Co-creating with AI
Co-creating with AI is Martin's podcast, produced through Multiply and co-hosted with co-founder Rasmus Adler Wahlberg, running weekly from January 2023 through late 2024 across roughly 60 episodes. It premiered under the working title "CoCast" (▶ 0:01) before being renamed specifically to foreground its organizing idea — human–AI partnership — from the show's inception (▶ 0:00). Martin has said the whole project grew out of one core belief: that co-creation with AI, not command-and-response prompting, should become the default mode of working with these systems (▶ 21:24) — and the show became a running, dated commentary on that belief as the industry (and Multiply itself) lurched through GPT-4 Turbo, the rise of agents, and the first wave of AI coding tools.
Premise: a podcast co-created, not broadcast
The founding episode is a small meta-exercise: Martin and Rasmus set out to make a show about co-creation by co-creating the show's own format live on air, deliberately uncertain of the outcome. Their explicit framing question was whether a podcast could be genuinely participatory — listeners contributing, not just consuming — rather than one-way broadcast (▶ 1:32). Martin's stated approach was to "plan without planning" — enough intention and preparation to show up prepared, but with real room left for emergence (▶ 15:08), so that the unexpected could surprise hosts and listeners in equal measure (▶ 15:33). Rasmus's version of the same ambition was structural: episodes should feed listener perspectives directly into their creation rather than arrive pre-packaged, moving podcasting away from a mainstream-media broadcast model toward something closer to participatory media (▶ 16:52; ▶ 17:17). In the first episode Martin also thanks Rasmus specifically for the initiative and energy that turned the idea into an actual recording — itself a small demonstration of the co-creative dynamic the show would spend two years examining (▶ 19:02).
A podcast about co-creation, made through co-creation. That founding episode frames Martin's creative philosophy in miniature: the show enacts the principle it discusses, treating co-creation as a lived practice rather than an abstract theme — professional, domestic, and creative all at once (▶ 0:07).
The core thesis: co-creation as the lens on the AI revolution
Martin repeatedly returns to a single claim across the run: that "co-creation" is the concept for understanding what AI is doing to society, not just a feature of one product (▶ 3:12). Two distinctions recur as its practical unpacking:
- Assistants vs. agents. Rasmus draws a sharp line between AI "assistants" — tools that execute one-off instructions, where the human stays a micromanager — and true "agents," which learn from ongoing feedback and synthesize a role description, objectives, tools and process over time, so the initial prompt doesn't need to be perfect (▶ 15:32). He grounds this in a concrete case: onboarding a new community manager at Multiply, whose autonomy grew through iterative back-and-forth rather than a fixed script — see The Community Manager's Growing Autonomy below.
- Single-player vs. multiplayer AI. Martin's parallel distinction is that existing AI tools are "single-player" — built for one person working alone — while Multiply's bet is on "multiplayer" AI that assists collective collaboration and productivity, not just individual output (▶ 15:08). He treats collaborative, multi-person co-creation with AI as one of the directions closest to his own interest (▶ 29:08), and elsewhere frames co-creation itself as the mechanism for realizing "unlimited human potential" — individual and collective at once, through networks and relationships (▶ 2:23; ▶ 2:49).
Practically, real co-creation requires AI-native interfaces where both human and AI have equal access to the same tools (▶ 2:49) — because prompting in natural language alone is too imprecise, Martin notes that working prompts increasingly contain pseudocode for its semantic density (▶ 14:53) — while the design goal is the opposite: users shouldn't have to become prompt engineers just to co-create (▶ 15:19). The idealized end state is an agent that asks permission to explore a risky creative direction rather than waiting passively for instructions (▶ 4:28), with trust between human and AI evolving gradually to determine who handles what autonomously (▶ 13:07). This philosophy is developed further in AI co-creation and AI-native interfaces & products, both drawn substantially from this show.
