The recent Google I/O keynote highlighted a shift in how AI is being applied to scientific research, focusing particularly on agentic systems. This change is exemplified by the introduction of Gemini for Science, a package aimed at integrating various LLM-based systems to enhance scientific inquiry.
Demis Hassabis, CEO of Google DeepMind, described this moment as “standing in the foothills of the singularity,” indicating a future where AI surpasses human intelligence. While this ambitious outlook generates excitement, practical applications like WeatherNext—used to forecast Hurricane Melissa's landfall—underscore the real-world benefits of specialized AI tools.
The contrast between these two approaches is striking. Tools like WeatherNext are specifically designed to tackle distinct scientific problems. In contrast, agentic systems are beginning to show their ability to conduct research with minimal human input. Pushmeet Kohli, Google Cloud’s chief scientist, recently stated, “We are moving toward AI that doesn’t just facilitate science but begins to do science.” This evolution raises questions about the future relevance of highly specialized AI tools in a landscape where autonomous AI agents might drive scientific discovery on their own.
Despite the excitement around agentic systems, Google remains dedicated to its existing AI tools. The impact of AlphaFold, for example, continues to resonate within the scientific community, with more than three million researchers relying on its protein structure predictions. Yet, signs suggest a strategic shift. John Jumper, a Nobel laureate recognized for his work on AlphaFold, has shifted his focus to AI coding, which may indicate Google’s prioritization of agentic systems where coding skills are crucial.
OpenAI recently demonstrated the capabilities of its general-purpose reasoning model, which successfully disproved a long-standing mathematical conjecture. This achievement showcases the potential for such models—like the upcoming GPT-5.5—to extend their influence into fields beyond their original design. If generalist AI can achieve breakthroughs in mathematics, similar progress could be expected in scientific research, despite the complexities involved.
Central to Google’s new direction is the Gemini for Science initiative, which unites several LLM-based systems under one framework. The AI Co-Scientist and AlphaEvolve are intended to work collaboratively, and early testers have shown optimism about their capabilities. One Stanford geneticist likened the AI Co-Scientist to “consulting the oracle of Delphi,” reflecting the perceived value of these systems in supporting researchers.
While Google continues to refine specialized tools, the focus is clearly shifting toward a future where AI serves as a collaborator rather than just an assistant. Hassabis articulated this vision, suggesting that for the next decade, AI should be seen as a powerful asset for scientists. However, he also recognized the possibility that these systems could evolve into more autonomous partners beyond that timeframe.
As AI technologies advance, the emergence of superhuman agentic scientists raises important questions about the future of research. Hassabis’s comments on the stagnation of progress in physics since the 1970s present a compelling narrative: could AI help overcome human limitations in scientific exploration?
Google’s trajectory suggests a growing alignment with agentic AI systems capable of making meaningful contributions to scientific discovery. As Gemini for Science prepares for a wider release among researchers, the scientific community may soon find itself at a crucial juncture, adapting to a new era of AI-driven collaboration.
The stories that move AI & crypto markets — before the market reacts.
Free. 7am ET. Five stories. 62,400 readers.
