Scaling Agentic Intelligence from Pre-Training to RL
summary
Covers how scaling from pre‑training to reinforcement learning uses long chain of thought examples plus RL with verifiable rewards (calculator checks, unit tests, constraint verifiers) to improve math and coding. Coding agents show deterministic gains and can solve longer tasks; median solvable task length is doubling every seven months but reliability remains ~50% for many long‑horizon tasks. Highlights context windows >100k tokens, need for credit assignment over thousands of steps, and synthetic interactive environments for training and evaluation.
AI-written summary · about
chapters (5)
- 0:00Scaling and reasoning
- 4:29RL and agentic benchmarks
- 13:01Coding agents and tool use
- 17:49Challenges in long-horizon tasks
- 21:09Future directions and hiring
related talks
19mTraining Agentic Reasoners
26mCan Reinforcement Learning Lead to AGI?
38mRetrospective on multi-agent learning as a path to intelligence
19mRL for Autonomous Coding
14mHard Won Lessons from Building Effective AI Coding Agents
29mHow AI Models Get Smarter
36mOn the Road to Artificial General Intelligence
30mReinforcement Learning - ChatGPT, Playing Games & More