Stanford released Lecture 1 of CS 329S 'Self-Improving AI Agents' to the public on X. Taught by Aakanksha Chowdhery and Azalia Mirhoseini, the course targets the frontier-lab playbook.
Key facts
- CS 329S Lecture 1 released publicly via X
- Chowdhery: ex-Google Brain, DeepMind
- Mirhoseini: ex-Anthropic, DeepMind, Google Brain
- 'Large Language Monkeys' samples thousands of outputs
- Verification named biggest unsolved challenge
Stanford opened the first lecture of "Self-Improving AI Agents" (CS 329S) to the public on X, according to @_vmlops. The course is taught by Aakanksha Chowdhery, formerly of Google Brain and DeepMind, and Azalia Mirhoseini, who worked at Anthropic, DeepMind, and Google Brain. The public release suggests the university is betting that the mechanics of agent self-improvement are now a teachable discipline, not a trade secret.
Key Takeaways
- Stanford publicly released CS 329S Lecture 1 on self-improving AI agents, taught by Chowdhery and Mirhoseini.
- The course centers verifier-based sampling and names verification as the key unsolved challenge.
The core techniques
The lecture frames "Large Language Monkeys": instead of asking a model once, sample thousands of outputs and use a verifier to pick the best one. The instructors note that models can solve problems this way that they fail in a single attempt. This is the inference-time compute scaling popularized by OpenAI's o1, which improves simply by thinking longer at inference time — no new parameters, just more compute.
The key insight, per the lecture, isn't longer chain-of-thought. It's learning when deep reasoning is actually worth the cost. That framing matters for operators who must budget inference spend per task.
Why coding agents got better
Coding agents like Claude Code and Codex became dramatically better, the instructors argue, because stronger base models plus reinforcement learning with verifiable rewards created a compounding improvement loop. The reward signal is automatic — tests pass or fail — which makes coding a natural fit for RL.
The open problem: verification
The biggest unsolved challenge the course identifies is verification. Math and code scale because answers can be checked automatically. Open-ended tasks still rely heavily on human feedback, which doesn't scale the same way. The course does not disclose a specific benchmark score or a named frontier-lab partnership; the source is a single tweet summarizing the lecture.
What to watch
Watch for Stanford to release subsequent CS 329S lectures and any associated GitHub repo with code. If the course publishes verifier benchmarks or agent evaluation harnesses, that would signal a shift from frontier-lab internal practice to open academic tooling. Also track whether Mirhoseini's Anthropic background surfaces specific agent-training recipes.
