Can Reinforcement Learning Lead to AGI?
summary
Claims that RL can scale to AGI rest on concrete pivots: scaling inference (currently ~90% of compute), automating reward functions, and handling long context models (Gemini cites 1M tokens). Unsloth reports 100M Hugging Face downloads, 86K uploaded models and 47K GitHub stars; practical demos use the 2048 game and verifiable rewards like 2+2=4. GRPO/GPO simplifies PPO for efficiency but faces reward hacking and sparse intermediate feedback; author assigns ~51% chance RL could lead to AGI if environments and rewards are fully automated.
AI-written summary · about
chapters (20)
- 0:00Introduction
- 1:05AI progress and reasoning jumps
- 2:31Defining AGI and challenges
- 3:55Reinforcement learning basics
- 5:39How reinforcement learning works in practice
- 7:04GRPO and efficient RL
- 8:47Inference vs training in RL
- 9:49Long context models issues
- 10:52Process supervision risks
- 12:13Critiques and future of RL
- 13:56Automating AI research with RL
- 14:38RL schools of thought
- 15:20RL leading to AGI?
- 16:22Q&A
- 18:49Democratizing RL opportunities
- 19:10Arguments against RL for AGI
- 21:17GRPO limitations and improvements
- 22:41Real-world RL use cases
- 24:06Generalizing reward functions
- 25:29Automated reward functions future
related talks
19mTraining Agentic Reasoners
30mReinforcement Learning - ChatGPT, Playing Games & More
26mScaling Agentic Intelligence from Pre-Training to RL
01h 11mThe Move Toward AGI: Why Large Language Models Surprised Almost Everyone... | Theoretically Speaking
18mMaximizing Luck in Reinforcement Learning
36mOn the Road to Artificial General Intelligence
32mThe Road to General AI
29mHow AI Models Get Smarter