reinforce
A basic policy-gradient estimator that scales the log-probability of an action by its observed return.
As defined in Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch
Other definitions
The process of repeatedly engaging with a concept through practice, explanation, exploration, and variation to ensure knowledge stays in long-term memory and becomes automatic.
Focusing Your Unconscious Mind: Learn Hard Concepts Intuitively (And Forever)
Where it comes up · 2
- Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch2 claims
- Focusing Your Unconscious Mind: Learn Hard Concepts Intuitively (And Forever)1 claim
Most important claims · 3 of 3Search all claims for “reinforce” →
- Naive win-reinforcement self-play RL is extremely high variance because in a matchup between evenly matched policies, only one critical move in one game out of ~100 (each ~300 moves) actually drove the win, so you have one useful supervision signal buried in ~99×300 neutral labels; the gradient variance grows quadratically with trajectory length T.Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch
- Knowledge must be reinforced through multiple reinforcement steps—invention, practice, explanation, and exploration—otherwise it will not remain in your brain for extended periods.Focusing Your Unconscious Mind: Learn Hard Concepts Intuitively (And Forever)
- Current LLM RL treats an entire generated sequence as a single action with T=1, summing token log-probs into one log-prob, because decomposing into per-token rewards introduces cross-correlations between actions that magnify gradient variance—this is why LLMs do one-step rather than multi-step RL.Dwarkesh Patel Podcast - Eric Jang – Building AlphaGo from scratch