Gary Marcus

Rank 12 of 47
|
Score 83
5/18/2024, 12:10:15 AM
X
In reply to:
Ali Mortazavi
@AAMortazavi
·
613d
@amahabal @GaryMarcus I’m surprised by this. It should be better at chess given its architecture.
Abhijit Mahabal
@amahabal
·
613d
@GaryMarcus Similarly with chess moves. “Chess move” it learns, but even the category “legal chess move” stumps it, as shown by this amazing video https://www.reddit.com/r/Anarchy… of ChatGPT playing Stockfish, where it appears unconstrained by legality, happy even to take its own material. 5/n
Abhijit Mahabal
@amahabal
·
613d
@GaryMarcus Similarly, those that apply to prime numbers but not to composite numbers. There are a few, such as “__ is even”, but only a tiny fraction of the number of patterns that apply to both. 4/n
Abhijit Mahabal
@amahabal
·
613d
@GaryMarcus But subsets of these categories are altogether a different kettle of fish. Think of the subcategories even number, odd number, prime number, or legal chess move.. Try to think of a few different patterns that apply to odd numbers and not to even numbers. 3/n
Abhijit Mahabal
@amahabal
·
613d
@GaryMarcus The category “Number” offers a panoply of syntactic cues. For instance, these patterns: has _ pencils,  a table for _ please, and so on. Number is a category ChatGPT has no trouble with. Likewise, the category chess move, such as Rxf3+,  is a piece of cake for ChatGPT. 2/n
Abhijit Mahabal
@amahabal
·
613d
@GaryMarcus I believe I have a simple theory of why it struggles at chess (and math). In a nutshell, although LLMs trivially learn the categories "number" and "chess move", what they don't learn well is "even number" or "a move that opens a file for a rook". 1/n

The statement '@AAMortazavi @amahabal You shouldn’t be surprised.' is a brief response in a technical discussion about the capabilities of AI in understanding and executing chess moves. The conversation revolves around the performance of AI models like ChatGPT when playing chess, particularly in terms of recognizing legal moves. The statement itself, however, does not substantively engage with broader public issues or contribute to civic dialogue. It is primarily a part of a technical discussion specific to AI behavior in chess games, which is more of a specialized topic rather than a matter of public concern or policy debate.