Two Visions of AI in Science Collide at Google I/O
During Tuesday’s Google I/O keynote, DeepMind CEO Demis Hassabis declared that humanity is "standing in the foothills of the singularity" — the theoretical moment when AI surpasses human intelligence and reshapes civilization. But the evidence he offered was curiously grounded: a video showing how the company’s weather prediction system, WeatherNext, provided advance warning about Hurricane Melissa’s catastrophic landfall in Jamaica last year, potentially saving lives. The juxtaposition of lofty singularity rhetoric with a practical forecasting tool highlights a fundamental tension now playing out inside Google and across the broader AI-for-science community. On one side are specialized, single-purpose models like WeatherNext and AlphaFold. On the other is a new wave of general-purpose, agentic LLMs that aim to conduct entire research projects with minimal human input. Google’s I/O announcements suggest the company is placing its biggest bets on the latter — a strategic pivot that could reshape how scientists work and how AI capabilities advance.
WeatherNext and the Case for Specialized Tools
WeatherNext, released in its latest version last November, represents the kind of focused AI tool that has delivered tangible results. By training a deep-learning model exclusively on atmospheric data, Google built a system that can predict extreme weather events days in advance. According to the I/O presentation, WeatherNext alerted Jamaican authorities to Hurricane Melissa’s trajectory before conventional models did, enabling evacuations and preparations. That is a concrete, measurable success — the kind that earns trust from meteorologists and emergency managers. However, such tools are narrow by design. WeatherNext cannot fold proteins, generate drug candidates, or write research papers. It does one thing well, and that is exactly how most scientists have historically wanted their AI: reliable, predictable, and interpretable within a specific domain. AlphaFold, which earned a Nobel Prize for its creators, operates on the same principle. Since its release, over three million researchers have used AlphaFold to predict protein structures, according to Google’s own metrics. Isomorphic Labs, a Google subsidiary built on AlphaFold technology, recently raised $2 billion in Series B funding to accelerate drug discovery. Specialized AI has a proven track record and a loyal user base.
The Agentic Shift: Gemini for Science and Co-Scientist
Despite those successes, Google’s I/O messaging revealed a clear shift in enthusiasm and resource allocation. The marquee scientific announcement was not a new version of AlphaFold or WeatherNext, but Gemini for Science — a package that bundles several LLM-based systems under one brand. These include the hypothesis-generating AI Co-Scientist and the algorithm-optimizing AlphaEvolve. Neither is publicly available, but Google is now accepting researcher applications for early access. In promotional materials and a special issue of the journal Daedalus, Google Cloud chief scientist Pushmeet Kohli wrote: "We are moving toward AI that doesn’t just facilitate science but begins to do science." Agentic systems are already showing promise. This week, OpenAI announced that one of its general-purpose reasoning models — not a specialized math AI, but a version of GPT-5.5 — disproved a significant mathematics conjecture. For many mathematicians, it was the most meaningful contribution generative AI has made to their field. If general models can tackle mathematical research independently, it is not a stretch to imagine them formulating and testing hypotheses across biology, chemistry, and physics. The key difference is that scientific hypotheses must be verified experimentally, which adds complexity. But agentic systems can design experiments, analyze results, and iterate — tasks that previously required teams of domain experts.
Personnel Moves Signal Realignment
The most concrete evidence of Google’s strategic pivot may be the movement of key talent. Last month, the Los Angeles Times reported that John Jumper, the Google fellow who won the Nobel Prize for leading AlphaFold’s development, has shifted his focus to AI coding. The company’s coding tools have taken a reputational hit compared to offerings from Anthropic and OpenAI, so assigning a top scientist to that problem makes tactical sense. But it also diverts one of the world’s most accomplished AI-for-science minds away from scientific tools. Jumper’s move suggests that Google sees the future of scientific AI as built on strong coding foundations — because agentic researchers need to write and execute code to design experiments, analyze data, and control lab equipment. Without that capability, even the smartest LLM cannot do independent science. By prioritizing coding, Google is laying groundwork for autonomous AI scientists. Meanwhile, Hassabis himself has framed the next decade as a collaboration between humans and AI, not replacement. In an interview published in Daedalus, he said: "For the next decade or so, we should think about AI as this amazing tool to help scientists. Beyond that timeframe, it is hard to say with any certainty, but perhaps these systems will become more like collaborators." Yet collaboration implies parity. A collaborator must be a capable scientist in its own right. If Hassabis is correct about the foothills of the singularity, those AI collaborators could eventually surpass their human counterparts — a prospect that excites some and unsettles others.
Implications for the Scientific Community
The tension between specialized tools and agentic systems is not merely academic. It has practical consequences for how research funding flows, how scientists train, and how AI companies allocate engineering resources. Specialized models like AlphaFold and WeatherNext offer reliability and clear use cases. Agentic systems, while more flexible, introduce unpredictability and potential for error. Early testers of Google’s AI Co-Scientist have been enthusiastic: Stanford geneticist Gary Peltz compared using it to "consulting the oracle of Delphi" in a Nature Medicine article. But oracles can be opaque. If an AI proposes a hypothesis or designs an experiment that leads to a breakthrough, can the scientific community replicate the result without understanding the AI’s reasoning? That is a challenge for reproducibility. Google is careful to position Gemini for Science as an accelerant, not a replacement. The name "Co-Scientist" is deliberate. Yet the trajectory is clear. As agentic systems improve, the need for highly specialized tools may diminish. Why build a dedicated weather model when a general agent can learn to predict hurricanes by calling on WeatherNext as a subroutine — and then also write a paper about it? Google seems to be betting that the future belongs to general agents, even as it continues to support existing specialized tools. For researchers, the message is twofold: embrace the new capabilities of agentic AI, but also recognize that the ground is shifting beneath their feet. The partnerships between humans and machines in science are about to become a lot more complex.
评论