CS329A Self-Improving AI Agents, Part 1: Course Overview

Name: CS329A Self-Improving AI Agents, Part 1: Course Overview
Uploaded: 2026-09-28T00:00:00.000Z
Duration: 4182 s
Channel: Stanford Online
Description: The opening lecture frames the whole course: scaling laws from GPT-2 to GPT-4, few-shot learning and chain-of-thought, then inference-time scaling, where repeated sampling plus a verifier lifts accuracy without retraining. It closes by mapping the shift from chatbots to agent workflows.

Stanford CS329A Self-Improving AI Agents | Part 1 | Course Overview

SOStanford Online@stanfordonline

Full transcript

English

Summary：The opening lecture frames the whole course: scaling laws from GPT-2 to GPT-4, few-shot learning and chain-of-thought, then inference-time scaling, where repeated sampling plus a verifier lifts accuracy without retraining. It closes by mapping the shift from chatbots to agent workflows.

Watch on YouTube

Core points (3)

1Scaling laws tied test loss to parameters, compute, and data from GPT-2 through GPT-4.
2Repeated sampling with a verifier improves accuracy without any retraining.
3Agent workflows grew out of prompt chaining, routing, and orchestrator-worker patterns.