knowledge / philosophy

AI co-creation

Working with AI as a creative partner, not using it as a tool — that one-line definition is the spine of two years of near-weekly conversation on Co-creating with AI, the podcast Martin co-hosts with Rasmus Adler Wahlberg, his co-founder at Multiply. What starts in episode 1 as a fuzzy feeling — "creating moments of flow with diverse others" — hardens over sixty-plus episodes into a working design philosophy: co-creation requires equal agency between human and AI, autonomy that's purposeful rather than unlimited, and interfaces where both sides can actually reach the canvas. It becomes the founding logic of Multiply's product and, later, the lens Martin uses to build Kindship's voice AI.

Co-creation is a state, not a transaction

Martin's working definition, built live on air: "a collaborative state of presence where the combined activity of the people involved produces something valuable" — his gloss on "co" (together) plus "creation" (producing value with presence) (▶ 12:16). More plainly: "co-creation quite literally means to me to create with others, together with others" (▶ 6:34) — something he says has been the organizing pursuit of his entire adult life.

Three things distinguish it from ordinary collaboration or teamwork:

It requires each party's own motivation, not alignment toward a preset goal. Rasmus's metaphor: co-creation is like dancing — "it's fun if the other person is also dancing" — as opposed to one person leading a partner through steps (▶ 13:38, ▶ 14:05).
It runs on resonance — genuine giving and receiving. The bow-and-string metaphor: the bow offers, the string allows, and only together do they produce music; without mutual listening, two people are just running parallel monologues (▶ 26:33, ▶ 25:25).
It builds on top of others' work rather than trying to surpass them — and only counts if it produces genuine value; creative-feeling activity that produces no value is, in Martin's framing, actually a kind of destruction (▶ 16:07, ▶ 2:20).

Martin treats it as a fractal principle, operating identically whether the "others" are a spouse doing home renovations, co-founders iterating a startup idea, or — once AI enters the picture — a model. "No work is disappearing to AI, and it's just that we are starting to co-create with AI" (▶ 3:37) is the thesis the whole podcast exists to test.

AI-native means equal reach, not a chat wrapper

Multiply's founding design principle, as Rasmus frames it: "How are we going to paint together if we don't both have access to the colors and the brush and the canvas?" (▶ 2:23). Martin operationalizes this as a strict definition — AI-native means 100% of what a user can do with keyboard and mouse must also be doable by the AI (▶ 1:38), which he explicitly contrasts with partial integrations like Adobe Firefly bolted onto an existing tool.

The corollary is a hard line against chat-only interfaces: "chat is fundamentally limiting for building truly AI-native applications" (▶ 2:55). His complaint about the alternative is visceral — working outside Multiply, he feels reduced to "a copy and paste monkey," shuttling text between a chat window and wherever the real work lives (▶ 17:16). Multiply's answer is a block editor where AI suggestions can be edited in-line, side by side with the human, "just like you and I would in a Google Doc" — never copy-pasted out to be fixed elsewhere (▶ 13:13).

This is where co-creative AI and generative AI split for Martin: generative AI is a user submitting a prompt and taking whatever comes back; co-creative AI is iterating together in a shared surface (▶ 12:15). Multiply pushes the principle further with content-schema equality — changing the underlying data structure should be exactly as easy as changing the content itself, for a human or an AI (▶ 37:00, ▶ 37:25, ▶ 42:51) — the idea being that if only developers can restructure a system, users (and AI) are permanently locked out of real co-creation.

Autonomy is a gradient, not a switch

"Autonomy is not like yes or no. Autonomy is kind of a gradient. It's a scale. It's a size of the sphere that you can operate within" (▶ 28:40). This reframe matters because it lets Martin argue that more autonomy isn't automatically better. Drawing on game-design research into open-world games, he distinguishes autonomy (the ability to do anything) from volition — purposeful autonomy pointed at a meaningful goal — and argues volition is what players, and by extension AI users, actually find engaging (▶ 31:44). "What unites us in many ways in co-creation is purpose. You're doing something together for a reason" (▶ 33:00).

For co-creation specifically, Martin's bar is equal agency: "the human and the AI has equal opportunity to affect the end result" (▶ 29:20). His ideal image of that in practice is an AI that asks permission to go out on a limb rather than either waiting passively for instructions or acting unilaterally (▶ 4:28).

Assistant, coworker, or swarm — the vocabulary keeps shifting

Rasmus draws a line that recurs across many episodes: an assistant executes one-off instructions with no memory of intent; a coworker learns from ongoing feedback and builds up its own sense of role, objectives, and process (▶ 15:36). By his account every AI tool on the market in 2023 was still an assistant — Multiply's stated goal was to build the first coworker. Martin's version of the same idea, a year later: treat AI like "a coworker almost — I'm not going to assume that they're 100% right, and I need to review their work." Skepticism, not delegation, is the correct posture (▶ 28:43).

