The opening lecture of Stanford CS329A, taught by Akansha (adjunct professor, Reflection AI) and Azalia Mirhoseini (ex-Anthropic Claude, Google DeepMind Gemini) — both ex-Google Brain. The arc: how we got from scaling laws to reasoning models to agents, and where the frontier sits now. Three movements:
1. Scaling → inference-time scaling. Pre-training scaling (compute/data/params) hit saturation around 2024, so the frontier moved to inference: with the model fixed, repeated sampling + a verifier extracts far more capability — “Large Language Monkeys” showed 7B-70B models with 10K samples beating GPT-4o asked once (“models already know a whole lot more than what you get out of them when you just ask them once”). Reasoning models (o1, DeepSeek, Gemini thinking) then showed log-linear test-time scaling on pass@1, using learned skills: problem analysis, task decomposition, self-correction, backtracking.
2. The self-improving loop. DeepSeek and Gemini thinking combined test-time scaling with fine-tuning: generate tons of synthetic solutions/reasoning traces during inference, then fine-tune the model on them. That loop — test-time scaling producing training data for the model that produced it — is the course’s “self-improving” core, and the instructors call it the most exciting open area.
3. Agents and verification. Chatbots are single-turn; agents (Claude Code, Codex, Deep Research) accomplish tasks end-to-end: goal → plan → act → feedback → stop, with tools and memory. Today most workflows are hand-constructed (chaining, routing, parallelization, orchestrator, LLM-as-judge, verifiers). The binding constraint is verification: easy to generate, hard to verify — human feedback is the bottleneck in creative domains, while RL with verifiable rewards (math/code) is what made coding agents reliable this year.
Watch the lecture on YouTube — full summary in the vault note.