Sarah Guo

Sarah Guo is an investor, founder of Conviction and host of No Priors. She founded Conviction in 2022 around AI-native companies, with a thesis that building products around foundation models would require new technical and commercial practices. Her later writing asks where applications can retain value as models absorb more of the work surrounding them.[1][4][6]

On October 4, 2022, Guo announced Conviction, saying it had raised a $100 million first fund. She described the firm as purpose-built for “Software 3.0”: in her framework, Software 1.0 involved human-written code, Software 2.0 dataset labeling, and Software 3.0 manipulating foundation models. Her proposed investment scope extended from chips and developer tools to research labs and AI-native applications.[4]

Guo proposed hands-on support measured through talent, customers, and access to data and compute. She identified concrete product challenges: handling model errors, managing hallucinations and abusive prompts, and reducing per-customer fine-tuning. These were areas in which she intended the firm to develop practices alongside founders. She also committed to taking technical risk and remaining patient through cycles of enthusiasm and disillusionment.[4]

Alongside investing, Guo hosts No Priors. At NeurIPS 2024 in Vancouver, she and Conviction partner Pranav Reddy delivered the opening keynote on AI startups at the first Latent Space LIVE! miniconference, reviewing developments across startups and enterprises.[1]

On January 31, 2025, Guo announced Mike Vernal as a general partner. She explained the decision through four criteria—character, motivation, alignment and ability—writing, “The bar for a partner is really the bar for a cofounder.” She said she had known Vernal for eight years and that he had advised the fund during the preceding year. Their shared priorities, she wrote, placed reputation and performance above fund scale and required trading short-term economics for long-term potential.[5]

In her June 10, 2026 essay The Untrainable, Guo argued that cheaply verifiable work is vulnerable to commoditization: “A benchmark is a thing you can measure, and a thing you can measure is a thing you can train against.” Compilers and test suites provide inexpensive feedback for coding agents, she explained, but passing a test does not establish that a change suits a mature codebase with undocumented dependencies and operational constraints.[6]

Her defensibility thesis centers on private ground truth, permission and accountability. A more capable model does not automatically gain access to a bank’s production systems, acquire a professional license or assume liability. Applications earn their position, she argued, by arranging a customer’s private information so models can act on it, providing tools, and helping change workflows. Integration and maintenance continue throughout the customer relationship.[6]

Guo credited Gabe Pereyra with the framework that automation requires the product, model, workflow and firm to move together; three of those four move at an organization’s pace. She also used Matt MacInnis’s comparison between tokens spent answering generic questions and more valuable tokens spent reasoning over a company’s own data. Both ideas support her emphasis on customer context and the slow work of changing how people operate.[6]

Guo argued that focused applications can tune a workflow to consume fewer tokens and retain the savings, while private data and evaluations can support specialized models that outperform general models on narrow tasks. She qualified this opportunity: the “absorption frontier” keeps rising as more work becomes measurable, so companies must continually reassess their defenses. She distinguished narrow specialization from competing broadly with frontier models, which she described as a capital contest dominated by compute.[6]

Beyond defending an application, Guo emphasized deciding what is worth building. Models, in her account, can pursue a chosen goal without determining which goal deserves pursuit. She ended that argument with the suggestion, “Maybe intent is an even scarcer input than compute.”[6]

Before founding Conviction, Guo spent almost a decade incubating and investing in startups at Greylock Partners. Her October 2022 launch essay placed that work, and her childhood experience inside a technology company, in the background of her approach to founders.[4]

At Greylock, Guo began in infrastructure and security before working across consumer and enterprise businesses, fintech and crypto. Reflecting on that period, she singled out finding product-market fit and early-stage team building as work she particularly enjoyed, recalling late-night whiteboard sessions, Sunday meetings, crises and company milestones.[4]

Guo recalled growing up inside her parents’ technology company and building its first website at fourteen under the supervision of its head of hardware engineering. Her account included office bug bashes and worries about failed deals, litigation, reliability problems and difficult hires. She said those experiences shaped her respect for the courage and resilience required to build a company.[4]