What It Actually Takes to Build an AI Company [Ft. Francois Chaubard, Y Combinator]

Francois Chaubard co-created CS224D at Stanford with Richard Socher and others, the course that replaced the entire NLP curriculum two years after launch and still carries his name. He then spent nine years as a founder before closing a $60M enterprise deal at Focal. He is now a Visiting Partner at Y Combinator and a PhD researcher working on alternatives to backpropagation. He is not a generalist with opinions.

In this conversation with Chad, Francois opens with a provocation: in terms of fluid intelligence, humans still dominate. The ARC-AGI 3 benchmark makes this concrete. The best LLMs are 50x less efficient than humans on novel tasks. Nucleus sampling structurally prevents LLMs from ever being funny, not as a solvable limitation but as a mathematical consequence of how the word "pretend" sits at 2-in-1,000 token probability. Stanford's entire compute cluster for 4,000 CS students is 250 H100s, what a single OpenAI new hire gets on day one. Each of these points has an implication, and Francois spells them out.

Topics discussed:

• The Stanford vs. MIT founder model and why novelty is structurally a startup liability

• YC's "agency and taste" funding criteria and what it looks like in a three-week window

• Why vibe coding and Claude Code collapse activation energy for international founders

• GPU-per-student as a research moat metric and the compute gap between academia and industry

• Why in-context learning does not monotonically improve and what that means for agents

• The batch size one problem: why humans learn stably from a single example and models cannot

• The token probability argument for why nucleus sampling structurally prevents LLM humor

• Cartridges: learning a compressed KV cache from random initialization using SGD, with the model frozen

• Auto Research with good ideas: a priority queue over ideas.md ranked by expected value, with a Gemini Pro reflection loop

• EVO's DNA base-pair prediction model: capped at 80B parameters due to funding, and what scaling to 800B might mean