Naval's bike analogy for SFT vs RL explains why DeepSeek R1 shocked the industry
McDonaghMatthew · x · 2026-09-07
- Naval offers a one-line mental model for DeepSeek-R1's training: you can hand a kid a bike manual (supervised fine-tuning), but they'll learn faster by falling off and trying again (reinforcement learning).
- The referenced writeup explains that DeepSeek-R1-Zero was trained with RL directly on a base model, no initial SFT, spontaneously developing chain-of-thought, self-verification, reflection, and even an "aha moment" of re-evaluating its own approach.
- Core takeaway: general reasoning ability can be grown via RL incentives rather than distilled from supervised data — the detail Naval calls the most industry-wobbling part, beyond the model's performance or cheap training cost.
Related event: Naval Uses Bike-Riding Analogy to Explain SFT vs RL(2 posts)→
More from AGI Musings
- New paper finds AI agents negotiate badly under uncertainty and break deals on purpose — xuanalogue · 2026-10-26
- NBER: Only 2.9% of US consumers pay for AI; retention as low as 20%, pricing power at risk — avicgoldfarb · 2026-10-26
- AI researchers guessed robots wouldn't outrun humans until 2034 — it already happened — ben_j_todd · 2026-10-26
- AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
- ESR on the Bugpocalypse: locking down frontier models only hands the game to Chinese open weights — colinmcnamara · 2026-10-26
- Anthropic: junior tech hires fell from 27% to 9% as agent hours hit 6x employee hours — thione · 2026-10-26