
RL Environments and Post-Training Evolution

The $1B Al company training ChatGPT, Claude & Gemini on the path to responsible AGI | Edwin Chen
Topic 7 of 8
Published Dec 7, 2025 · Clip 33:09–1:04:52 (31:44)
LLMs may plateau per Bitter Lesson; need new methods mimicking human learning diversity (e.g., RL environments simulating real worlds like startups with tools/Slack). Post-training evolution: SFT (mimic masters), RLHF (rank outputs), rubrics/verifiers (graded feedback), now RL for end-to-end tasks with trajectories rewarding efficient paths. Surge's research team (forward-deployed/internal) collaborates on customer models, builds better benchmarks, and trains own models to refine data ops.