On Tuesday at Google I/O, Demis Hassabis put Google’s scientific AI pitch in unusually grand terms: humanity is “standing in the foothills of the singularity,” he said, moments after highlighting a weather model that warned of Hurricane Melissa’s landfall in Jamaica last year.
That contrast is the story. Google’s science strategy is no longer just about building specialized tools like WeatherNext or AlphaFold. It is increasingly about agentic systems that may generate hypotheses, optimize algorithms, and push research forward with less human direction, according to MIT Technology Review.
Tuesday’s I/O stage turned weather forecasting into a singularity argument
Hassabis’ “foothills of the singularity” line landed because the example on stage was not a self-improving superintelligence. It was WeatherNext, Google’s weather prediction software, shown in a video about an advance alert before Hurricane Melissa’s catastrophic landfall in Jamaica last year.
If that warning helped people leave danger or protect homes, the achievement is concrete. But it is not, by itself, proof that AI is about to outrun human intelligence.
That gap exposes the tension inside AI-driven science:
| Approach | Example from the source | Core promise | Constraint |
|---|---|---|---|
| Specialized scientific AI | WeatherNext, AlphaFold, AlphaGenome, AlphaEarth Foundations | Solve defined scientific prediction problems | Needs domain-specific design and validation |
| Agentic scientific AI | AI Co-Scientist, AlphaEvolve, OpenAI’s general reasoning model | Generate hypotheses, optimize methods, contribute to research | Harder to verify, especially outside math |
The sharper reading: Google used I/O to show that science is becoming one of the strongest public arguments for frontier AI. Consumer tools can impress. Scientific tools can justify ambition.
That distinction matters as Google also pushes agentic AI into mainstream products. Its Search team said AI Mode is rolling out in the U.S. and uses a custom version of Gemini 2.5 for AI Mode and AI Overviews, while Deep Search can issue hundreds of searches and create a cited report in minutes, according to Google. For readers tracking the consumer side of that push, MLXIO has separately covered Google I/O’s Gemini race, Google Sparks Search Revolution with Gemini 3.5 Flash AI, Cheap AI Agents: Google’s Gemini 3.5 Flash Bets Big, Gemini Mac agent, and Fara1.5 AI web tasks.
From AlphaFold’s Nobel halo to Gemini for Science’s agentic bet
Google is not abandoning specialized scientific AI. The source explicitly says AlphaGenome and AlphaEarth Foundations were released last summer, and the newest version of WeatherNext came out in November.
The company’s record here is not rhetorical. AlphaFold changed structural biology enough that DeepMind scientists won a Nobel Prize. Google reported last year that AlphaFold’s protein structure predictions had been used by over three million researchers worldwide. Isomorphic Labs, the Google subsidiary using AlphaFold and related technologies for drug development, raised a $2 billion Series B funding round.
Those are not keynote vibes. They are adoption signals.
But the center of gravity appears to be moving. MIT Technology Review points to the Los Angeles Times report that John Jumper, who won the Nobel for AlphaFold, is now working on AI coding rather than science-specific AI tools. The source frames that as partly unsurprising: Google’s coding tools have taken a reputational hit against those from Anthropic and OpenAI. It also may matter for science, because coding ability is central to some agentic research systems.
The I/O science announcement fits that shift. Google introduced Gemini for Science, a package bringing several LLM-based scientific systems under one brand, including AI Co-Scientist and AlphaEvolve. They are not publicly available yet, but Google is allowing any researcher to apply for access.
That is a platform move, not just a model demo.
The data says specialized AI still has the strongest proof points
The hard numbers in the source still favor domain-specific systems.
AlphaFold has the cleanest evidence of reach: over three million researchers worldwide using its predictions. Isomorphic Labs has the clearest capital signal: a $2 billion Series B tied to AlphaFold-related drug-development technology. WeatherNext has the clearest public-interest use case: an alert tied to Hurricane Melissa’s landfall in Jamaica last year.
Agentic science has a different kind of proof. It is more provocative, but less settled.
OpenAI announced this week that one of its models disproved an important mathematics conjecture. MIT Technology Review says some mathematicians described it as perhaps the most meaningful contribution generative AI has made to mathematics so far. The key detail: OpenAI’s model was not built specifically for mathematics or research. It was a general-purpose reasoning model in the vein of GPT-5.5.
That is the strongest evidence for Google’s strategic pivot. If general reasoning agents can produce real mathematical contributions, then science-specific AI may become only one part of the stack. Agents could call specialized tools when needed, then stitch results into broader research workflows.
“We are moving toward AI that doesn’t just facilitate science but begins to do science.”
That line came from Pushmeet Kohli, Google Cloud’s chief scientist, in a special AI and science issue of Daedalus. It captures the shift more precisely than the singularity line. The near-term question is not whether AI becomes omniscient. It is whether AI systems can move from assisting scientists to performing pieces of scientific work.
Google is choosing “co-scientist” language for a reason
Google’s naming is careful. The company calls its hypothesis-generating system AI Co-Scientist, not “AI Scientist.”
That framing keeps humans in charge. It also lowers the temperature around autonomy. Hassabis used similar language in the Daedalus issue:
“For the next decade or so, we should think about AI as this amazing tool to help scientists,” Hassabis said. “Beyond that timeframe, it is hard to say with any certainty, but perhaps these systems will become more like collaborators.”
The word “collaborators” does a lot of work. A collaborator is not a calculator. A collaborator proposes, critiques, searches, tests, and revises. In science, that also means being wrong in productive ways and being checked by experiment, peer review, and reproducibility.
This is where math and science diverge. A model disproving a conjecture can be assessed within formal structures. A model proposing a biological hypothesis still needs wet-lab validation. An AI-generated drug lead does not become medicine because a model says it is promising.
MLXIO analysis: Google’s careful language suggests it wants the upside of autonomous research without triggering the full burden of claiming autonomous science. That balance may get harder to maintain if these systems keep producing stronger outputs.
The next credibility test is validation, not keynote language
The most important post-I/O milestone is not another phrase about the singularity. It is whether Gemini for Science, AI Co-Scientist, and AlphaEvolve produce repeatable results that researchers outside Google can inspect, test, and build on.
Early testers are enthusiastic. Gary Peltz, a Stanford geneticist, compared using AI Co-Scientist to “consulting the oracle of Delphi” in a Nature Medicine article. That is praise, but it is also revealing. Oracles are powerful because they suggest answers. Science still has to prove them.
The thesis to watch: Google appears to be shifting its science narrative from specialized breakthroughs to general agents that can coordinate discovery. Evidence that would strengthen that thesis includes wider researcher access to Gemini for Science, published results from AI-generated hypotheses, and more personnel or product emphasis moving toward agentic systems. Evidence that would weaken it would be continued dominance of narrow tools like AlphaFold and WeatherNext, with agents remaining useful wrappers rather than original contributors.
Google’s I/O message was not simply that AI will help scientists work faster. It was that science may become the proving ground for claims about transformative AI. That is a higher bar than a keynote can clear.
Impact Analysis
- Google is positioning scientific breakthroughs as a key justification for frontier AI development.
- The shift from specialized tools to agentic systems could change how research is conducted and validated.
- Real-world successes like weather warnings show promise, but broader claims about autonomous scientific progress remain harder to prove.
