At a recent Google I/O keynote, Demis Hassabis, the visionary CEO of Google DeepMind, made a profound declaration: we are currently “standing in the foothills of the singularity.” This statement alone was a showstopper, referencing the theoretical moment when artificial intelligence rapidly surpasses human intellect, fundamentally reshaping our world. Yet, what truly resonated was the specific context in which he uttered these powerful words.
Hassabis was on stage discussing scientific AI, having just showcased a video about WeatherNext, Google’s weather prediction software. The video highlighted WeatherNext’s crucial role in providing advance warnings for Hurricane Melissa’s devastating landfall in Jamaica last year, potentially saving lives. While this is an undeniable and significant achievement, helping people prepare for or escape a catastrophic storm is a far cry from an impending singularity.
This striking juxtaposition of Hassabis’s grand, futuristic rhetoric with the tangible, real-world utility of WeatherNext perfectly illustrates a core tension in the realm of AI for science. On one side, we have AI tools meticulously designed and trained to tackle specific scientific challenges. On the other, there’s the more ambitious vision of agentic, large language model (LLM)-based systems that could eventually conduct cutting-edge research autonomously, without human intervention.
The Shifting Landscape of AI in Science
This latter vision, of AI systems becoming independent researchers, fuels much of the current excitement in the AI community. It includes the fascinating concept of recursive self-improvement, where AI systems become the primary drivers of their own advancement, accelerating progress exponentially as they grow smarter. Indeed, agentic systems are already making genuine research contributions, sometimes with minimal human guidance.
Just recently, Pushmeet Kohli, Google Cloud’s chief scientist, articulated this shift in a piece for Daedalus, stating, “We are moving toward AI that doesn’t just facilitate science but begins to do science.” With the prospect of autonomous AI scientists on the horizon, it’s becoming harder to justify immense efforts solely on highly specialized tools. This is true even for groundbreaking systems like AlphaFold, which earned DeepMind scientists a Nobel Prize, or the life-saving potential of WeatherNext.
Such a future also hints at a far more extraordinary scientific landscape, one where humans and AI systems collaborate as peers, or where AI drives scientific progress entirely on its own. To be clear, Google is not entirely abandoning its work on specialized AI tools for science. They released AlphaGenome and AlphaEarth Foundations, trained for genetics and Earth science, last summer, and an updated version of WeatherNext arrived in November.
Furthermore, these specialized tools remain incredibly popular among the scientific community. Last year, Google reported that over three million researchers worldwide have utilized protein structure predictions from AlphaFold. Moreover, Isomorphic Labs, a Google subsidiary focused on drug development using AlphaFold technology, recently secured a staggering $2 billion in Series B funding.
From Tools to Teammates: Google’s AI Evolution
Despite the continued success of specialized tools, there are clear indications of a strategic realignment, both in terms of enthusiasm and resource allocation within Google. Last month, the Los Angeles Times revealed that Google fellow John Jumper, a Nobel laureate for his work on AlphaFold, is now focusing on AI coding rather than science-specific AI tools. This shift isn’t entirely surprising, as Google has recently faced reputational challenges regarding its coding tools compared to offerings from competitors like Anthropic and OpenAI.
However, this move may also signal a deeper prioritization of agentic science on Google’s part, given that robust coding abilities are fundamental to the success of many such autonomous systems. Across the wider industry, agentic researcher systems are demonstrating remarkable potential. OpenAI, for instance, recently announced that one of its models successfully disproved an important mathematics conjecture, hailed by some mathematicians as generative AI’s most significant contribution to the field thus far.
Crucially, the OpenAI model used in this mathematical breakthrough is not a specialized tool for math problems or even general research; it’s a general-purpose reasoning model, akin to GPT-5.5. If general agents can make independent contributions to complex mathematical research, their ability to do the same in scientific domains—though experimental verification presents a tougher challenge—could follow suit. Google is certainly dedicating substantial attention to an agent-driven scientific future.
The standout scientific announcement at I/O was the introduction of the new Gemini for Science package, which unifies several of the company’s LLM-based scientific systems under a single brand. This suite includes the hypothesis-generating AI Co-Scientist and the algorithm-optimizing AlphaEvolve, both of which are not yet publicly available but may soon see wider adoption as Google now allows any researcher to apply for access.
Early testers have expressed immense enthusiasm for their potential; Gary Peltz, a Stanford geneticist, likened using the AI Co-Scientist to “consulting the oracle of Delphi” in a Nature Medicine article. It’s important to note that Gemini for Science isn’t incompatible with specialized tools; in fact, agentic systems can be designed to seamlessly integrate and call upon such tools when needed. For instance, no agentic system can yet predict protein structures without AlphaFold’s expertise.
However, Google appears to be strategically shifting its public narrative, and at least some personnel and resources—like John Jumper—away from the exclusive development of these highly specialized tools. Even though it’s only been five years since AlphaFold revolutionized protein folding, both the underlying technology and the surrounding discourse have rapidly evolved beyond that once-groundbreaking achievement.
Beyond Collaboration: The Horizon of Autonomous AI
Google has carefully positioned this new generation of scientific agents as an accelerant for human scientists, rather than a replacement. The naming choice of “AI Co-Scientist” over “AI Scientist” seems quite deliberate in this regard. Hassabis himself employs this human-centric framing when discussing the evolution of scientific AI.
In a Daedalus interview, Hassabis shared his perspective: “For the next decade or so, we should think about AI as this amazing tool to help scientists.” He continued, “Beyond that timeframe, it is hard to say with any certainty, but perhaps these systems will become more like collaborators.” Yet, to truly be an effective scientific collaborator, an entity must also be a skilled scientist in its own right.
If Hassabis’s “foothills of the singularity” comment holds any predictive power, then AI scientists could eventually surpass the capabilities of their human counterparts. Hassabis himself was inspired to pursue AI after observing the stagnation in physics progress since the 1970s. He pondered whether human minds had reached their limits in that domain, and if AI could provide the breakthrough needed.
Superhuman agentic scientists would undoubtedly fit that bill, offering a path to overcome such barriers. While we may never reach such a point, Google seems to be resolutely charting a course towards that ambitious summit, aiming to redefine the very nature of scientific discovery.
Source: MIT Tech Review – AI
