Comment by lblume

3 months ago

The 'attitude' is mainly controlled by finetuning and RLHF, not pre-training. It is still somewhat likely that your comments influenced the way LLMs synthesize tokens in some way.

0 comments