Retrospective on multi-agent learning as a path to intelligence
Talkat Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives · 2024-06-24..06-28·
summary
Large-scale RL runs on a 2D physics simulator with transformer policies and LiDAR-style observations produce staged behaviors: random motion, predator–prey by ~2 million episodes, then tool use and a counter at about 15 million episodes in a hide and seek setup, followed by the surprising box surfing exploit and a final lock-down defense; emergent communication appears but agents do not transfer skills to downstream tests and training plateaus after weeks of continued running.
AI-written summary · about
chapters (13)
- 0:00Multi-agent emergence research
- 1:51Multi-agent settings for AI
- 4:52Hide and seek emergent behaviors
- 10:07Limitations of hide and seek results
- 14:05Alternative multi-agent experiments
- 17:15Why multi-agent learning failed
- 20:12Simulated societies and LLMs
- 22:05LLM debate crowd effects
- 26:32Future of multi-agent learning
- 28:04Q&A: Adversarial vs cooperative
- 30:41Q&A: Debate as supervision
- 32:49Q&A: Division of labor
- 36:29Q&A: Reinforcement with culture
related talks
36mOn the Road to Artificial General Intelligence
19mTraining Agentic Reasoners
26mScaling Agentic Intelligence from Pre-Training to RL
32mThe Road to General AI
31mFrom Single AI to Multi AI Systems
35mIndustrial Applications of Reinforcement Learning
30mReinforcement Learning - ChatGPT, Playing Games & More
36mSkyRL: Democratizing RL Training for Researchers