https://www.youtube.com/watch?v=JHq4EKMg7fI
Mr°hard word. Rozman°hard word doesn't°hard word know anything°hard word about how to get the best results from LLMs°hard word. But he won't°hard word let that stop him. In fact, his job is to do everything he can to present the machine as°hard word incompetent°hard word, in an amusing°hard word way.
----
He can't°hard word rely°hard word on the machine generating°hard word legal°hard word moves. ⚔️ he probably could, but he doesn't°hard word. ⚙️ he defines°hard word what a "flank°hard word pawn°hard word" is. because he assumes°hard word there will be a lot of non-regular°hard word viewers°hard word.
And he describes the machine explaining the history of the English°hard word Opening. Because long-windedness°hard word is a feature°hard word.
----
The machine can't°hard word (or at least doesn't°hard word) generate a list of legal°hard word candidate°hard word moves. 💡 it probably should. Human players°hard word don't°hard word explicitly do that, because they do it subconsciously°hard word faster than they can notice.
----
I can't°hard word watch all of this. His schtick is too aggravating°hard word. So I jump ahead ten minutes. The machine is now hallucinating°hard word pieces.
Because, there is no "board", and there are no constraints°hard word. 🔥 this would be less non-sensical°hard word if you couldn't°hard word see the board.
There is more of the logic°hard word that isn't°hard word logic°hard word in its output°hard word.
----
Question: is there motivation for an LLM°hard word that can actually analyze°hard word chess positions well?
That is: to give an accurate°hard word summary°hard word of a position? 💡 by mechanically°hard word generating°hard word the first two layers°hard word of processing, ChatGPT°hard word could probably do it now. ⚙️ for example: the white knight°hard word on c6 is attacking the (defended°hard word) black rook°hard word on e7, the (defended°hard word) black pawn°hard word on b4, and defending°hard word the white bishop°hard word on d8. 💡 of course, it won't°hard word be able to go down long lines of calculation°hard word without doing the calculation°hard word. which, without a lot of fine-tuning°hard word, it can't°hard word do.