4 ms·
I don’t think anyone is really suggesting they stole COT or that it is leaked, but rather that the final o1 outputs were used to train the base model and reason
by natdempk 2y ago
I don’t think anyone is really suggesting they stole COT or that it is leaked, but rather that the final o1 outputs were used to train the base model and reasoning components more easily.
- valine 2y agoThe RL is done on problems with verifiable answers. I’m not sure how o1 slop would be at all useful in that respect.