Google reports an AI system that combines large language models with tree search to autonomously write expert‑level scientific software and invent novel methods. In tests, it created 40 new single‑cell analysis methods that beat the human leaderboard and 14 epidemiological models that set state‑of‑the‑art for COVID‑19 hospitalization forecasts.
— If AI can originate superior scientific methods across fields, it shifts research from AI-as-assistant to AI-as-inventor, with implications for funding, credit, safety, and the pace of discovery.
Alexander Kruel
2025.10.16
70% relevant
The post’s headline result—an LLM proposing a previously unreported CK2→MHC‑I mechanism that Yale validated—extends the theme that AI can originate new, high‑value scientific contributions, not just assist, akin to prior reports of AI inventing superior scientific methods.
Alexander Kruel
2025.10.01
86% relevant
The post highlights Google’s 'AI as a research partner' (AlphaEvolve) that not only generated candidate structures but also invented verification optimizations (branch‑and‑bound and system-level tweaks) yielding up to 10,000x speedups—directly aligning with the claim that AI can originate superior scientific methods, not just assist.
Scott
2025.09.27
76% relevant
Aaronson reports that GPT5‑Thinking supplied a key technical step in a new QMA amplification optimality proof, paralleling prior claims that AI can originate high‑level scientific methods rather than merely assist execution.
Tyler Cowen
2025.09.17
78% relevant
The post cites Arc Institute reporting 'the first viable genomes designed using AI,' a concrete instance of AI originating frontier‑level biological designs rather than merely assisting analysis.
Brent Orrell
2025.09.16
80% relevant
The article cites a paper where an AI system produced and evaluated 1,773 neural‑network architectures—yielding 106 state‑of‑the‑art linear‑attention models in ~20,000 GPU‑hours—showing AI can originate high‑performing methods rather than merely assist humans.
Alexander Kruel
2025.09.10
100% relevant
ArXiv paper (2509.06503) claiming cross‑domain, AI‑devised methods that outperform human state‑of‑the‑art in bioinformatics and epidemiology.