Google I/O Reveals a Shifting Path for AI-Driven Science

Google I/O’s recent keynote saw Demis Hassabis, CEO of Google DeepMind, declare that humanity stands in the "foothills of the singularity"—a theoretical future where AI surpasses human intelligence. This striking pronouncement, however, was delivered within a context that highlighted a fascinating tension. Hassabis was closing a segment on scientific AI, showcasing WeatherNext, Google’s weather prediction software, which provided an early warning for Hurricane Melissa, potentially saving lives in Jamaica. While a significant achievement, this practical application seemed a world apart from the grand vision of an impending singularity. This juxtaposition underscores a fundamental divergence in the approach to AI for scientific advancement. On one hand, there are highly specialized AI tools like WeatherNext, or even AlphaFold—for which DeepMind scientists won a Nobel Prize—designed and trained to solve specific scientific problems with remarkable precision. On the other, a growing enthusiasm surrounds agentic, large language model (LLM)-based systems. These systems are envisioned to autonomously execute cutting-edge research, potentially even driving AI advancement through recursive self-improvement, a process that accelerates as AI grows smarter. As Pushmeet Kohli, Google Cloud’s chief scientist, noted, we are moving towards AI that "doesn’t just facilitate science but begins to do science." Concrete signs of a strategic realignment within Google, in terms of both enthusiasm and resource allocation, are becoming apparent. Reports indicate that John Jumper, the Google fellow and Nobel laureate behind AlphaFold, has shifted his focus to AI coding rather than developing science-specific AI tools. This move, while partly driven by Google's need to bolster its coding tools against competitors like Anthropic and OpenAI, also strongly suggests a prioritization of agentic science, given that robust coding abilities are crucial for such autonomous systems. This trend is not unique to Google; OpenAI recently announced that one of its general-purpose models disproved a significant mathematics conjecture, marking a substantial contribution from generative AI to the field. Google is clearly dedicating considerable attention to an agent-driven scientific future. The major scientific announcement at I/O was the introduction of the new Gemini for Science package, which consolidates several of the company’s LLM-based scientific systems under a unified brand. This suite includes the hypothesis-generating AI Co-Scientist and the algorithm-optimizing AlphaEvolve. While not yet publicly available, Google is now accepting applications for access, hinting at their imminent wider adoption within the scientific community. Early testers, such as Stanford geneticist Gary Peltz, have expressed profound enthusiasm, likening the experience of using AI Co-Scientist to "consulting the oracle of Delphi." It's important to note that Gemini for Science and other agentic systems are not inherently incompatible with specialized tools; indeed, they can be designed to leverage such tools when beneficial. For instance, no agentic system can currently predict protein structures without AlphaFold's assistance. However, Google's public image and, crucially, its allocation of resources and personnel, are visibly shifting away from the sole development of these highly specialized instruments. Just five years after AlphaFold revolutionized protein folding, both the underlying technology and the broader scientific discourse have rapidly evolved. Google carefully positions this new generation of scientific agents as accelerants for human scientists, rather than replacements, a sentiment underscored by the choice of "AI Co-Scientist" as a name.