GPT-5 supplied the key step in a proof in quantum complexity theory
By Scott Aaronson (UT Austin)Built and shared by Scott Aaronson (UT Austin), not by me. Summarised from the public sources below. Check them before you rely on a number.
- Who
- Scott Aaronson (UT Austin)
- Category
- Research & science
- Source
- Web
- Dates
- Published 27 Sept 2025 · Found 17 Oct 2026
Computer scientist Scott Aaronson says that a central technical step in the proof in his new paper on QMA amplification came from GPT-5 Thinking. It is the first time this has happened in his career.
What they built
Aaronson and his co-author needed to show an upper bound on how close QMA protocols can get to perfect completeness. In a dialogue of about half an hour, after several wrong attempts, GPT-5 Thinking suggested using the function Tr[(I−E(θ))^−1]. The rational function encodes how close the largest eigenvalue is to 1, and the researchers verified themselves that it worked.
The result
The paper was posted on arXiv with a proof in which a key step came from AI. Aaronson compares it to a clever suggestion from a talented PhD student.
Sources
Novo Nordisk cuts clinical study documentation from 10+ weeks to 10 minutes
By Novo Nordisk (Waheed Jowiya, Louise Lind Skov)
Novo Nordisk built NovoScribe, an AI documentation platform that writes regulatory-grade clinical study reports and other trial documents. According to Anthropic's case study, producing clinical study documentation went from more than ten weeks to ten minutes.
Swedish MASAI trial: AI-supported mammography finds more cancer with 44% less reading
By Kristina Lång and colleagues (Lund University, Skåne University Hospital)
Lund University ran a randomised trial with about 106,000 women in the Swedish screening programme, comparing AI-supported mammography with standard double reading. AI found 29% more cancers, cut radiologists' reading workload by 44%, and a 2026 Lancet follow-up showed higher sensitivity with no rise in interval cancers.
DTU researcher co-designs AI proteins that protect mice from lethal snake venom
By Timothy Patrick Jenkins (DTU Bioengineering) and David Baker (University of Washington)
Timothy Patrick Jenkins of DTU and David Baker's lab at the University of Washington used deep learning to design new proteins that bind and neutralise cobra toxins. The designed proteins protected mice from lethal doses, which could lead to cheaper and safer antivenoms.
Want this in
your business?
Tell me which case caught your eye. I will tell you straight whether it fits, and what it would take.
Start a conversation →