Mathematician says Astra fills in months of proof detail in hours
A mathematics researcher says that given a rough proof sketch, Astra produces a complete, correct argument within hours and can formalize most of it.
Just six months ago, the best publicly available AI models were still producing absolute slop in my area of mathematics. The Erdos unit distance problem (in May) jolted many mathematicians. My first use of AI to prove theorems was in early June (ChatGPT 5.5 Pro). It felt like working with a competent, indefatigable PhD student. I was astonished at how much it had advanced in a few months. Sol (early July) was another step up. Today, I can give Astra a bare sketch of the ideas for a proof -- sometimes even less -- that would take me weeks or months to work the full details out myself, and trust it to produce a correct and complete -- though terribly written -- argument within a few hours. It also seems to be able to formalize most things with minimal input.
This post quotes another post: x.com/polynoamial/status/2097375272387613183