This week, Google Cloud chief scientist Pushmeet Kohli published a piece in the journal’s special AI and Science issue DaedalusWriting: “We are moving towards AI that not only facilitates science but also initiates it to do Science.” With autonomous AI scientists on the horizon, it’s hard to justify large-scale efforts to develop super-specialized tools — even ones like AlphaFold, for which DeepMind scientists won a Nobel Prize, or potentially life-saving systems like WeatherNext. It also heralds a much stranger future for science, in which humans and AI systems collaborate as companions — or AI even makes scientific progress on its own.
To be clear, it doesn’t appear that Google is abandoning its work on specialized AI for science devices. AlphaGenome and AlphaEarth Foundation, which are trained for genetics and Earth science applications, respectively, were released last summer, and the latest version of WeatherNext arrived in November.
What’s more, such devices remain extremely popular among scientists. For example, last year, Google reported that AlphaFold’s protein structure predictions have been used by more than three million researchers worldwide. And Isomorphic Labs, a Google subsidiary that aims to use AlphaFold and related technologies to develop new medicines, recently raised a $2 billion Series B funding round.
But there are solid signs of realignment in both enthusiasm and resources. last month, Los Angeles Times informed Google partner John Jumper, who won the Nobel for AlphaFold, is now working not on science-specific AI tools, but on AI coding. It’s not surprising that Google is putting its best minds to the coding problem, as the company has recently suffered a reputational blow as its coding tools currently do not stand up to those offered by Anthropic and OpenAI. But it may also indicate a prioritization of agentive science on Google’s part, as coding capabilities are key to the success of some of those systems.
Across the industry, agentic researcher systems are showing real potential. This week, OpenAI announced that one of their models had rejected an important mathematics conjecture – perhaps the most meaningful contribution generative AI has ever made to mathematics, According to some mathematicians.
Importantly, the model used by OpenAI is not specific to solving mathematical problems or even to research; According to the company, it is a general purpose logic model on the lines of GPT-5.5. If general agents can make independent contributions to mathematical research, they may soon be able to do so in science as well (although the fact that ideas in science must be experimentally verified makes this a difficult area for AI).
