GPT-5.6 fills a 35 year gap in math

The model has found a solution to a mathematical problem that has baffled experts for 35 years. It did so in 168 minutes of autonomous work by finding new theoretical bounds on an optimization problem. This is a subfield of math that plays a key role in physics and machine learning. It beats all previous publications since 1991 and approaches the results of eight years of unpublished research conducted by the Microsoft Research team.

AI’s potential has always been limited to replicating existing knowledge. It could never create anything truly original, like groundbreaking scientific discoveries. But now, AI has reached a new height, and we can ask ourselves what tasks will still be left for us humans.