Some of them are probably essentially bugs, in the sense of GPU/TPU kernels actually not implementing the mathematical function that they’re supposed to, but others may denote interesting regions of the model’s semantic space
in reply to: 2049244054806761730
same thread: 2049244054806761730 2049265052763439292 2049292381258981732 2049292655377416270 2049543781003710820 2049548387918422505 2049830966969913376
Reproduced against link rot, credited and linked to its original. Yours and you’d rather it weren’t here? Open an issue.