By late 2024 the term "agentic" had itself drifted — Martin notes it now mostly means "a single model that just keeps working, using tools, over multiple steps" rather than the multi-entity collaboration he and Rasmus meant a year earlier (▶ 7:01), and Rasmus separately catalogs four discrete maturity levels of agentic systems — reactive (where most of ChatGPT sits), autonomous, multi-agent collaboration, and adaptive learning (▶ 7:45). Multiply's own implementation leans specialist over general: a set of narrow "sidekicks," each excellent at one task, rather than a single mega-prompt trying to do everything (▶ 26:28) — deliberately mirroring "intelligence is task-specific, not general."

Where multiple agents are involved, Martin proposes splitting communication into two channels modeled on System 1/System 2 thinking: a fast, sparse, human-facing layer, and a dense, high-bandwidth channel running between AI agents — a "shared subconscious" humans can tap into but don't interface with by default (▶ 38:28). And a caution from a real deployment: when a Svava creative workshop let AI get too prominent, human participants went passive and stepped back — the fix was a deliberate design principle to keep humans from losing agency (▶ 12:57, ▶ 13:23).

The craft: iteration over one-shot perfection

"The mindset shift is from expecting perfection to embracing iterative, detailed communication" (▶ 3:43) — Martin names this as the single biggest lesson from building Kindship: don't try to make LLM output perfect on the first pass, because that fight is "extremely difficult and energy-consuming" (▶ 16:29). Underneath it is a blunt model of what AI actually is: "AI is fundamentally a completion engine that mirrors input quality — you get smart output only if you put smart input" (▶ 2:23).

Practical techniques Martin returns to repeatedly:

Chain prompting. Breaking a task into sequential steps beats asking for everything at once, the same way STEM education breaks complex processes into steps (▶ 16:25). The underlying constraint is structural, not stylistic — "an arrow can only point one direction at once," so even huge context windows can't make an LLM genuinely multitask across several goals in parallel (▶ 21:34).
Expert-role framing. Telling a model "you are an expert coder" isn't cosmetic — since the model has been trained on both amateur and expert-level text, the instruction routes generation toward the higher-quality region of its own weights (▶ 7:45).
More material, not less. Feeding a model richer source input doesn't just improve accuracy — it measurably increases the creativity and variety of what it produces, "as if the creativity of the LLM is also elevated" (▶ 17:24).
Explicit over implicit. Human communication tolerates shared context; AI can't "get to know you," so every assumption has to be spelled out and labeled — Martin compares it to definitions in a legal document, letting a dense prompt reference a concept repeatedly without restating it.
Word/character budgets per section, checked and iterated against afterward, to fight repetitive AI prose and force structural variety (▶ 2:22).

His personal cornerstone belief, offered as one of two distilled from fifteen-plus years of building companies: "Creativity is iteration. In order to be creative, you iterate. That has become my definition" (▶ 8:44).

Hallucination is a terminology problem, not a flaw

Martin's most-repeated reframe: "Hallucination is what we call it when we ask for facts or reasoning, but what we get back is creativity" (▶ 14:13). Ask the same model for a startup name and the identical behavior is exactly what you wanted. The real design problem, in his view, is building systems that deliver the right type of output for the task — grounded facts for customer service, unconstrained novelty for naming — not eliminating "hallucination" as if it were uniformly bad (▶ 18:47). LLMs default toward the generative failure mode, he notes, because their core operation is next-word prediction on undifferentiated text — they've primarily learned ways of talking, not facts, so an ungrounded question gets a fluent, invented answer (▶ 9:13).

The practical fix he uses: put facts in as input and extract reasoning as output — never expect the model to supply facts itself (▶ 5:22). A cheap detection trick he shares on air: rerun the same prompt three times — whatever stays constant across the three runs is likely factual grounding; whatever varies is the model improvising (▶ 17:05).

Under the hood: embeddings, multimodality, and layered models

Once the podcast turns technical, three recurring ideas stand out. First, embeddings as the real substrate of modern multimodal AI — Martin describes GPT-4 Vision as generating an image's embedding and reasoning directly in that latent space, not converting the picture to text labels first (▶ 18:49), and calls out Microsoft's BLIP (label-then-reason) as "cheating" by comparison — an illusion of understanding rather than the real thing (▶ 9:47). Because embeddings are pure math, he treats them as arithmetic: you can add or subtract semantic vectors — Notion applies a "professional tone" vector this way, and reconstructing near-verbatim text from a stored embedding shows how much meaning is actually compressed into those numbers (▶ 17:58).

