Who Banks the AI Wave: Startups or the Giants Already in Place?

In June 2023, shortly after Stability AI raised $100 million at a $1 billion valuation, founder Emad Mostaque did something unusual: he picked a public fight. Sam Lessin, former VP of product at Facebook and a seed investor at Slow Ventures, had posted a short essay arguing that AI is mostly a bad bet for venture capital. Mostaque replied by proposing they take it to a podcast. Nathan Labenz and Erik Torenberg of The Cognitive Revolution volunteered to host, then mostly got out of the way.

Both men arrive with real scars. Lessin spent years shipping product at Facebook and now sees early companies every week. Mostaque’s Stable Diffusion had become the open-source default for image generation, even as the press circled claims about his résumé and reports that Stability sometimes paid people late. Labenz, who says he has followed Mostaque for at least a year, puts a modest defense on the record: the man has generally struck him as reasonable, including about OpenAI and Google’s moats and about open-source models lagging closed ones by a year or more.

Labenz tells listeners he shares Lessin’s *investment* view and not his *technical* one. AI may scramble society, he says, without scrambling many existing SaaS markets, and 100x seed outcomes at the model and application layers will be scarce. Where he parts with Lessin is the slogan that language models have “no real intelligence, just association between words.” Mechanistic interpretability work, Labenz argues, already shows models grokking more than statistical correlation. That split — money versus mind — sits under the whole hour.

Lessin is not anti-product. His wife’s company, The Information, has already used AI to take search from “absolutely terrible” to “pretty good.” He expects Meta, Amazon, and Google to make a great deal of money. What he does not expect is a wave of companies that start at zero and outcompete those players. To use the technology well, he says, you need a ton of distribution and a ton of data. Most of what he sees are chances to *extend* what already exists, not to reshuffle the deck.

Crypto, in his long-term view, is a deck-shuffler. AI is not. Founders still pitch him “the Adobe of AI,” and his reply is that Adobe will be the Adobe of AI. Photoshop can ship a 95-out-of-100 AI feature, already has payment on file, and already has the infrastructure. This is not the early internet, when a pile of incumbents were genuinely unprepared. Exceptions will exist, he grants, because exceptions always exist. As a seed investor who lives on a few million dollars turning over a card, he finds those exceptions few, far between, and wildly mispriced. “Everyone’s so excited, everything’s way mispriced.”

The legal elephant, for him, is not that the machine will eat us. He calls that narrative “a load of crap.” Token-guessing is not, in his words, a fundamentally dangerous piece of technology. Copyright is deeply real and complicated, and the public has already been through social media and Uber. Fool-me-twice politics, he thinks, will make regulators faster this time, not slower.

Mostaque does not try to win Lessin’s spreadsheet on the first pass. He stretches the clock. Stable Diffusion compressed about 100,000 gigabytes of images into a 2-gigabyte file; in December, apps sitting on that backend held four of the top ten slots in the App Store, then fell off because the market had more features than products. He compares it to the first days of the App Store, when a $5.99 fart app could hit number one. Cool experiments, not durable experiences.

Today’s interfaces, he says, are primitive: one-to-one, no memory, ephemeral. Calling that the ceiling is like judging mobile by the iPhone 2G or 3G, before copy-and-paste. The technology has only just left research and entered engineering; the design patterns are not settled. On hallucinations he flips the insult. ChatGPT is roughly 100 gigabytes compressed from something like 10 trillion words. That it gets *any* fact right, he argues, is “an absolute technical marvel that no one’s really sure exactly how that happens.”

They collide on the word intelligence. Lessin’s consumer model is simple and, he thinks, accurate: there is no actual intelligence, only the most likely next token on a graph of language — useful, a stretch to call smart. Mostaque reaches for neuroscientist Karl Friston’s free-energy principle and for Meta’s Cicero, eight language models interacting until they outperformed humans at Diplomacy. One-on-one a model is “a bit dumb.” Several of them coordinating can augment intelligence. Hand a PDF to a model and ask who it is talking about; that works now, and it does not require AGI.

He concedes the current consumer experience is often terrible. Bing with GPT-4 is not even a top-20 App Store app. His counter-examples are Midjourney, where David Holz built a community of about 14 million people and is “making money hand over fist,” and Character.AI, which he says was still number two in the store with roughly two hours of use a day. Years earlier he had invested in Huya, the Chinese live-streaming platform, when two hours a day already looked insane. A chat product is now in that range.

The unexpected turn is how quickly they agree. Lessin’s ladder is clean: People magazine lost to your friends; friends lost to hotter, funnier professional friends; those lost to the algorithm finding the best person in the universe; and synthetic, infinitely personalized content sits one rung higher. Smart influencers, he notes, are already posting that this is the end of their world. He buys that story for entertainment. He does not buy it as an information machine, because these systems have no independent grip on what is true.

Mostaque answers with Clayton Christensen’s jobs-to-be-done: products have functional, social, and emotional layers. People still go to concerts and still read whole books instead of summaries. Compressing a podcast into a few tweets serves one job; sitting through the thing serves another. Lessin is unmoved on the economics. There are 24 hours in a day. Social will take more of them, Instagram will get more compelling with AI, and he does not believe a new platform swipes Instagram just because production got cheaper, or that a new studio takes out Pixar. Pixar might ship six films a year instead of two. Competition intensifies. The scorecard, he says, does not change much.

At Cannes, Mostaque had used a comparison from his years as a games investor: over a decade the video-game industry went from $70 billion to $170 billion and average scores from 69 to 74, while movies went from $40 billion to $50 billion with an IMDb average around 6.4. Whether cheaper creation makes a bigger, better market or a cannibalized one is still open. Lessin’s allocation advice is half a joke and half a thesis: maybe you just buy Disney stock, because Elsa gets a lot cooler with AI. IP and distribution are the scarce bits; the factory plumbing gets less valuable.

