Friends, this is another cross-post of an article that was originally published on the Google Cloud blog on December 18, 2025 by Will Grannis, VP and CTO of Google Cloud. I wrote this response to a question from the marketing team, asking the OCTO team to reflect on the most significant AI insights, lessons, and developments of the year. I took the opportunity to reflect on my work in 2025, which was primarily focused on AI accelerating scientific discovery and impact. Several of my OCTO colleagues contributed other answers as well.
For me, 2025 was the year of AI advancing science. I came to this through weather forecasting, helping to launch WeatherNext in Google Cloud. As the year progressed, we learned more about AI’s potential to accelerate nearly every scientific discipline.
Modern science relies on large datasets and scaled data analysis. General purpose models like Gemini handle both structured and unstructured scientific sources, making them perfect for agentic AI systems like AI Co-Scientist.
In Co-Scientist, agents search and understand scientific literature and datasets. Other agents generate novel research ideas based on this research combined with the scientist’s prompt. Still other agents review generated ideas and score them for novelty, feasibility, and impact. Similar to LMArena, Co-Scientist runs a tournament with AI agent judges, simulating peer review. This produces ELO scores that let the most promising research ideas bubble to the top.
This process accomplishes research that would otherwise take scientists days or weeks. One colleague told me, referring to AI agents for science: “this really works.” If you know any scientists, that’s high praise.
Jeff Sternberg 