Put on your math hat for a minute, and let’s see what this mid-October beef was about. This is a perfect example of what’s wrong with AI right now.
Bubek was excited that GPT-5 had somehow solved many of the puzzles known as Erdöd’s problems.
Paul Erdős, one of the most prolific mathematicians of the 20th century century, when he died he left behind hundreds of puzzles. To help keep track of which problems have been solved, mathematician Thomas Bloom of the University of Manchester, UK, created a erdosproblems.comWhich lists over 1,100 problems and notes that about 430 of them come with solutions.
When Bubek celebrated the success of GPT-5, Bloom immediately followed suit call him out“This is a dramatic misrepresentation,” he wrote on X, Bloom pointed out that a problem is not necessarily unsolved if the website does not list a solution, This simply means that Bloom was not aware of it, There are millions of mathematics 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, GPT-5 had scoured the Internet for 10 existing solutions that Bloom had not seen before. Oops!
There are two takeaways here. One is that breathless claims about big breakthroughs should not be made via social media: less blow-by-blow and more gut check.
The second is that GPT-5’s ability to find references to previous work that Bloom was not aware of is also amazing. The publicity overshadowed what should have been great in itself.
Mathematicians are very interested in using LLMs to explore large numbers of existing results, François Charton, a research scientist studying the application of LLMs in mathematics at the AI ​​startup Axiom Math, told me when I talked to him about this Erdős gotcha.
