I have no idea what he's planning to do, but this kind of thing just now becoming feasible.
I get the impression doing something like the title is a dream of many Chinese AI researchers, and why we [edit:see] them focusing on things like the mathematics competition datasets. I am slightly in this direction myself.
I could review his published research, but in the interim, I had GPT painstakingly review the math in the image, and extrapolate the direction, which it considered to be a suggested hybrid successor to the Transformer for structured and scientific domains: https://chatgpt.com/share/6931c915-88c8-8012-9ca4-b0097e53e8...
Yeah, that’s not right. I’m not sure about painstakingly… it said it couldn’t make out the notation, and spat out what it thought it could read, and you never checked it - nor read the articles for context, just assumed it was to do directly with further AI work.
It picked up on the polynomial, then what it thought was a scheme/sheaf being defined is actually the finite field with six elements. It also misread “Thue” as “the”.
If you had corrected what it read from the board, then gave it the context that he was a number theorist now working for a company trying to get AI to work through proofs, then you may have got the correct answer that this appears to be them crafting problems on polynomial reduction to test how the LLM reasons about proof.
> If you had corrected what it read from the board, then gave it the context that he was a number theorist now working for a company trying to get AI to work through proofs
It was just a quick and dirty chat. A proper evaluation will consider his published research to date.
I get the impression doing something like the title is a dream of many Chinese AI researchers, and why we [edit:see] them focusing on things like the mathematics competition datasets. I am slightly in this direction myself.