongside chat interfaces, exemplified by Anthropic, v0, Cursor, Bolt, and Replit agents.31:19factMartin predicts new AI interface paradigms in 2025, including Trello-like task boards and Gantt charts operated by AI agents.32:10thesisNew user interface paradigms combining chat, documents, and workflow visualization tools are emerging and will proliferate.32:10factMultiply plans to implement Kanban board, table, and canvas interfaces for AI agents in 2025.32:44factMartin and his co-host plan to take a holiday break and resume the Co-creating with AI podcast in late January or early February.33:27