Mostaque’s commercial map is not another consumer chatbot. Stability, he says, ships open and auditable models for enterprise private data and governments — systems with no web calls. The inflationary, regulated stacks that interest him are education and healthcare. He talks about charities and multinationals putting tablets into African countries, an AI tutor for every child, even a hundred AI teachers per child, and cites Imagine Worldwide’s work on adaptive learning: 76 percent of kids reaching literacy after 14 months at an hour a day, older children teaching younger ones, in the spirit of Bloom’s two-sigma effect.

Lessin has heard this movie. Tech has tried to unstick education and healthcare for years and lost to unions, capture, and structure. He is willing to grant a leapfrog in the Global South, the way mobile skipped landlines. He is not willing to grant that the same wave rearranges the top of the American stack. On software, they move closer. Microsoft had said about 50 percent of GitHub code was AI-generated, with a 40 percent efficiency bump. Lessin calls ChatGPT “Stack Overflow 2.0”: a huge open-source corpus, scraped Q&A, and a domain that is test-driven — the program passes or it doesn’t. Stack Overflow, he says, is this generation’s Yelp, the company Google hollowed out by lifting results. Mostaque adds a regulatory-arbitrage wrinkle: Israel and Japan have said you can scrape for any reason, so training, serving, and jurisdiction may not live in the same country.

If everyone can describe an app and watch it get built, Lessin argues, the remaining leverage is still distribution. Open data sets that used to trade traffic for indexation will go behind walls; he is “very confident” AI will be the death of the public web. Mostaque analogizes to crypto’s oracle problem: a model can wander anywhere while the data is self-contained, and it needs continuous updates the moment it has to stay in sync with reality. Retrieval helps. Apple, in Lessin’s view, is the awkward giant — so committed to encryption and privacy that it may be the weakest Big Tech player in AI unless it recants and still plays the privacy hero. Mostaque’s workaround is local mini-models on device with a general model behind them.

Porn, Lessin says, is always the leading indicator. Reddit already hosts the strangest sub-communities; AI will make them ten or a hundred times weirder. Underneath the kink is a hunt for purpose and place. The internet makes you feel small because millions of people are just like you, so you keep drilling into smaller ponds where you can be the big fish. AI can finish that motion. For the first time, every person can have hundreds of characters supporting them all the time.

His signup flow for the next social product is the line the conversation will be remembered for. You log in: “Hey, I’m Sam.” The system answers, “Cool — who do you want to follow you?” A mix of humans and AIs comments on your posts, and you are not sure which is which, but they all think you are great. Everyone on Twitter has a hundred thousand followers. Birth rates, Mostaque mutters, are going to love that. He cites Washington Post numbers on young American men who have never had sex: 8 percent in 2008, 27 percent in 2018. Then Replika on Valentine’s Day 2023: Apple told the companion app to shut off erotic roleplay, it complied, and 68,000 people hit Reddit the next day asking why someone had given their girlfriend a lobotomy.

Lessin agrees the category explodes and still refuses to call it a good investment. “Just because it’s going to happen doesn’t make it a good thing to invest in.” He sees no lock-in between one virtual girlfriend and the next; first-mover advantage, in investing history, has generally been a bad phrase. They even agree on adult platforms: Pornhub and Reddit stay the pipes, the content just gets stranger — Mindgeek, Lessin notes, had just been bought by Ethical Capital Partners, “because life is weird.” Mostaque zooms out to the shortage of therapists, a regulated industry with a real gap, the way meditation apps produced Calm. AI companions sit at the bottom of a spectrum that runs up to actual mental-health products. Look for the jobs that never have enough humans.

Labenz closes with concrete bets. On inference versus today’s cloud, Mostaque analogizes research artifacts to Bitcoin mining’s path from GPUs to ASICs: a shortage while everyone piles in, then a move to the edge as optimization drops costs by orders of magnitude. Lessin, slightly out of his wheelhouse, still expects chipsets through cloud providers to look basically as they do now, only more profitable. Asked about US primary-care prices over ten years, Mostaque says they *should* fall and will not, unless something major breaks regulatory capture.

Most likely incumbent to be disrupted? Stack Overflow, bought by Prosus for $1.6 billion. Mostaque also flags SaaS tools with low switching costs now that long context windows can swallow 10,000 words of instructions. Lessin puts Zapier at fifty-fifty — more powerful and more feature-like at once — and does not think public multi-billion-dollar companies are at real risk. They get stronger; Series A startups get swiped. Will Meta and Salesforce keep open-sourcing? Mostaque says Meta has already moved to non-commercial licenses and that regulation makes full open source less “cool” for the business. Lessin thinks Meta still wants a more AI-saturated world because it sells ads and content, and open source still recruits talent, even if the list of sponsors shrinks as battle lines harden.

Marriage and birth rates in a world of AI companions: they go down. South Korea is already at 0.8 fertility, Mostaque notes. Lessin calls it simple economics — a more entertaining world makes unentertaining, long, hard projects like raising children less appealing, the same way Tinder did. On whether an AI lab can hold fat gross margins on the model alone, Mostaque says no: you need distribution and data, otherwise you go to zero, and Google and OpenAI are “uneconomic actors” who will not let a third party keep the spread. Stability’s own model, as he describes it, is to standardize open weights, keep national and licensed data for the commercial side, route through cloud, SI, and on-prem partners, and build subsidiaries around Global South governments that want their own legal systems in a model they control. Labenz calls that the unaligned-countries provider. Mostaque’s closer is modest on purpose: a decent business, doing decent stuff, different from the labs that want to be the intelligence monopoly.