The February 2023 vision dump: agentic AI, ubiquitous computing, and its dark side
One early episode (ep5, "As AI Enters Our Lives") is an unusually dense, freewheeling brainstorm — Martin and Rasmus running through implications of agentic AI months before "agents" became an industry buzzword. Worth reading as a single arc rather than isolated bullets:
- Defining agency. They define agentic AI simply as AI with agency — systems that can go and do things on a user's behalf, not just answer (▶ 1:46), citing Adept's ACT (Action Transformer) model as the leading edge — including a demo of Adept autonomously finding a suitable house in Houston by operating web applications itself (▶ 1:13; ▶ 1:46). Rasmus's read: if Adept succeeds at teaching AI to operate software generally, AI could eventually do anything a human does in front of a computer (▶ 2:42).
- The web as a language interface. Martin frames the web itself as fundamentally a text interface to human society, and sees the core AI shift as giving machines native access to natural language and semantic understanding (▶ 3:10; ▶ 4:01) — leading toward ubiquitous computing, where agentic, language-capable AI becomes an assistant to the web itself, eventually removing the need for screens entirely as people simply talk to the things around them (▶ 4:01; ▶ 4:27).
- The darker mirror. The same capability, scaled, worries them: AI-calculated behavioral patterns could be used to control people at scale, and Rasmus frames agentic AI as "the ultimate bot problem" — systems that can impersonate and generate misinformation across thousands of fabricated identities with distinct personalities via API access (▶ 5:37; ▶ 6:13; ▶ 6:39). Martin's counterpoint is not reassurance so much as continuity — humans already control each other through technology, so AI mainly amplifies existing dynamics rather than inventing new ones (▶ 5:17).
- Deficient technology. Martin's recurring frame for "skills AI makes obsolete": phone-number memorization or a sense of direction aren't fundamental human capacities worth preserving — they're artifacts of deficient technology (bad interfaces, no GPS), and losing them to a better tool is neutral or good, freeing attention for human interaction and creation instead (▶ 14:36; ▶ 17:34). He illustrates it with road networks and turn-by-turn navigation directly: Google Maps augments cognition rather than "amputating" it, since it supplies real-time traffic knowledge no human could hold unaided (▶ 15:28; ▶ 15:02) — though Rasmus pushes back with his own experience that leaning on Maps has visibly eroded his sense of direction while driving (▶ 12:35).
- Personal AI, mirrored back. They sketch a "virtual coworker" AI (Martin nicknames it "Steve") that grows alongside a user over time (▶ 23:49), including using AI to identify areas for personal development and mirror a user's own negative behaviors — lateness, sloppy grammar — back at them as a behavior-change tool (▶ 24:14; ▶ 24:40). Rasmus extends the idea to AI avatars that act and even decide on a person's behalf — planning social engagements without the user knowing the details — and to trusting AI with autonomous driving and other high-agency domains (▶ 26:26; ▶ 26:51).
- Generative entertainment. Martin imagines personalized VR entertainment with AI as a real-time "game master" generating immersive worlds and switching themes without engine constraints, plus AI-assisted "autocomplete" for video generation (▶ 18:15; ▶ 19:31); Rasmus counters that active, creative play is inherently richer than passive consumption, and that easily-created co-creative experiences could compete with passive screen time on those terms (▶ 20:55; ▶ 21:20).
- Surveillance, scaled up. The same monitoring logic that could track individual wellbeing on a team, Martin notes, extends unsettlingly to society-wide data use for steering collective goals — a thought he and Rasmus leave open rather than resolve (▶ 28:28; ▶ 29:25).
Startups, no-code, and the shape of AI-era work
A later episode with guest Rasmus [Oller] — a Klarna product manager of eight years, initially skeptical of AI hype until ChatGPT's November 2022 release (▶ 1:33) — turns into an extended, opinionated discussion of how AI changes company-building:
- The three-buckets framework. For most companies, AI implementation breaks into three buckets: augmenting individual staff productivity, automating recurring processes where bots now often outperform humans on consistency, and building genuinely AI-enriched products (▶ 4:56).
- No-code lowers the floor, not the ceiling. No-code tools (Make.com, Zapier) combined with multimodal LLMs put what used to be "hardcore engineering work" within reach of non-engineers — visible in the rise of automation agencies built entirely on these tools (▶ 12:33). Martin separately names training a custom GPT on how you work and wiring it into Zapier as a generalizable, powerful recipe for AI-native workflows without deep engineering (▶ 12:16).
