mit-hyung-won-chung-incentivize

MIT EI Seminar: “Don’t Teach. Incentivize.” — Hyung Won Chung (OpenAI)

Date: 2024-09-26 (saved to Readwise)

Summary

OpenAI researcher Hyung Won Chung argues for incentivization over direct teaching in AI training — let the model discover solutions through reward signals rather than explicit instruction.

Key Insight

“Don’t teach. Incentivize.” — foundational philosophy behind RLHF and RL-based post-training.

Concepts

• RL Infrastructure
• Reward Hacking

See Also

• post-training-101 — post-training pipeline
• the-bitter-lesson — scale and general methods win