AI’s 250 Key Builders: Inside Their Beliefs with Invest Like the Best

The episode centers on a long-form conversation between Invest Like the Best host Patrick O’Shaughnessy and early-stage investor Sarah (founder of Conviction). The discussion focuses on how the roughly 250 people building frontier AI models and companies actually think about the next few years, and what that might mean for investors.

According to Sarah, the current AI boom leaves investors torn between two dangers: slamming on the brakes and missing a generational opportunity, or repeating the classic boom-bust mistakes of past tech cycles. She describes the environment as frenzied, with a sense that the chaos is likely to intensify rather than calm down.

Within that turbulence, Sarah positions herself as “long” the idea that a small set of labs and founders are doing something fundamentally important. But she strongly rejects a future where one to three frontier models “consume the economy.” Her goal, as she frames it, is to support individuals and teams who can bend the trajectory toward a more competitive, plural AI ecosystem rather than a monolithic one.

Sarah’s thesis, as described in the conversation, rests on a “great man and great woman” view of history applied to AI. She argues that high-agency founders and researchers, given the right capital, networks, and infrastructure, can materially change outcomes in a field that otherwise looks destined for concentration among a few huge labs.

She and her partners built their firm around a focused bet: if they deeply understood the underlying models, the community, and the specific workflows those models could transform, they could gain better access and make sharper decisions than more generalized investors. Rather than just chasing “AI” as a broad theme, they looked for application areas where the current generation of models was a tight fit.

One example she highlights is Harvey, a company aimed at the legal profession. Sarah notes that if you believed in late 2022 that large language models could do next-token prediction on structured text, then law — as structured language with abundant precedents — stood out as a rational early target. Her framework starts with what is technically possible now, then narrows to what is economically valuable, and finally to which founders have both the ambition and alignment to attack that space.

To make the thesis concrete, Sarah cites specific portfolio cases and patterns she observes across AI builders. In law, she points to the founders of Harvey, Winston and Gabe, who she says were “AI-pill” long before it was fashionable. According to her, they believed AI would eventually handle enormously complex legal work, moving from trivial tasks like landlord–tenant questions in California to taking on something as intricate as an Activision Blizzard M&A deal and doing most of the work.

She also discusses Sunday Robotics, founded by Tony Zhao and Chenxi, whom she met as PhD students. Sarah and her partner Pranav concluded that these two had contributed a remarkable share of the “interesting ideas in robotics AI” over roughly four years, particularly around using modern AI to solve robustness and generalization problems under real-world data constraints.

In under two years, Sarah says Sunday Robotics went from “cardboard in a Stanford basement” to manufacturing semi-humanoid robots and iterating hundreds of times on hardware and data collection. She relays the team’s internal belief that they will have general semi-humanoid robots in people’s homes in beta by the end of the year discussed in the episode, something she notes almost no one thought was that close. Nothing is “true until it is shipped,” she cautions, but she sees the speed as evidence of how quickly AI-native teams can execute.

Sarah emphasizes several risks that complicate any bullish narrative. First is the compute bottleneck. She describes leading researchers who now feel that either (1) what they do no longer matters because future models will improve themselves, or (2) only sheer compute scale matters. Both attitudes, in her view, are disempowering and sap individual ownership at large labs.

She recounts a conversation with an infrastructure leader at a hyperscaler who said “nothing is going to move the needle for us at sufficient scale before 2030,” with issues like natural gas sourcing and nuclear deployment as key constraints. Sarah argues this is not a technology or capitalism problem but a regulatory and public-alignment problem around data centers and energy projects.

On the capital side, she worries about investors making large, research-heavy bets primarily based on pedigree and social proof rather than grounded intuition about the underlying technical and business logic. She describes debates with investor friends where the entire bull case rested on “do you know the quality of this person?” while the actual technical thesis did not make sense to her. She calls this style of decision-making “very dangerous,” especially when non-specialist investors proxy judgment to others without forming their own view.

Looking forward one year from the episode’s 2026 framing, Sarah offers several speculative but pointed expectations. She hopes to see Jevons paradox play out in knowledge work as AI agents and tools automate more mundane tasks. Drawing on software engineering as a template, she notes companies — including in her portfolio — where engineering velocity has increased dramatically with AI assistance, and expects analogous shifts across other functions.

She gives a concrete example: a portfolio company where the marketing department is effectively “a person and a half” running a traditional workload with the help of an “autonomous marketing department” that the head of marketing assembled using AI. Her expectation is not less work, but more; Sarah remarks that as individuals become more productive with AI, they tend to work more, not less, repurposing freed-up time into additional output.

For investors, she implies several forward-looking signals to monitor: how quickly agent-like tools diffuse beyond software development; whether compute bottlenecks are eased by regulatory progress on energy and data centers; and whether talented researchers and founders continue to feel empowered to drive frontier research, rather than viewing themselves as cogs in compute-constrained machines. She stresses that she prefers to spend energy on understanding “the next 99% of diffusion” over trying to predict which abstract “layer” of the AI stack will win.