Guo bets AI's frontier stays open while compute stays scarce
Guo argues the AI frontier is still open to entrepreneurs, and an investor's edge comes from a grounded view of each business rather than founder pedigree. The hard limits are physical, meaning compute, energy and siting, and these move far slower than software. With only about 1% of diffusion done, the biggest opportunity lies in applying AI across fields like law, robotics and biology.
Takeaways
- Entrepreneurs, not a few labs, will shape AI. Guo rejects the idea that one to three frontier model owners will define the future, since that would consume the economy. She argues high-agency people change outcomes, and says her goal is to be the best investor, not the person who made things happen.
- Compute and physical supply chains are the real bottleneck. Compute over the next five to ten years is Guo's biggest worry. A hyperscaler infrastructure leader told her nothing would move the needle at scale before 2030. She sees regulatory and social limits, such as data center siting and nuclear costs.
- The research frontier is fiercely competitive and unsettled. Many researchers now believe recursive self-improvement could bring exponential intelligence within about two years, a view Karpathy has held for a decade. At large labs many feel disempowered. Guo says at most 5% to 10% of her backed companies would not have been funded without her.
- Law shows why language-based work is ripe for AI. Conviction approached Harvey on the logic that law is structured language. Guo projects legal AI could one day handle an Activision Blizzard acquisition and do 85% of the work. The firm also covers biology, defense and robotics.
- Backing research founders on pedigree is a mistake. Guo says many investors judge research companies by founder pedigree and referrals instead of a grounded view of the business. Conviction rates deals from one to 10, sends memos to outside investors like John Lilly and Dylan Field, and studies for days or weeks.
- Restrict open models less and test them more. Competitive open models from China, the US and Europe, including Mistral and NVIDIA, already exist. Guo argues restricting them would hurt American businesses rather than attackers, and favors rigorous safety testing. She also says the US must automate to rebuild its industrial base.
- AI changes biology economics, and diffusion has barely begun. Conviction wrote the first check into Chai Discovery, which works with pharma companies, before there was evidence its approach would work. Guo expects a wave of investment once AI clearly changes a drug's trajectory. She says we are about 1% into diffusion.
Conviction's bets and habits