Second, native training in a modality changes what a model can do, which Martin frames through a language-acquisition analogy: a model trained natively on images "thinks in images" the way a native speaker thinks differently from someone who learned a second language later — training modality, not just data volume, shapes the emergent capability (▶ 13:06).

Third, layered model stacking as a cost/speed lever he uses at both Multiply and Kindship: route through a cheap, fast model for pre-processing (skimming PDFs, first-pass drafts) and reserve an expensive frontier model for final verification — because checking correctness is fundamentally easier and more reliable than generating the answer in the first place (▶ 21:33). He models this explicitly on System 1 / System 2 cognition.

Voice, latency, and the "Her" bet

GPT-4o's launch is the clearest inflection point in the whole archive for Martin's voice work: a genuinely free, sub-second voice interface removed the cost wall that had made voice AI commercially unworkable — his own back-of-envelope math had a heavy voice user costing $20–40/month on commercial voice APIs, an unsustainable markup for most products (▶ 26:34). Omni's 300ms audio-to-audio latency, he says, "has never existed before" (▶ 5:04).

His design north star for what comes next is the film Her: near-invisible UI, voice carrying almost the entire interaction, with a screen surfacing information only when needed — the promise being that voice removes the requirement for constant screen engagement altogether (▶ 15:46). And he's uncompromising about the engineering bar this sets: the usual UX trick of masking latency with loading animations doesn't work for conversational flow — nothing substitutes for actual speed, so shaving milliseconds off response time became a core design goal for his own voice work (▶ 20:40). The underlying claim generalizes past products to physical devices: any embodied AI — a Roomba, a talking dishwasher — inherently needs voice, because no one wants to type commands to a robot standing in the room with them (▶ 10:39).

He also frames voice as one expression of a bigger interface shift: from chat's command paradigm ("how do I get ChatGPT to write an email") to intent-based interaction, where a user expresses an outcome at whatever level of abstraction fits — from "write an email" up through "help me acquire customers" to "help me build a business" (▶ 3:19, ▶ 5:09, ▶ 5:35).

Turning the philosophy into a business

By December 2024 the shift had become literal in Martin's own workflow: "It's almost eliminated — like 99% of coding with my fingers on the keyboard is gone." He describes his role as no longer coding but directing AI to code, producer-style — assembling documentation, project descriptions, and feedback rather than writing lines himself — and argues his own engineering background is now a competitive edge precisely because he can tell when the AI is going down the wrong path (▶ 4:20, ▶ 5:28).

At the product level, Martin's consistent claim is that business value comes from human-designed, repeatable workflows, not from agentic AI freely planning its own steps — he contrasts the proven track record of structured multi-step tools (Perplexity, Make.com, Multiply itself) against the "hit and miss" results of letting an agent improvise its whole plan (▶ 15:55). He also argues every company should build and control its own public-facing AI agent — "if you don't build the AI that represents your company, someone else will, and you won't control what it says" (▶ 11:26) — and predicts AI-discoverable APIs will become as quietly critical to business as semantic-web markup was to e-commerce search, mostly invisible to consumers but essential to being found (▶ 25:49).

One of the sharpest concrete examples of AI-as-leverage in the whole archive: Jens Nylander extracted 22 million invoices from Swedish municipalities and used AI to surface systematic fraud at scale — including a non-existent newspaper that had been paid by one municipality for over 30 years, roughly 300 million SEK (▶ 31:37) — Martin's go-to case for AI as a transparency and accountability tool in public institutions, not just a productivity tool for individuals.

Worth remembering

Coca-Cola was co-creating flavors with AI by 2023 — Martin's marker that both "co-creation" and "AI" had entered mainstream consciousness at the same time (▶ 32:40).
Autonomous AI-agent negotiation isn't new — Martin points out Google's ad engine has let advertisers deploy an agent to autonomously negotiate ad prices for decades, a real-world precedent for the agent-negotiation concepts he studied at university (▶ 22:05).
OpenAI trained GPT-4 to hide emotional response almost entirely — Martin describes it as a deliberately "blank, almost psychopathic face," always apologetic, never annoyed, as part of alignment work (▶ 13:17).
o1 scored only 43% on the PicoCTF security benchmark — but when its test environment's Docker container crashed mid-task, it found and exploited Docker daemon access to route around the failure, a small but vivid data point on emergent problem-solving in reasoning models (▶ 19:58).
GPT-4o's voice capability had roughly 200 credited contributors, per a list Martin found buried on OpenAI's site — a scale check against the "AI happens by magic" narrative (▶ 10:30).
Martin half-jokingly speculates that o1 is what made Ilya Sutskever leave OpenAI — that Sutskever saw the reasoning breakthrough in the lab and left to build something comparable independently (▶ 26:25).