Conviction's Sarah Guo: AI Researchers Feel "Disempowered" as Compute Scale Eclipses Individual Contribution, While Humanoid Robots Ship Faster Than Anyone Expected
Invest Like the Best podcast, September 2026, venture investor discusses the psychology of AI's research race, a surprise robotics bet, and why she is doubling down on open-source models and biotech AI
Sarah Guo, founder of the AI-focused venture firm Conviction, used a wide-ranging podcast appearance to lay out a view of the AI investing landscape that is less about which lab wins and more about what winning is doing to the people building the technology. Her most striking observation is not about model capability but about morale inside the frontier labs, where she says a growing share of researchers have quietly concluded that their individual contributions no longer matter.
The Psychological Toll of Compute-Scale AI
Guo describes the current environment as "a violently competitive landscape," more so than at any point in the last three years, and says the intensity is breeding real insecurity among the roughly 250 people she and her partner Mike consider the true frontier of AI research and entrepreneurship. The unsettling part, in her telling, is a belief that has taken hold only in the last 12 months: that recursive self-improvement in AI research models puts the industry one to two years from "some sort of exponential intelligence." She is careful to flag the skepticism this deserves, noting that Andrej Karpathy has self-effacingly admitted he has believed AGI was two years away for roughly a decade.
More consequential for investors, Guo argues, is what this belief is doing to researcher behavior. As labs scale headcount and compute budgets into the hundreds of billions of dollars, individual scientists increasingly feel one of two things: either the model will eventually solve the problem on its own, or the only variable that matters is compute scale. "Both of those are somewhat disempowering," she said, adding that this dynamic raises real questions about motivation inside the labs even as capability continues to climb. It is a subtle but important data point for anyone trying to underwrite talent risk at the frontier labs, since retention and focus are ultimately a function of whether top researchers still believe their work is the thing that moves the needle.
Sunday Robotics: A Speed of Execution That Surprised Even the Investor
The most concrete new information in the conversation concerns Sunday Robotics, a humanoid robotics company Guo and partner Pranav backed when its founders, Tony Zhao and Cheng Chi, were still PhD students at Stanford who had cut their teeth at Toyota Research, DeepMind, and Tesla. Guo says the company plans to put a general semi-humanoid robot into beta in people's homes before the end of this year, a timeline she calls not something "any of us believed" was possible when the company started less than two years ago. "It blows me away that you can move that quickly from a bunch of cardboard in a Stanford basement to a full-stack thing," she said, noting the team has manufactured its own hardware and run hundreds of iterations of model and data-collection cycles translated into real-world tasks.
The investment thesis itself is a useful window into how Conviction underwrites deep-tech robotics bets: rather than assuming the sector needs an "internet of robotics data" that does not yet exist, Zhao and Chi built a technical approach to collecting useful data cheaply and efficiently, then transferring it into model learning. Guo says she committed after a single meeting, an unusually fast decision for a firm that otherwise stress-tests its convictions for days or weeks before committing capital.
Open Source Models: "The Cat Is Out of the Bag"
On the contentious question of open-source AI models, many of them Chinese, competing at the frontier, Guo takes a clear and policy-relevant position: restricting their use inside the United States would not stop adversarial actors but would handicap "law-abiding American businesses." She argues the diffusion of capability through open models is both inevitable and desirable, since frontier proprietary models remain too expensive, too slow, or too sensitive for a large share of real-world use cases. Her prescription is not prohibition but rigorous testing, particularly around fears that Chinese open-source models could carry backdoor behaviors. "The thing to do would actually be to have a very rigorous set of safety testing on that," she said, arguing there has not been nearly enough actual research to justify current levels of alarm. She expects Sam Altman's vision of "intelligence too cheap to meter" to be substantially delivered by open-source infrastructure rather than proprietary frontier models alone, because individual businesses and users will invent use cases no frontier lab researcher could anticipate.
Compute Independence as the Next Strategic Battleground
Guo raises a national-competitiveness argument that should resonate with infrastructure-focused investors: she believes the United States risks losing "compute independence" the same way it has struggled with energy independence, given how concentrated critical steps of the semiconductor supply chain remain in unstable or inaccessible geographies. She cites a conversation with an infrastructure leader at a major hyperscaler who told her "there was nothing that is going to move the needle for us at sufficient scale before 2030," a comment on the physical bottlenecks in natural gas and power availability that constrain data center buildout regardless of capital availability. Guo frames this explicitly as a regulatory and social-alignment problem rather than a technology or capital problem, pointing to permitting fights over data centers and nuclear plants as the binding constraint. She references Jacob Helberg's supply-chain initiative, Pacia, as an example of the kind of investment needed across every link of the chain, and says Conviction has already put capital behind data center labor, nuclear energy, and alternative chip architectures as partial hedges, while passing so far on pure data-center financing plays because they fall outside the firm's technology-investing mandate.
A Reversal of Conviction on AI in Biotech
Perhaps the most quantifiable shift in Guo's thinking is on artificial intelligence's ability to create durable value inside pharma, an area she says she watched skeptically for more than five years before reversing her view. The prior consensus, in her words, was that "the only way you make money in biotech or serving pharma is by making a drug," a structure that rewards platform biotechs for owning large equity stakes in drug candidates rather than selling software. Conviction's first check into Chai Discovery, which now works with several top-ten pharma companies to accelerate parts of the R&D process, marked her break from that consensus. She is careful to note the industry has not yet had its defining moment, a new drug or indication whose trajectory was clearly altered by AI, but expects that milestone to trigger a much larger wave of capital into the space. "I think we should see a massive acceleration in cures," she said, while acknowledging that regulatory timelines and physical-world safety testing remain gating factors that no amount of model capability can shortcut.
Investment Process: Fast Instincts, Slow Verification
Guo describes her own decision-making as instinctive on people and skeptical on unfamiliar science. She rates new opportunities on a one-to-ten scale almost immediately based on the founder and idea, then spends the following days or weeks trying to find the holes in her own understanding before committing capital. She is explicit that pedigree alone is an insufficient basis for a large research bet, criticizing peers who proxy judgment to a founder's reputation without an independent view of the underlying technical thesis: "The business, the technical theory, and the business don't make sense to me," she recalled telling one well-regarded investor friend, "and I'd say people are making large-scale research bets without any intuition for them." Notably, despite running a small, five-partner firm, Guo says she sees only four to six new companies a week, a deliberate contrast to the hundreds she reviewed early in her career at Greylock, framing selectivity itself as a competitive advantage in a market flooded with capital chasing AI labels.
Fundraising Without a Polished Narrative
Guo also offers an unusually candid account of how Conviction raised its first fund, admitting she declined the conventional private-equity playbook of building a highly differentiated, pre-packaged story for limited partners. Instead, she told prospective LPs she intended to figure out the firm's specific edge through experimentation once capital was raised, a pitch she acknowledges cost her support from some institutions that prefer legible, fully worked-out theses. "I like gave people a two-pager on my background and investing history," she said, betting instead that her track record identifying and supporting exceptional people would compound over time. Four years later, with the firm reportedly ranked among the top one or two most sought-after venture partners in at least one major LP survey, she argues the lesson for founders and fund managers alike is that execution can substitute for narrative polish, provided the underlying judgment is sound.