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

AI development has shifted from merely increasing model size to optimizing how systems "think" during use and how they function as autonomous units capable of executing multi-step workflows.

• Scaling laws — Increasing model parameters, training data, and compute power systematically lowers error rates in language models [03:02].
• Emergent capabilities — Abilities like multi-step reasoning appear suddenly once models reach a high parameter count, allowing them to handle tasks without being trained on them [05:35].
• Alignment techniques — Developers turn raw predictors into useful assistants by steering outputs using instructions and reinforcement learning based on human feedback [11:45].
• Inference scaling — Performance is boosted by generating many potential answers at runtime and using a verifier to select the best one, rather than relying only on model size [21:10].
• Agentic systems — Modern workflows move beyond simple chat interfaces toward autonomous agents that can plan steps, correct their own mistakes, and use external tools to reach a goal [42:32].
• Feedback loops — Improvement relies on methods like unit testing generated code to verify correctness, which helps the system refine its performance over time [50:10].

How does inference-time scaling differ from increasing model parameters during training? What defines an agentic workflow?