<aside> ⚡
EVOLVING AI · GUIDE · SELF-IMPROVING AI AGENTS
</aside>
Stanford put its graduate seminar on self-improving AI agents on YouTube, free. CS329A ran as a limited enrolment graduate course in Autumn 2025, taught by Aakanksha Chowdhery and Azalia Mirhoseini. Nine recordings are now public, roughly ten hours covering scaling laws through to agents that write their own training data and propose their own problems. This page lists all nine in order, with what each one actually covers and a direct link to the video.
INDEX · click any lecture to jump straight to it
HOW TO USE THIS GUIDE
01 · Watch in order the first time
The lectures build on each other. Lectures 2 and 3 set up the vocabulary of test time compute and verifiers that every later lecture assumes you already have.
02 · Or jump to one topic
Each entry lists what that lecture covers before you commit an hour to it. Want verification, go to 03. Want RL training, go to 06. Want evaluation, go to 08.
03 · Budget the time honestly
Roughly ten hours in total, and it is not a beginner tutorial. Stanford lists NLP with deep learning or systems for ML as prerequisites, plus Python and LLM API fluency.
FOUNDATIONS · LECTURE 1 OF 9
Sets up the vocabulary the other eight lectures assume you already have. Start here even if the basics feel familiar.
What it covers
· Scaling laws that link model size, compute, and data · Why step by step reasoning appears in bigger models · Instruction tuning and human feedback (RLHF) · Getting better answers without retraining (verifiers) · The shift from chatbots to agent workflows
Watch Part 1 · Course Overview →