M9: Applications, Case Studies, and Frontiers
The course closes by looking at what game-theoretic ideas do in the largest AI systems built so far. The first lecture covers game-playing agents at scale: AlphaZero, AlphaStar's league, OpenAI Five, Cicero in Diplomacy, and Pluto, the 2026 StarCraft bot trained entirely by self-play. We read their training methods as the algorithms from earlier modules run at scale. Self-play and fictitious play become leagues and PSRO, exploitability becomes a practical weakness, and Cicero's piKL brings language into a game that is not two-player zero-sum. The second lecture asks how we know an agent is good: benchmarks such as SWE-bench, WebArena, and GAIA, and the cost, variance, contamination, reproducibility, and shortcut problems that make their numbers less reliable than they look. It closes with multi-agent benchmarks through a game-theoretic lens and an 11-point checklist for evaluating any benchmark. After this module, class time goes to your projects. Recorded presentations are due November 30 and final reports December 3.
Module preview
The course closes by looking at what game-theoretic ideas do in the largest AI systems built so far. The first lecture covers game-playing agents at scale: AlphaZero, AlphaStar's league, OpenAI Five, Cicero in Diplomacy, and Pluto, the 2026 StarCraft bot trained entirely by self-play. We read their training methods as the algorithms from earlier modules run at scale. Self-play and fictitious play become leagues and PSRO, exploitability becomes a practical weakness, and Cicero's piKL brings language into a game that is not two-player zero-sum. The second lecture asks how we know an agent is good: benchmarks such as SWE-bench, WebArena, and GAIA, and the cost, variance, contamination, reproducibility, and shortcut problems that make their numbers less reliable than they look. It closes with multi-agent benchmarks through a game-theoretic lens and an 11-point checklist for evaluating any benchmark. After this module, class time goes to your projects. Recorded presentations are due November 30 and final reports December 3.
Lectures and materials
L20: Game-Playing AI at Scale: From Self-Play to Leagues to Language
AlphaZero, AlphaStar, OpenAI Five, Cicero, and Pluto read as game-theory algorithms at scale: self-play, leagues and PFSP, exploitability, and piKL.
L21: Why PFSP?
Fictitious self-play, opponent selection, neural regret minimization, and last-iterate convergence.
L22: Agent Benchmarks and Evaluation
How do we know an agent is good? We examine benchmarks like SWE-bench, WebArena, and GAIA, and the cost, variance, contamination, reproducibility, and shortcut problems that make benchmark numbers less reliable than they look. The lecture closes with multi-agent benchmarks through a game-theoretic lens and an 11-point checklist for evaluating any benchmark.