Thrilled to release AI in Science: Early Insights into the world. This was a project co-led with the excellent @JMateosGarcia and @m_codreanu and an incredible team of co-authors. This collaboration between @Google and @GoogleDeepMind teams is the first, initial look, representing the beginning of a research agenda for us. The promise of AI for economic growth and flourishing moves directly through its impact on science, and this paper is the start of a brand new effort in our group on the study of AI's impact on science. Lots of results, lots of insights, lots of questions still to answer. There is great excitement -but also concerns- about the impacts of AI on science, but so far little data. We provide early insights on this from three complementary data sources: 1) a sample of 15 million Gemini interactions, 2) an inventory of over *2,600* specialized AI models (e.g., AlphaFold) across disciplines, 3) a new survey of over 600 scientists. But what does this tell us about how scientists actually use AI in their workflow, and is the impact of AI? To answer these questions, we map these data to: a) a new taxonomy of scientific tasks from @ProfNeilT and his lab; b) bibliometric data tracking scientific publications, and citations (including to specialized AI models). Four main findings emerge. 1) We find broad adoption and coverage: scientists use AI more than most other occupations. Specialized AI models have huge disciplinary coverage and are highly cited. Nearly half of the scientists surveyed report using some form of AI every day. 2) There is evidence that LLMs and specialized models act as *economic complements*: LLMs are used for general analysis, coding, and manuscript preparation, while specialized models push the frontier through domain-specific predictions, data generation and classification. The figure below illustrates this nicely: Each model class reinforces the other in the scientific process. 3) Surveyed scientists report large productivity gains from using AI: a saving of nearly 7 hours per week, time which is primarily reinvested in more research. Scientists also report having better access to interdisciplinary insights and increased capacity for synthesis. 4) But as some stages of scientific research become easier, bottlenecks shift downstream to non-automated tasks. Scientists report an increased backlog of untested hypotheses and substantial time spent on output verification. Half of scientists also report AI is pushing them towards safer, more incremental question, evidence of a potential “streetlight effect”. What do we make of this? Our findings suggest that AI holds significant potential to increase scientific productivity. However, as with other sectors, its ultimate impact will be governed by complex task interdependencies and investment into eliminating emerging bottlenecks. Investment into infrastructure and the organization of scientific production is necessary to realize AI’s full scientific potential for economic and societal gains. Finally, this is work in progress. There are many limitations, which we discuss in section 6 of the paper, at length. Link: ai.google/static/documents/A…