ChatGPT has taken the First World by storm the past few months (and probably not just the First World), wowing people with its capabilities while worrying many who are concerned that a future version may be superintelligent and malicious, or at least blithely unconcerned about human well-being, even if what it perceives as good will have dire consequences for homo sapiens. (Incidentally, while I’m not inclined to believe that there is or will be anyone “there”, I don’t think it matters if a future chatbot really is conscious or only acts as if it is. Its behavior is what will matter.)
Anyway, while it would be interesting to write more about the various LLMs in a general way (and readers are welcome to offer their thoughts about it, and how likely you think it is that we’re on the fast track to our doom), let’s stick to chess. The first generally available version of ChatGPT that came out late last year wasn’t able to play a real game of chess. It would make a few proper opening moves, but then its output was fairly random—so much so that its moves were as likely as not to be illegal. Various chess personalities had some fun with that, but as usual with computers the amusement was short-lived.
It seems that the latest edition, ChatGPT-4, is now actually playing chess. (Does anyone know how? This doesn’t seem explicable by the “ChatGPT is a glorified autocomplete” theory. Has it somehow run a neural net within itself, a la AlphaZero or Leela, to learn how to play?) Have a look here and see for yourselves. It doesn’t play well, and the author of that piece grossly overestimates its strength, but it doesn’t matter. If it really is playing chess now, it’s not going to be long before it’s playing very well - too well for humans to keep up. On the other hand, it might manage to handicap itself far more successfully than standard chess programs do when set to play at a more human-friendly level.
Regardless, this blog continues to be produced by a human, and that’s how it’s going to remain until the Terminators take over. So at least until the next post, don’t worry:
I’LL BE BACK.
I don't have GPT-4 but just tried it with the free version and it said that as a language model, it can't play chess but can give me advice on how to improve. I wonder whether future versions (4 or beyond) will be able to parse chess notation and use it to generate coaching advice for the improving player.