Simply this week, Pushmeet Kohli, Google Cloud’s chief scientist, printed a bit in a particular AI and science challenge of the journal Daedalus, writing: “We’re transferring towards AI that doesn’t simply facilitate science however begins to do science.” With autonomous AI scientists on the horizon, it’s more durable to justify large efforts to develop super-specialized instruments—even one like AlphaFold, for which DeepMind scientists gained a Nobel Prize, or a probably life-saving system like WeatherNext. It additionally heralds a far stranger future for science, by which people and AI techniques collaborate as friends—or AI even makes scientific progress by itself.
To be clear, Google doesn’t seem like abandoning its work on specialised AI for science instruments. AlphaGenome and AlphaEarth Foundations, that are skilled for genetics and Earth science purposes respectively, have been launched final summer time, and the most recent model of WeatherNext got here out in November.
What’s extra, such instruments stay extraordinarily in style amongst scientists. Final yr, as an example, Google reported that protein construction predictions from AlphaFold have been utilized by over three million researchers worldwide. And Isomorphic Labs, a Google subsidiary that goals to make use of AlphaFold and associated applied sciences to develop new medication, simply raised a $2 billion Sequence B funding spherical.
However there are concrete indicators of realignment, in each enthusiasm and assets. Final month, the Los Angeles Instances reported that Google fellow John Jumper, who gained the Nobel for AlphaFold, is now engaged on AI coding, not on science-specific AI instruments. It’s not stunning that Google is assigning its finest minds to the coding downside, as the corporate has just lately taken a reputational hit as a result of its coding instruments don’t at the moment stand as much as these supplied by Anthropic and OpenAI. However it might additionally sign a prioritization of agentic science on Google’s half, as coding skills are key to the success of a few of these techniques.
Throughout the trade, agentic researcher techniques are displaying actual potential. This week, OpenAI introduced that one in every of their fashions had disproved an essential arithmetic conjecture—maybe essentially the most significant contribution that generative AI has made to arithmetic thus far, in keeping with some mathematicians.
Importantly, the mannequin utilized by OpenAI shouldn’t be specialised for fixing mathematical issues, and even for analysis; in keeping with the corporate, it’s a general-purpose reasoning mannequin within the vein of GPT-5.5. If common brokers could make unbiased contributions to mathematical analysis, they could quickly be capable of do the identical in science (although the truth that concepts in science should be verified experimentally makes it a more durable area for AI).
