Pretraining sets the floor, RL raises the ceiling: on how the two divide the work
JacquesThibs · x · 2026-11-23
JacquesThibs weighs in on the pretraining-vs-RL relationship: pretraining solidifies concepts into crisper representations and stores facts precisely, which RL can leverage via elicitation to complete complex, long-horizon tasks. Both determine the ceiling, but pretraining helps set the floor.
More from Models
- Voyager releases open-source shot-reader-8b multimodal model on Hugging Face — ravisparikh · 2026-11-23
- Matthew Berman baffled: Gemini weekly limits stay at 100% after a weekend of heavy Neon use — MatthewBerman · 2026-11-23
- 5 free uncensored GGUF models on Hugging Face: MiniMax, Qwen, Kimi, Nemotron — Promptmethus · 2026-11-23
- Google AI Studio's periodic-table model selector reveals Gemini 4 Neon: 1M context, $3/$15 pricing — patloeber · 2026-11-23
- Domingos: Log-Perplexity Fell to 0.75 Bits While Cost Rose Orders of Magnitude — pmddomingos · 2026-11-23
- Nathan Lambert: RL Is 'Frying' Models and Reducing Their Diversity — natolambert · 2026-11-23