CS329A Self-Improving AI Agents, Part 9: Future Research Areas

Name: CS329A Self-Improving AI Agents, Part 9: Future Research Areas
Uploaded: 2026-09-28T00:00:00.000Z
Duration: 4062 s
Channel: Stanford Online
Description: The finale maps open problems: Multi-Agent Fine-Tuning pairs generator and critic agents, Deep Math V2 automates proof checking via meta-verification, and Absolute Zero generates its own tasks. An intelligence-per-watt lens shows small local models covering most chatbot queries.

Stanford CS329A Self-Improving AI Agents | Part 9 | Future Research Areas

SOStanford Online@stanfordonline

Full transcript

English

Summary：The finale maps open problems: Multi-Agent Fine-Tuning pairs generator and critic agents, Deep Math V2 automates proof checking via meta-verification, and Absolute Zero generates its own tasks. An intelligence-per-watt lens shows small local models covering most chatbot queries.

Watch on YouTube

Core points (3)

1Meta-verification lets Deep Math V2 check proofs without reference solutions.
2Absolute Zero trains on self-proposed coding tasks with no external data.
3Models of 20B parameters or fewer already handle 88.7 percent of chatbot queries.