The "random factoids" were verbatim training data though, one of their extractions was >1,000 tokens in length.
> GPT4 never merely regurgitated
I interpreted the claim that it can't "regurgitate training data" to mean that it can't reproduce verbatim a non-trivial amount of its training data. Based on how I've heard the word "regurgitate" used, if I were to rattle off the first page of some book from memory on request I think it would be fair to say I regurgitated it. I'm not trying to diminish how GPT does what it does, and I find what it does to be quite impressive.
Anecdotally, there were also a few examples I tried earlier this year (on GPT3.5 and GPT4) of being able to directly prompt for training data. They were patched out pretty quick but did work for a while. For example, asking for "fast inverse square root" without specifying anything else would give you the famous Quake III code character for character, including comments.
That is pretty interesting and also I didn't realize you could share chats like that.
That GPT is so bad at tic-tac-toe and relatively good at other games like chess is one of the main things that contributes to me having a lower opinion of its ability to generalize than I would have otherwise.
I think any human with GPT's abilities in chess (but somehow no prior knowledge of ttt) would have zero issue becoming an expert with a single explanation of the game. Even very young children can learn to play ttt well and at least consistently make valid moves if nothing else.
Oddly just like the text version it is still really bad at tic-tac-toe. Gave it a picture of a completed game and "Who won?" It told me "X won with a vertical line through the middle column" when in fact O won and there was only one X in the middle column.
Very impressive with almost everything else I gave it though.