Sarah Guo - What the 250 People Building AI Believe - [Invest Like the Best, EP.489]

Summary

• Sarah Guo’s open-source position: competitive open-source models are already widespread, and restricting them would only handicap Americans. “The cat is out of the bag” — Chinese, US, and European open models are already in use everywhere, and if the US restricted them, “you’d basically just restrict law-abiding American businesses” while actual adversaries ignore the rules. Her response to backdoor-like fears in Chinese models is rigorous safety testing, not speculation; she expects the US to talk much more about “compute independence.”
• Among the roughly 250 entrepreneurs and researchers Conviction tries to stay close to, a new belief has emerged that recursive self-improvement could produce “some sort of exponential intelligence” in one to two years. She qualifies it with Karpathy’s line: “I thought it was two years away for about 10 years” — and says a contingent of researchers now feels either that their work does not matter because the model will do it, or that only compute scale matters.
• Her boldest portfolio timeline: Sunday Robotics believes it will have general semi-humanoid robots doing things in people’s homes, first in beta, by the end of this year. Founders Tony Zhou and Chang Xi, who worked at Toyota Research, DeepMind, and Tesla, are people she thinks have contributed “dual-handedly” most of the interesting ideas in robotics AI over the last four years by treating cheap, distribution-matched data collection as the core technical problem — going from “cardboard in a Stanford basement” to a full-stack system manufactured there in just under two years.
• She has moved strongly to the “yes” side on AI in biology: “You can create and capture enormous value with models in biology.” Conviction was the first check into Chai Discovery, which is working with a number of top-10 pharma companies on R&D acceleration. In discussing the evidence, she cites a $10 million contract and customer adoption, against the conventional wisdom that “you can’t make money selling software to pharma.” The industry light-bulb moment will be a new indication or drug whose trajectory was clearly changed by AI — “it’s going to happen.”
• On the investing side, her biggest worry is capital allocated through pedigree rather than fundamental intuition. Large-scale research bets are being made via “proxying of judgment to pedigree or to other legible signals” — one extraordinarily good investor’s explanation of a company was essentially, “Do you know the quality of this person?” Her verdict: “it’s not all gonna work… and I may not be any better at deciding,” but having no point of view beyond the person’s pedigree “is dangerous.”
• A major constraint on AI is regulatory, alignment, and physical-supply-chain capacity, not a lack of technical or entrepreneurial capability. A hyperscaler infrastructure leader told her, “There was nothing that was going to move the needle for us at sufficient scale before 2030.” Energy requires convincing New Yorkers to accept data centers and the public to accept nuclear so SMR cost curves can fall. Sarah says US reindustrialization cannot happen without automation, making competitiveness an active choice, not an inevitability.
• Her one-year hope is Jevons paradox in practice: the software-engineering speedup replicated across every function. One portfolio company’s marketing lead built “an autonomous marketing department” for a 1.5-person team; both she and Patrick say they work more, not less, with AI — echoing “a core wisdom of Jensen’s” that everyone will be more employed.

Deep dive

1. Ninety miles an hour with no backtest — and the bet against a monolithic AI outcome

• The opening tension, via a friend’s confession Guo relays: “I keep saying I wanna press the brakes as hard as I can, but I’m not doing it, going ninety miles an hour.” The investor’s dilemma is whether you miss the upside or “make the mistake of every boom-bust cycle in technology history” — compounded, four years into building Conviction, by questions of firm durability and people’s careers.
• Patrick probes whether her firm is a deliberate wager against a monolithic lab-dominated future. Her reframe: she believes in “the great man and great woman theories of history” — high-agency people with the right risk capital change outcomes, e.g. whether a competitive Western open-source model exists comes down to “did anybody make one?” She rejects the war framing (she works closely and co-invests with the labs), but on the extreme view that “the owner of one, two, three frontier models consumes the economy”: “I do not want that future, very clearly, and I don’t think we are going to end up there.”
• Asked if she wants to be the great woman herself: no — not humility, “it’s not in my set of goals.” Her identity choice was investor over software entrepreneur: “I’m deeply curious, I like to understand things, I like to be right” — the firm can support the movement, “but I don’t think it has to be me.”