- What a founding team still needs. Guest Rasmus argues the one non-negotiable capability for a software startup is backend engineering — frontend can largely be handled by no-code and AI-assisted methods — with sales remaining the real gate on market validation (▶ 32:19).
- Co-location beats tooling. Martin's contrarian claim, against the AI-productivity narrative of the same conversation: the single most important factor in getting a startup off the ground is still putting people in a room together and focusing — more important than whether today's tools deliver 2x or 10x productivity over his first company in 2010 — because co-location produces accountability and a shared, undistracted sense of truth about what's actually being built (▶ 38:24; ▶ 39:48).
- Risk appetite scales with your customer base. Guest Rasmus's operating rule from Klarna: risk tolerance for shipping AI features should be relative to what's at stake — five customers can absorb more risk than five million, which need automated testing, review processes and gradual rollout (▶ 45:20). He pairs this with a caution against compounding narrow A/B-test wins into a "Frankenstein interface" that loses the strategic product vision underneath it (▶ 46:52).
- What doesn't automate. Guest Rasmus's closing argument: as routine work gets automated, human-centric skills — sales, relationship-building, an old-fashioned phone call — become more valuable, since human attention becomes the actual limiting factor in an information-saturated market (▶ 35:32).
- Martin also argues, elsewhere, that AI-enabled speed doesn't necessarily mean shrinking headcount: in a globalized market that's functionally infinite, companies can instead move faster and serve a much larger market with the same or similar team (▶ 27:11) — and that GPT-4, structured with the right process and accountability, already reasons about as well as many people on an average office day, cutting against a binary "AI isn't really intelligent" framing (▶ 14:29).
Industry moments the show tracked in real time
Because it ran weekly across two turbulent years, the podcast doubles as a dated log of the AI industry's inflection points, several covered live:
- GPT-4 Turbo and OpenAI's first Dev Day (Nov 2023) — see OpenAI Dev Day aftermath below; Martin separately argues in the same window that voice and speech are what make AI a true workplace participant rather than a screen-bound chatbot — necessary for it to sit in meetings and real-time conversation (▶ 24:07) — and frames AI more broadly as a new workforce inside companies, the framing he says the show's whole mission aligns with (▶ 11:26).
- The GPT Store and GPT Team launch (Jan 2024) — OpenAI opened shared organizational workspaces for building and using custom GPTs alongside the public GPT Store, while deferring creator revenue-sharing to Q1 2024 and initially to US creators only, citing legal and cross-border payment complexity (▶ 3:39; ▶ 19:55). The same episode notes early, unconfirmed signs (leaked screenshots, A/B tests) of ChatGPT gaining persistent cross-conversation memory alongside a temporary/incognito chat mode (▶ 18:34).
- The Super Bowl goes AI-mainstream (Feb 2024) — 2024's Super Bowl ad slate featured Microsoft Copilot and Google's Pixel 8 "smart" camera features prominently, alongside mainstream brands using AI to generate recipes, read as a signal of AI crossing into consumer mainstream marketing (▶ 1:16).
- The AI coding-assistant war (Dec 2024) — active competitive escalation among coding tools, with Windsurf shipping its own VS Code fork and Cursor and Aider introducing "architect" modes, tracked as a defining feature of the year's developer tooling (▶ 1:39).
- 2024 in review, 2025 forecast (Dec 2024, season-closing episode) — Martin's read on 2024: models across vendors had visibly differentiated and caught up with one another rather than one lab staying years ahead (▶ 2:47), and AI-assisted coding matured into a genuinely new developer workflow (▶ 6:10) — alongside voice AI, which Martin had spent much of the prior winter building toward, finally reaching real traction, with Multiply customers already asking for it directly (▶ 17:44; ▶ 19:28). For 2025, Rasmus bet on multi-agent collaboration as a major value driver (▶ 25:52), Multiply planned Kanban, table and canvas interfaces for its agents (▶ 32:44), and a hoped-for ~10x drop in inference cost was pitched as the unlock for genuine "tree of thoughts" reasoning — agents branching across multiple approaches before presenting a result, rather than following one linear chain (▶ 28:36).
