LLMs Get Trounced at Chess

While Large Language Models (LLMs) have demonstrated they can do many things well enough, it’s important to remember that these are not “thinking machines” so much as impressively competent “writing machines” (able to figure out what words are likely to follow).

Case in point: both OpenAI’s ChatGPT and Microsoft Copilot lost to the chess playing engine of an old Atari game (Video Chess) which takes up a mere 4 KB of memory to work (compared with the billions of parameters and GB’s of specialized accelerator memory needed to make LLMs work).

It’s a small (yet potent) reminder that (1) different kinds of AI are necessary for different tasks (i.e. Google’s revolutionary AlphaZero probably would’ve made short work of the Atari engine) and (2) don’t underestimate how small but highly specialized algorithms can perform.