AI tools cited speeding scientific discovery

OpenAI, Google DeepMind, Harvard researchers and FutureHouse all posted fresh examples last week of AI systems being used not just to summarize papers or write code, but to generate hypotheses, produce research software and surface candidate treatments. The burst of announcements ran from May 19 to May 20 and covered discrete geometry, liver disease, scientific programming and eye disease. Several of the claims came with outside validation or lab follow-up, while others were framed as early research tools rather than finished products. ### Which result drew the most attention? OpenAI said on May 20 that one of its internal reasoning models had disproved a longstanding conjecture in the planar unit distance problem, a question first posed by Paul Erdős in 1946. The company said the model produced an infinite family of examples that improved on the conjectured optimum and that the proof was checked by a group of external mathematicians. (openai.com) Tim Gowers, writing in the companion paper cited by OpenAI, called the result “a milestone in AI mathematics,” and Arul Shankar said the paper showed current AI models can have “original ingenious ideas.” OpenAI said the system was a general-purpose reasoning model rather than one trained only for mathematics. ### What did DeepMind say its system actually did in biology? Google DeepMind said on May 19 that its Gemini-based Co-Scientist system had been published in Nature and was designed to iteratively generate, debate and refine scientific hypotheses. (openai.com) The company said researchers had been testing it on problems including antimicrobial resistance, plant immunity and liver fibrosis. DeepMind said Co-Scientist helped identify drug-repurposing candidates and combination therapies for acute myeloid leukemia that were validated in vitro, according to the Nature paper summary. In a separate DeepMind post on liver fibrosis, the company said the system highlighted repurposing candidates, including one that blocked 91% of a scarring-linked response in lab tests. ### Where does the Google-Harvard coding work fit in? (deepmind.google) Harvard SEAS said on May 19 that a Google-led team co-led by Michael Brenner and Google DeepMind’s Shibl Mourad had built an AI system called Empirical Research Assistance, or ERA, that can automatically write scientific software programs. Harvard said the system outperformed software written by experts and was published in Nature. (nature.com) Harvard said ERA combines Google’s Gemini model with tree search to explore and refine thousands of code variants for “scorable tasks” such as disease forecasting or protein-shape prediction. The university said the aim is to automate a software design cycle that can otherwise take human researchers months or years. ### What was FutureHouse claiming on blindness research? (seas.harvard.edu) FutureHouse said in a May 20 research announcement that its multi-agent system Robin identified ripasudil, a glaucoma drug, as a candidate treatment for dry age-related macular degeneration, a leading cause of irreversible blindness. The group said Robin generated the hypotheses, proposed experiments, analyzed data and selected follow-up candidates, while human researchers carried out the physical lab work. (seas.harvard.edu) Sam Rodriques, writing for FutureHouse, said code, data and full agent trajectories would be released on May 27. FutureHouse described the result as its first AI-generated discovery using an integrated system that combined literature search, evaluation and data-analysis agents. ### How much of this is deployed versus still experimental? (futurehouse.org) Google DeepMind said Co-Scientist would begin rolling out “in the coming weeks” through a Hypothesis Generation tool, with researchers invited to register interest at labs.google/science. OpenAI linked its geometry announcement to a broader effort to test whether advanced models can contribute to frontier research. Harvard framed ERA as a system that could accelerate discovery across domains, while FutureHouse tied its next step to a May 27 release of underlying materials. (futurehouse.org) May 27 is the next concrete date in the sequence: FutureHouse said it plans to release Robin’s code, data and full agent trajectories that day. Google DeepMind, meanwhile, said researchers can register now for Co-Scientist access through labs.google/science as the rollout begins in the coming weeks. (futurehouse.org) (deepmind.google)