- Multiply's own early technical bets, previewed on the show: ChatGPT plugins integrated as an "ultimate API connector" to link Multiply to outside services (▶ 5:22); Perplexity and Metaphor's web-scale vector search evaluated as an alternative to routing search queries through GPT-4 directly (▶ 5:40); pre-computed Wikipedia embeddings as an off-the-shelf vector database for grounding answers (▶ 2:27); and Martin's own hands-on work with the Sentence Transformers library to place multilingual text in a shared embedding space (▶ 0:17). Full platform detail lives on the Multiply page.
Guests
Most episodes are Martin and Rasmus alone, but recurring guest slots brought in outside expertise: Malcolm Sparks, founder/CTO of Juxt and creator of XTDB — the database underlying Multiply — on data flexibility and graph relationships (ep17); Oskar Beijbom on practical AI deployment (ep38); Johan Salo on AI, design and strategy (ep35); and a Klarna product manager also named Rasmus (surname Oller) on startups and no-code, distinct from co-host Rasmus Adler Wahlberg (ep30).
Notable episode narratives
OpenAI Dev Day aftermath: the industry awakening. Recorded the morning after OpenAI's first Dev Day, Rasmus is visibly hyper about the announcements while he and Martin work through what changed: GPT-4 Turbo's 128K context window, a two-thirds price cut, and new capabilities — vision, function calling, code interpreter, the Assistants API. The conversation lands on a real pivot point: the entire industry being reset overnight, forcing every company to choose between seeing opportunity or existential threat. "This is a big divisor of the entire world — either you're 'holy fuck, I'm going to do so much more now,' or you're 'holy fuck, my business is in ruins.'" (▶ 1:12)
From API Constraints to Self-Hosted Models. Years into using commercial AI APIs — fast, production-ready, his default path — Martin hits a wall building a conversational AI system: he needs millisecond-level control over conversation flow, and chaining six or seven API calls together makes cost prohibitive. That pushes him to rent GPUs and self-host models, discovering along the way that flexibility, not cost, is the real driver: APIs cap how granularly you can shape data in and out, in ways self-hosting doesn't. "It was the flexibility, because you run into limitations with the API where you actually can't influence how they receive data or send data to the granularity that I need." (▶ 6:36)
The Marketing Agency's AI Celebrity Search. A marketing agency's celebrity brand ambassador drops out days before a campaign, and they need a fast, nuanced replacement — similar appeal, not A-list budget, an existing online following, matching audience overlap. Multiply's research agent returns ten candidates from just a couple of prompts, three of them genuinely strong — a task that would normally demand significant expertise and time, compressed to minutes. "They got, basically, out of like 10 suggestions just from the first couple of prompts — there was like 3 really good ones." (▶ 20:27)
The Community Manager's Growing Autonomy. Rasmus describes onboarding a new community manager at Multiply, starting from strict process documentation for the podcast workflow and weekly product updates. Rather than staying scripted, the manager's autonomy and initiative grew through natural iteration — his questions decreased not because he'd become a robot, but because shared understanding deepened; he began taking on work unprompted, the way a trusted colleague does, and collaboration increased rather than shrank. It becomes the template Martin and Rasmus use to imagine AI agency itself: specify role and context, iterate on feedback, let autonomy and partnership co-evolve. "I can already feel his autonomy increasing... it doesn't necessarily mean we collaborate less. It's like working with another person." (▶ 13:56)
A podcast about co-creation, made through co-creation. See Premise above — the founding episode's meta-structure is itself one of the show's clearest stories. (▶ 0:07)
Worth remembering
- Multiply's platform demo on the show — the Podcast Pro app generating episode descriptions, Stable Diffusion images, and social copy across four platforms from one transcript in a single pass — was an early, concrete proof of multimodal, multi-model orchestration, months before "agentic AI" was a common term (▶ 19:24).
- Martin's most speculative aside from the ep5 brainstorm has aged unusually well: an AI that notices itself doing repetitive work and decides, unprompted, to build another app to automate it — a small autonomous, self-optimizing loop he flagged as "something really exponential" in mid-2023, well before autonomous coding agents existed (▶ 29:53).
- Martin also floated AI as a mediator for human misunderstanding — detecting when two collaborators have talked past each other and helping clear it up — a use case distinct from the show's usual productivity framing (▶ 12:24).
Related: Multiply · AI co-creation · AI-native interfaces & products