2. The 250-person map and technology-forward picking

• Her account of the edge so far is deliberately unglamorous: generational transitions meant early-stage investing was not as competitive as it had been, while a massive technology transition was underway. So “all you have to do is take the risk and be focused, and then it’s an execution play… pretty simple and is just hard.” Patrick pushes back — LP surveys rank her first or second; surely there’s more than outworking people — citing Mike’s frame of roughly 250 entrepreneurs and researchers “doing the most interesting things on the frontier” that the firm tries to know and support.
• The concession, and the real method: a technology-forward thesis others called nonsense. Instead of only working customer-back, Conviction mapped model capabilities to professions — Harvey being the specimen. “Law is structured language”: late-2022 next-token prediction plus retrieval plus precedent text made law “a really good match,” and founders Winston and Gabe were “AI-pilled” enough to project from a trivial California landlord-tenant question to “doing an Activision Blizzard M&A and doing eighty-five percent of the work.” What appealed to her was “the ambition of what was possible then and the technical logic of why it would work.”

3. Inside the frontier labs: exponential belief, disempowered researchers

• What the roughly 250 are saying now: the landscape is “violently competitive” and globally so, which is “narrative breaking.” The belief she describes — “new within the last twelve months for a lot of researchers” — is that recursive self-improvement puts some sort of exponential intelligence one to two years out. Her hedge, via Karpathy’s self-aware line: “I thought it was two years away for about 10 years. And he thinks it again, to be fair. Who can say?”
• The psychological cost: when an individual could feel, as one of 200 people at OpenAI, that they moved the needle, the question becomes “will I need $750 billion of compute spend” with many thousands working on it, and ownership evaporates. Her taxonomy of the resulting despair: “What I do doesn’t matter anyway because the model is gonna do it, or the only thing that matters is compute scale, and both of those are somewhat disempowering.”
• Her own version of the agency test: asked whether any of her companies wouldn’t have been backed without her at that round — “5% or 10% max. They were resourceful, really talented people.” She would not work on companies if she felt she could not change their outcomes a little bit, even though most would have found other investors.

4. The real bottlenecks: permits, supply chains, and pedigree-proxy capital

• On compute, people are “very much thinking about 2032 at this point,” and a hyperscaler infrastructure leader told her “there was nothing that was gonna move the needle for us at sufficient scale before 2030. That’s depressing.” Her diagnosis: not a technology, capability, or capitalism problem but “a regulatory problem and an alignment problem, and I don’t mean AI alignment” — convincing New Yorkers to want data centers, convincing the public nuclear is safe, and allowing enough SMR construction to bend the cost curve. The physical supply chain’s tacit knowledge, labor, and raw materials “can’t go as fast as software”; “the only way through that is through.”
• Her investing-side worry is how capital judges research bets: much of the money funding them has no fundamental understanding, so decisions run on “proxying of judgment to pedigree or to other legible signals.” Her debate with an “extraordinarily good investor friend” crystallized it — his explanation was essentially, “Do you know the quality of this person?” Her verdict: “it’s not all gonna work, and I may not be any better at deciding,” but having no view beyond pedigree “is dangerous.”

