Well first of all, you’re right that my comment about ChatGPT being “the most intelligent person in the world” was meant in jest. But I would argue that there’s a grain or two of truth there, too, because of the amazing ability of GPT to source information, and also – I will say this unequivocally – to solve problems in logic, and definitely problems it hasn’t seen before. I don’t know if you’ve seen this thread but it goes back to 2024 and ChatGPT 3.5, but it contains may examples of logic puzzles and similar questions that even older versions of GPT were able to answer successfully.
There appears to be a strong connection between extensive training in the structure and semantics of human language on a sufficiently large scale LLM that allows it to learn internal representations of logical abstractions so they can generalize situations they’ve never seen before. These are not simple composites of the training corpus, but much more profound instances of internalized logic.
This is a major reason that the best LLMs like GPT-4 and GPT-5.2 have done so well on human intelligence tests, professional skill tests, and general problem-solving.
Now to get to your specific examples. Sorry, you’re wrong when you say that “everybody is familar with Connect 4” – I’m not! So I just skipped over that example for the moment, but I want to get to the chess one and there may be some insights there that relate back to Connect 4, too.
I gave ChatGPT-5.2 the URL you posted, and simply asked it to find the best move for Black (I didn’t say there was a mate-in-1 possibility).
First of all, you have to admit it’s impressive that you can give it a web URL that contains an image of a chessboard, and have it analyze the image and understand the piece positions. But the response it gave me on the best move for Black was wrong, and certainly not optimum.
I pointed out that Black could checkmate in one move. GPT said, nice observation, but no, that’s a check, not a mate, because White could interpose their queen. I said it was impossible for White to do that. It agreed that I was right.
I think what’s happening here is probably related to what I’ve seen with other GPT failures where it seems to have big trouble with spatial relationships. Despite its prowess at image processing and image generation, it still seems to have trouble with real-world spaces. It’s not that it doesn’t understand chess per se – far more trivial machines can play chess at a grandmaster level – it’s just this spatial manipulation problem that it still has.
Which has very little to do with emergent reasoning skills in other domains, and may also explain whatever problem it had with Connect 4. Why don’t you try asking it typical questions from an IQ test or something similar, like an SAT test, and see how it does on that. Because GPT generally handles those quite well.