3 weeks through Stanford CS329A: the generator has outrun the verifier
le_james94 · x · 2026-09-16
James Le spent 3 weeks working through all 9 lectures of Stanford's CS329A "Self-Improving AI Agents" (Azalia Mirhoseini, Aakanksha Chowdhery) and wrote a section-by-section recap.
- The through-line: the generator has raced ahead of the verifier — generating candidates is far easier than verifying them, and that gap is where agent research's hard problems live.
- METR's chart shows agent task-horizon doubling every 7 months (GPT-2 through Claude 3.7 Sonnet), but at 80% reliability the horizon collapses — both lines are real.
- Self-improvement is one loop everywhere: generate candidates → verify (unit tests, reward models, LLMs, humans) → select or train on survivors → repeat.
- Lecture 9 highlights DeepSeekMath-V2: treating verification as the product, scaling verification compute to stay ahead as the generator improves — whether that holds outside math (legal reasoning, experimental design, untestable codebases) is unknown.
- Topics span test-time compute scaling, robust verification, tools/code feedback, planning, RL train-time scaling, search and deep research agents, agentic evals, and long-horizon tasks.
Related event: Stanford Opens CS329A on Self-Improving AI Agents(5 posts)→
More from coding & agent
- Google Cloud API Gateway now acts as a remote MCP server for existing REST APIs — rseroter · 2026-09-16
- Dev ports Blender 5.1 MCP add-on to Blender 3.6, AI builds a full owl scene from one prompt — SpinachOk9137 · 2026-09-16
- Lyft cut support agent ship time from 6 months to 1-2 weeks with LangGraph and LangSmith — LangChain · 2026-09-16
- Text-only AI agent beats Doom at ~10 calls/sec, costing about $7 per hour — hackgoofer · 2026-09-16
- MCP vs plain REST APIs: developer asks if the protocol layer is worth the code — Onmas · 2026-09-16
- Removed from org, 5 years of commits gone: dev can't train agent on own history — DanielLockyer · 2026-09-16