Put your math hats on for a minute, and let’s take a look at what this beef has been since mid-October. This is a perfect example of what is wrong with AI right now.
Bubeck was excited that GPT-5 seemed to somehow solve some of the puzzles known as the Ardis problems.
Paul Ardes, the most accomplished mathematician of the 20th century Century, when he died, left behind hundreds of puzzles. To help track down who solved it, mathematician Thomas Bloom of the University of Manchester in the UK. erdosproblems.comwhich lists more than 1,100 problems and notes that about 430 of them come with solutions.
When Bubeck celebrated GPT5’s progress, Bloom was quick Call it. “This is a dramatic misrepresentation,” he wrote on X. It simply means that Bloom did not know anyone. There are millions of math papers out there, and no one has read them all. But GPT-5 probably has.
It turned out that instead of coming up with new solutions to 10 unsolved problems, GPT5 scoured the Internet for 10 existing solutions that Bloom hadn’t seen before. Oops!
There are two takeaways here. One is that breathless claims about major achievements shouldn’t be made via social media: less knee-jerk and more gut-check.
Second, GPT5’s ability to find references to previous works that Bloom did not know about is also surprising. The hype overshadowed something that should have been pretty cool in its own right.
Mathematicians are very interested in using the LLMS to trawl through a large number of existing results, a research scientist who studies the application of LLMS mathematics at the AI ​​startup Axiom Mathematics, told me when I spoke to him about this gotcha.