Deutsche Bank: best AI models will reach 'superintelligence' by 2030
sudoraohacker · x · 2026-10-14
- Deutsche Bank research argues the best AI models in 2030 will have 'superintelligence' qualities — beating human experts in all domains — and maps what happens when cognitive work gets abundantly cheaper.
- Early signs cited: GPT-5 running 36,000 experiments and cutting cell-free protein costs 40%; Google saying AI-written code went from 25% to 75% in about a year; GPT-5.4 running 10,080 robot-lab reactions and lifting drug-making yields.
- Also: chatbot alone outperformed humans in a 1,298-person medical trial (Nature Medicine); AI reading 500M words of state law let San Francisco cut a third of reporting requirements; a one-person bot won the Metaculus Cup; eight Nvidia agents trained robot arms to 99% success in two hours.
- Extrapolations include virtual cells, top-tier legal counsel for everyone, write-and-discard software, molecules designed to spec, and continuously audited law.
More from AGI Musings
- Reading books while your agents work: AI is giving users their attention back — martyamark · 2026-10-14
- Richard Socher says p(doom)=0 and explains how close Recursive SI is to automating science — jeremyakahn · 2026-10-14
- Andrew Ng blasts AI extinction fears, says 'made-up safety arguments' stifle competition — JHochderffer · 2026-10-14
- Biologist pushes back on 'automate all labs': only ~20% of assays make sense — anshulkundaje · 2026-10-14
- AI Bubble Check: Most People Still Don't Know Who Claude Is — RachelVT42 · 2026-10-14
- Timnit Gebru resurfaces Michael Jordan's 2018 essay debunking machine-god hysteria — AlexTensor · 2026-10-14