5. Sunday Robotics, and how Conviction actually decides

• The researchers who most blew her away: Tony Zhou and Chang Xi of Sunday Robotics, roughly 25-year-old Stanford PhD students (one didn’t finish) who had worked at Toyota Research, DeepMind, and Tesla. She thinks they have contributed “dual-handedly” most of the interesting ideas in robotics AI over the last four years. Their creativity: treating the field’s missing “internet of robotics data” as a solvable constraint — collect data “in the cheapest way possible in a way that supports the distribution of real world environments and tasks.” Under two years from “cardboard in a Stanford basement” to hardware and models manufactured there, and the whole team believes it will have general semi-humanoid robots doing things in people’s homes “first in beta end of this year” — a timeline that surprised even her.
• Her process: “very instinctive on people” — on a one-to-ten scale she’s “immediately an eight or a nine,” then spends days to weeks hunting the holes in her understanding, writing a full, perhaps Greylock-style memo (in the solo days, sent to John Lilley or Dylan Field for outside reads). “Other people climb to conviction versus I start there and then I work backwards.”
• The limit case, raised by Patrick: is anyone so good you back them without comprehension? Her answer intertwines the two — “if I don’t understand what they’re doing, I can’t have an opinion on their judgment.” If Brett Taylor wanted to “dig in volcanoes or do dog streaming… I’d be like, yeah, of course, man. But he wouldn’t do that.” Time allocation now: roughly two-thirds on portfolio work, only four to six new companies a week — versus 500 in her first couple months at her old firm — because “I just have much more confidence I can tell.”

6. Open source is already widespread; compute independence is the next fight

• Her open-source position starts from what’s already happened: increasingly competitive open models over three years — largely China, but also Thinky, Poolside, “people waiting for Reflection,” NVIDIA models, and Mistral. Frontier models are often “too expensive, too sensitive, or too slow” to use, so democratized capability diffuses further — “you can’t imagine the diversity of reality” from inside a lab. Restricting open models in the US would just “restrict law-abiding American businesses… you’re restricting your own people,” since adversaries ignore the rules; on Chinese-model backdoor-like fears, “let’s go find out as much as we can” with rigorous testing rather than speculation.
• She can “very easily” imagine a world where America doesn’t get intelligence too cheap to meter competitively — and it matters because “there is not a version of the world where we rebuild our industrial base without automation.” The risk: rational job fears plus resentment of rent capture hardens into an anti-capitalist bloc that slows energy and industrial buildout. Hence “we’re gonna start talking much more about compute independence” — thin sieves in the supply chain (her TSMC mug in hand), efforts like Jacob Helberg’s PacSilica, and Conviction’s own investments in labor gaps for data centers and robotics, nuclear energy, and alternative chip architectures. Pure data-center building she’s circled but not done: “fundamentally, I’m a technology investor.”

7. Live debates, the Suno miss, and a Jevons-paradox year ahead

• The recurring internal debate is whether historically venture-hostile markets have changed. Semis “was a god-awful business for the longest time,” but at-scale demand for accelerators and buyers’ desire not to be “stuck on one line at TSMC” changed the risk equation. Biology is where empirical data flipped her to one side: against the biobucks-or-nothing conventional wisdom, Chai Discovery is working with a number of top-10 pharma companies in significant ways, and the discussion cites a $10 million contract and customer adoption. She is now “a strong yes” that models create capturable value — regulation and physical-world speed remain constraints, “but I think we should see a massive acceleration in cures.”
• The firm’s name is aspirational: the ability to “suspend doubt and act with full belief until it’s true or not,” informed by companies like Sigma, Notion, and Rippling that “took a minute to begin to work.” On risk: no contrarian instinct, but “if you find the truth and it is wrongly priced and you hold onto that, you’re in a good position.” And decisions must be owned — collective ownership of investments is “nonsense to me.”
• Her confessed miss, kept as told: she knew Mikey Shulman at Suno, was asked to invest, “and I stupidly said no… I don’t think that many people want to make music.” Lesson drawn: “My intuition was just wrong,” or “somewhat wrong” because she underestimated expression and creation demand across AI tools. She also dismisses grand which-layer-wins frameworks: judge the labs’ actual priorities (ChatGPT, ads, coding) effort by effort, and spend energy on “if we’re 1% of the way in, what is the next 99% of diffusion?”
• Her one-year hope: Jevons paradox in practice — the software-engineering speedup replicated across functions, like the portfolio marketing lead who built “an autonomous marketing department” for a person-and-a-half team. Both speakers say they work more with AI, not less: “a core wisdom of Jensen’s, which is we’re all gonna be more employed” — provided people get access and education to the tooling.