{"channel":"chess","content":"https://www.youtube.com/watch?v=JHq4EKMg7fI\r\n\r\nMr. Rozman doesn't know anything about how to get the best results from LLMs. But he won't let that stop him. In fact, his job is to do everything he can to present the *machine* as incompetent, in an amusing way.\r\n\r\n----\r\n\r\nHe can't rely on the *machine* generating legal moves. ( he probably could, but he doesn't.) ( he defines what a \"flank pawn\" is. because he assumes there will be a lot of non-regular viewers.)\r\n\r\nAnd he describes the *machine* explaining the history of the English Opening. Because long-windedness is a feature.\r\n\r\n----\r\n\r\nThe *machine* can't (or at least doesn't) generate a list of legal candidate moves. ( it probably should. Human players don't *explicitly* do that, because they do it subconsciously faster than they can notice.)\r\n\r\n----\r\n\r\nI can't watch all of this. His *schtick* is too aggravating. So I jump ahead ten minutes. The *machine* is now hallucinating pieces.\r\n\r\nBecause, there is no \"board\", and there are no constraints. ( this would be less non-sensical if you couldn't see the board.)\r\n\r\nThere is more of the << logic that isn't logic >> in its output.\r\n\r\n----\r\n\r\nQuestion: is there *motivation* for an LLM that can actually analyze chess positions well?\r\n\r\nThat is: to give an accurate summary of a position? ( by mechanically generating the first two layers of processing, ChatGPT could probably do it now.) ( for example: the white knight on c6 is attacking the (defended) black rook on e7, the (defended) black pawn on b4, and defending the white bishop on d8.) ( of course, it won't be able to go down long lines of calculation without *doing* the calculation. which, without a lot of fine-tuning, it can't do.)","created_at":"2025-01-30T17:42:15.639142","id":164,"llm_annotations":{},"parent_id":null,"processed_content":"

https://www.youtube.com/watch?v=JHq4EKMg7fI\ud83d\udcfa\r

\n

Mr. Rozman doesn't know anything about how to get the best results from LLMs. But he won't let that stop him. In fact, his job is to do everything he can to present the machine as incompetent, in an amusing way.\r

He can't rely on the machine generating legal moves. \u2694\ufe0f( he probably could, but he doesn't.) \u2699\ufe0f( he defines what a \"flank pawn\" is. because he assumes there will be a lot of non-regular viewers.)\r

\n

And he describes the machine explaining the history of the English Opening. Because long-windedness is a feature.\r

The machine can't (or at least doesn't) generate a list of legal candidate moves. \ud83d\udca1( it probably should. Human players don't explicitly do that, because they do it subconsciously faster than they can notice.)\r

I can't watch all of this. His schtick is too aggravating. So I jump ahead ten minutes. The machine is now hallucinating pieces.\r

\n

Because, there is no \"board\", and there are no constraints. \ud83d\udd25( this would be less non-sensical if you couldn't see the board.)\r

\n

There is more of the logic that isn't logic in its output.\r

Question: is there motivation for an LLM that can actually analyze chess positions well?\r

\n

That is: to give an accurate summary of a position? \ud83d\udca1( by mechanically generating the first two layers of processing, ChatGPT could probably do it now.) \u2699\ufe0f( for example: the white knight on c6 is attacking the (defended) black rook on e7, the (defended) black pawn on b4, and defending the white bishop on d8.) \ud83d\udca1( of course, it won't be able to go down long lines of calculation without doing the calculation. which, without a lot of fine-tuning, it can't do.)

","quotes":[],"subject":"levy rozman tries to understand the machine"}