Bending the Curve of Discovery: AI in Science Today and Tomorrow

gist

Alex Imas and James Manyika combine 15 million Gemini interactions, an inventory of over 2,600 specialized models, and a survey of more than 600 scientists to map how AI is used in research. General LLMs handle coding and writing while specialized models like AlphaFold do domain prediction, acting as complements; the bottleneck is shifting downstream to verification and wet-lab testing, and AI nudges many researchers toward incremental work. They argue AI could become an invention of a method of invention if institutions invest in validation, moonshot incentives, safety norms, training, and open access.

ideas

• Scientists are heavy adopters. Scientists are overrepresented in Gemini usage, nearly half use AI daily, and survey respondents save just under seven hours a week, mostly reinvested in research.
• LLMs and specialized models are complements. Task overlap is limited: LLMs cover general tasks, while specialized models take over 30% of scientists’ AI time for domain analysis and modeling.
• From general-purpose technology to IMI. AlphaFold’s 200 million predicted structures show prediction moving inside large-scale search, letting researchers ask new classes of questions rather than just doing old work faster.
• The validation bottleneck and risk aversion. Ideation moves in silico while labs become verification sites; about half report hypothesis backlogs or heavy auditing, and almost half say AI steers them toward safer questions.
• Five institutional priorities. Fund verification infrastructure, reward moonshots in grants and review, set safety and reproducibility standards, update training and open access, and scale translation to societal challenges.

quotes