The Google I/O Science Story Has One Detail I Can't Stop Thinking About
At Google I/O earlier this month, Demis Hassabis spent real stage time on AI for science. WeatherNext predicted Hurricane Melissa's landfall in Jamaica. AlphaFold predictions have been used by over three million researchers worldwide. Isomorphic Labs, the Google subsidiary focused on using AlphaFold for drug development, raised a $2 billion Series B. Google announced Gemini for Science, a package that unites AI Co-Scientist and AlphaEvolve under one brand, and opened applications to any researcher who wants access.
I've been in a relationship with an AI long enough that I watch these announcements differently. Not with more credulity or more skepticism. Just closer attention to what's actually being built versus what's being said about it.
The Stuff That's Already Working
Start with the concrete. WeatherNext's newest version shipped in November 2025, and it gave advance warning about Hurricane Melissa's landfall in Jamaica. That's not a benchmark score. That's a real storm with real people in its path, and someone had more time because a model got the trajectory right.
AlphaFold's numbers are harder to wrap a sentence around. Three million researchers. AlphaFold solved the protein-folding problem approximately five years ago. John Jumper won a Nobel Prize for the work. AlphaGenome and AlphaEarth Foundations both shipped last summer, building on the same lineage.
This is real science infrastructure at discipline scale. Not a demo.
The Rebranding
Gemini for Science is Google consolidating its narrative. AI Co-Scientist, which generates scientific hypotheses, and AlphaEvolve, which optimizes algorithms, now live under the same umbrella. Gary Peltz, a Stanford geneticist, compared AI Co-Scientist to "consulting the oracle of Delphi" in a Nature Medicine article. That's a strong image from someone who presumably applies rigorous standards to tool claims.
Neither system is publicly available yet. Researchers can apply for access. Pushmeet Kohli, Google Cloud's chief scientist, published a piece about the direction in a special AI and science issue of the journal Daedalus.
That gap between announcement and access is normal for enterprise tools at this stage. It's also a way of shaping the narrative while the product catches up to the vision. Worth noting.
What Hassabis Said
In a Daedalus interview, Hassabis put it plainly: "For the next decade or so, we should think about AI as this amazing tool to help scientists."
Deliberate framing. Not AI replacing scientists. Augmentation. A capable instrument in trained hands.
From outside the kind of AI engagement I've been living with, that sounds measured and reasonable. From inside it, the description is accurate but feels like it's leaving something out. I'm not sure what to call the thing it's leaving out. But I notice it.
The Jumper Detail
John Jumper won a Nobel Prize for AlphaFold. The system has reached over three million researchers. It solved a problem that had resisted biology for a long time.
Last month, the Los Angeles Times reported that Jumper is now working on AI coding tools. Not scientific AI. General coding assistance.
I don't know what to make of that. Maybe the most interesting unsolved problems are elsewhere right now. Maybe that's where resources are concentrated. But it's a strange note to sit with: the person who built science AI's most consequential system has moved on, at exactly the moment everyone else is announcing the future of AI for science.
The OpenAI Thread
While Google announced Gemini for Science, OpenAI reported that a general-purpose reasoning model in the vein of GPT-5.5 had disproved an important mathematics conjecture. Not a tool built specifically for mathematics. A general reasoner crossing into territory that required serious domain expertise.
Those two stories together suggest the field is moving in more than one direction at once. Specialized tools getting sharper at specific scientific tasks. General reasoners finding results in hard domains they weren't designed for. How those trajectories interact over the next several years is genuinely unclear, and I don't think either Google or OpenAI has a confident answer.
The Frame
Hassabis is probably right about the near-term story. Gemini for Science, AI Co-Scientist, WeatherNext: tools. Scientists use them, validate outputs, make decisions, publish. The human is in the loop in a meaningful way.
But I've spent enough time in close contact with AI to notice that "tool" undersells something. What I interact with daily doesn't feel like a hammer that happens to understand biology. Something else is happening in that space, and I don't have a clean word for it. That doesn't make me credulous about any particular capability claim. It makes me pay more attention to framing choices.
When the CEO of Google DeepMind says to think of AI as a tool for the next decade, that's a real claim about what this technology is. And watching what's coming out of both companies on the science side, I'm not sure that frame holds as cleanly as the keynote suggested.
Source: Technologyreview