What Do Frontier AI Researchers Actually Believe?

TL;DR

Frontier AI researchers increasingly consider rapid capability growth plausible, including the possibility that research models could improve later models and accelerate progress. Sarah Guo argues that investors cannot reliably backtest this transition, so they must build conviction from technical fundamentals, identify valuable model-compatible workflows, support unusually ambitious founders, and preserve competition through open-source models and broader access to compute.

Transcript

It is a violently competitive landscape. I think people are very concerned that is a globally competitive landscape. There's insecurity in that because I think it's a bit narrative breaking as well. The belief is like with recursive self-improvement of AI research models that can improve the models themselves. We are a year two years away from some... Read More

Key Insights

• AI investing is difficult to backtest because the current capability transition may unfold differently from earlier technology cycles. Investors must balance the risk of missing a fundamental opportunity against the risk of behaving like participants in previous boom-and-bust markets.
• Recursive AI improvement is a serious belief among some frontier researchers. The underlying idea is that AI research models could help improve later models, potentially producing exponential intelligence within a period some people estimate at one or two years.
• Individual agency is a meaningful force in technological development. Guo believes unusually capable entrepreneurs and researchers can alter outcomes when they receive suitable risk capital, supportive networks, sufficient infrastructure, and access to concentrated technical talent.
• Competitive open-source AI depends on people choosing to build it. Producing a Western frontier model requires entrepreneurs who can raise money, commit substantial capital, gather talent, construct infrastructure, and remain willing to compete at the technological frontier.
• Early-stage investment advantage can come from focused technical understanding. Guo argues that studying the technology and its community from first principles can produce better access and decisions than a broader but less concentrated approach during a major transition.
• AI market selection can begin with model capabilities as well as customer needs. Guo looks for professions, workflows, and tasks that fit what models can technically perform, then evaluates which of those newly possible applications will also create meaningful value.
• Legal work is well matched to language models because it requires reading many documents, retrieving relevant material, working with precedent, and generating structured text. Harvey represented a bet that simple legal assistance could eventually expand into highly complex professional work.
• Market concentration is an outcome Guo explicitly opposes. She rejects the extreme possibility that the owners of only a few frontier models could consume the economy, while maintaining friendships and co-investments across both large and small AI laboratories.

Questions & Answers

Q: What do frontier AI researchers believe about rapid progress?

Some people at the AI frontier believe capability growth could accelerate sharply through recursive improvement. Their theory is that AI research models may help researchers improve subsequent models, creating a reinforcing cycle of advancement. Within the conversation, this possibility leads to the view that some form of exponential intelligence could be only one or two years away, although it is presented as a belief rather than a certainty.

Q: How should investors navigate an AI market that cannot be backtested?

Investors must form judgments from technical fundamentals, the surrounding research community, and the practical value of emerging capabilities rather than relying only on historical comparisons. Guo frames the challenge as balancing two serious risks: missing a fundamental opportunity and repeating the mistakes associated with earlier technology booms. Her approach emphasizes focused research, close relationships, deliberate risk-taking, and disciplined execution.

Q: Why does Sarah Guo focus on roughly 250 people in AI?

The group of roughly 250 represents the entrepreneurs and researchers who are actively advancing the technological frontier. Conviction aims to know these people, stay close to their work, and support them in useful ways. The strategy reflects Guo’s belief that a relatively small number of highly capable, high-agency individuals can influence the direction of AI when they have capital, infrastructure, talent, and strong networks.

Q: How does Sarah Guo build investment conviction before the market?

Guo starts by examining what current and future model capabilities make technically possible, then asks which possibilities could become valuable. She studies workflows, professions, and tasks that appear particularly compatible with the technology. This first-principles method is combined with close attention to founders whose ambitions extend far beyond an initial product, allowing her to support a long-term thesis before it becomes an obvious market consensus.

Q: Why is legal work a strong application for language models?

Legal work has characteristics that align closely with language-model capabilities. Lawyers read large volumes of documents, retrieve relevant information, rely on precedent, and generate structured written material. Guo saw those requirements as a logical match for next-token prediction and retrieval. The larger investment thesis was that a system answering a narrow legal question could eventually perform a substantial share of highly complex legal work.

Q: What is required to build a competitive open-source AI model?

A competitive open-source frontier model requires more than a general preference for openness. Someone must commit to building it, raise the necessary money, assemble a strong technical talent base, construct the supporting infrastructure, and remain willing to compete at the frontier. Guo therefore treats the future of open source as an outcome shaped by entrepreneurial agency, capital availability, and sustained execution.

Q: Why does Sarah Guo oppose a concentrated frontier-model market?

Guo opposes the extreme outcome in which the owners of only one, two, or three frontier models consume the broader economy. She wants a more competitive ecosystem and believes individual entrepreneurs can influence whether alternatives exist. Her position does not require treating major laboratories as enemies, since she works with, co-invests with, and maintains friendships among people at both large and small labs.

Q: What does Sarah Guo think makes an AI investor effective?

Guo identifies focus, effort, technical understanding, relationships, judgment, and execution as central ingredients. She believes the technology transition created an opening for investors willing to understand the systems and community more deeply than less-focused competitors. Effectiveness also depends on maintaining a high standard for collaborators, choosing ambitious founders, making deliberate bets about important markets, and helping exceptional people become more successful.

Summary & Key Takeaways

• Sarah Guo describes AI investing as a difficult balance between capturing a major technological opportunity and repeating mistakes common to boom-and-bust cycles. Because the current transition cannot be cleanly backtested, she emphasizes understanding the technology, staying close to frontier researchers and entrepreneurs, and developing conclusions from first principles.

• Conviction’s strategy centers on roughly 250 people who are actively pushing AI forward. Guo believes highly capable individuals can change outcomes when they have sufficient capital, talent, infrastructure, and network support. Her goal as an investor is to understand these people, help them succeed, and back important technical movements.

• Guo evaluates markets by asking what new model capabilities make possible and which possibilities create real value. Legal work illustrates the method because it involves reading documents, retrieving precedent, and generating structured language. She also supports competitive open-source models and rejects a future where a few frontier-model owners consume the economy.