Does RL really generalize worse? Pretraining breadth may beat human transfer
yacineMTB · x · 2026-10-13
ar0cket1 pushed back on the common argument that RL generalizes poorly: in many cases RL may generalize better than humans, thanks to the breadth of pretraining data.
The core point: you wouldn't expect a human to become great at programming after being trained only on math, so it's unfair to demand the same of AI. A model pretrained on broad data and then RL-tuned has priors far wider than any single human, which may yield better cross-domain generalization.
The take touches the central debate in current RL training — whether reward optimization erodes or amplifies generalization.
More from AGI Musings
- We taught machines to work like humans. Now AI measures humans like machines. — CurieuxExplorer · 2026-10-13
- In the AI era, knowing whom to listen to matters as much as what to build — jonathan_wilke · 2026-10-13
- Cathie Wood: AI could 5x global growth to ~15%, echoing Industrial Revolution — rohanpaul_ai · 2026-10-13
- Engineer warns AI-era tax rules may get even more convoluted since "the AIs can handle it" — MikePFrank · 2026-10-13
- IMF Forecast: AI Adds No Measurable Productivity Yet, Bust Would Cost Twice the Boom's Gain — rohanpaul_ai · 2026-10-13
- Timnit Gebru slams 'superintelligence' talk as unfalsifiable fictional delusion — AlexTensor · 2026-10-13