Not just combinatorics and counterexamples: GPT-5.5 solving selected problems in pure functional analysis
After witnessing numerous viral posts of GPT 5.6 in the past couple of days, concerning mathematics problems like
- (yet another batch of) Erdös problems, e.g. (#793)
a survey has been today uploaded to the arxiv, showing what the precursor model GPT 5.5 can accomplish in combination with human mathematicians... "in combination" is quite the euphemism though, as the authors make abundantly clear that apart from problem setting and proof verification the _conceptualisation and execution_ of any of the proofs was done by GPT 5.5 entirely on its own (modulo problem-independent harnesses). As one can judge for oneself, the proofs are _not_ "merely" meticulously constructed counterexamples, but rather a (presumably) firm understanding of definitions and smart synthesis of ideas, in the area of pure functional analysis / Banach space theory.
This should quiet down some goalpost-shifters who previously were claiming that 'Solving Erdoes problems is not equivalent to pure mathematics'. This branch of mathematics is as pure as it gets.
paper: https://arxiv.org/abs/2607.17388
post: LinkedIn post by author
(This is coincidentally the same person who proved an important open problem by Lindenstrauss [from the Johnson-Lindenstrass Lemma] in 1964)