RL can discover genuinely new things, but the line between new and old is blurry
AndrewLampinen · x · 2026-10-03
Researcher Andrew Lampinen argues that RL can indeed discover genuinely new things, though the boundary is not always clear: is making a novel connection between two areas of math "new" if both appeared in training data? He adds that most human mathematicians who invent something new are also deeply studied in relevant areas.
Related event: DeepMind Researcher Debates Neurosymbolic Camp on LLM Math Limits(4 posts)→
More from Research
- Yoav Goldberg: the 'post-training adds no knowledge' dogma is clearly no longer true — yoavgo · 2026-10-03
- Anil Seth pushes back on Steve Yegge: denying machine consciousness isn't human exceptionalism — anilkseth · 2026-10-03
- YacineMTB: PPO Is All You Need, With Tiny RNNs Under 1M Params — yacineMTB · 2026-10-03
- Yacine points RL world model debaters to PufferLib as the reference project — yacineMTB · 2026-10-03
- Hutter et al. Argue Generalization Requires Universal Induction in New arXiv Paper — examachine · 2026-10-03
- NVIDIA team lands 10th place among ~4,000 in Biohub cell tracking Kaggle competition — JFPuget · 2026-10-03