Hacker News new | past | comments | ask | show | jobs | submit
This may be news for some, but they actually train these things with RL now, not just pretraining on a library of text.
The thinking part is still not a “natural” or “organic” process happening during inference.