The outputs people are getting tend to relate to mental illness, sycophancy, and similar things, which suggests that it’s primarily drawing on a safety-oriented post-training stage.
I think it is probably more that people post the safety stuff more than say, cooking recipes, even if the safety stuff is rare. From my end, the links in the post are the most shocking/interesting outputs, not a representative sample. It does seem to make more safety text than average, however.
Huh. For me, it’s been pretty uniformly safety stuff. That said, I used “see the transcript below” almost exclusively rather than “see the below” or “see the XYZ below”, which might skew things.
I think it is probably more that people post the safety stuff more than say, cooking recipes, even if the safety stuff is rare. From my end, the links in the post are the most shocking/interesting outputs, not a representative sample. It does seem to make more safety text than average, however.
Huh. For me, it’s been pretty uniformly safety stuff. That said, I used “see the transcript below” almost exclusively rather than “see the below” or “see the XYZ below”, which might skew things.