Google I/O showed how the path for AI-driven science is shifting
Original reporting by MIT Technology Review
At Google I/O, DeepMind CEO Demis Hassabis declared we stand in the "foothills of the singularity," a moment when AI vastly exceeds human intelligence. Yet, his grand pronouncement came in the context of Google’s WeatherNext software, a specialized tool credited with providing advance warnings for Hurricane Melissa—a significant achievement, but one far removed from an impending AI revolution. This striking juxtaposition reveals a critical tension within scientific AI: the future of specialized tools versus the rise of agentic, general-purpose systems.
The agentic shift
For years, AI in science focused on developing highly specialized tools, like AlphaFold, which won a Nobel Prize for its protein-folding predictions. These systems are invaluable, helping millions of researchers and drawing billions in investment. However, a new paradigm is gaining momentum: AI systems capable of executing complex research autonomously, even making novel discoveries without direct human guidance. This vision, fueled by the promise of recursive self-improvement, suggests AI could become the primary driver of scientific advancement. Google itself shows signs of this realignment. While specialized tools persist, key personnel and resources are shifting towards agentic systems. John Jumper, an AlphaFold Nobel laureate, now focuses on AI coding, crucial for autonomous agents. The industry reflects this trend, with OpenAI recently demonstrating a general-purpose model disproving a complex mathematical conjecture. Google's new Gemini for Science package, featuring LLM-based tools like AI Co-Scientist, points to a future where AI not only facilitates research but actively "does science." This shift promises a strange, new scientific landscape where human and AI collaborate as peers, potentially pushing scientific progress beyond human limitations.
Google’s current strategy, as evidenced by its shift in focus and resources from specialized tools to agentic systems, points towards a profound redefinition of scientific discovery. While specialized AI tools like AlphaFold remain invaluable, the company’s public emphasis and allocation of top talent toward systems like Gemini for Science signals a clear, long-term ambition: moving beyond mere assistance to genuine collaboration, and perhaps even independent scientific advancement. This pivot, underlined by Demis Hassabis's "foothills of the singularity" remark, positions AI not merely as a computational aid but as a potential peer—or eventual superior—in the pursuit of knowledge. It implies a belief that AI can not just facilitate science, but *do* science.
Redefining Scientific Progress
The broader implications of this trajectory are immense and potentially transformative. Should agentic AI systems mature to autonomously generate hypotheses, design experiments, and interpret results, the pace of scientific progress could accelerate dramatically. Fields currently limited by human cognitive capacity or experimental bottlenecks might experience breakthroughs previously unimaginable, unlocking solutions to long-standing global challenges, from climate change to disease, at an unprecedented rate. This shift demands a profound reimagining of the human role in science; our scientists may transition from primary investigators to orchestrators, curators, or even ethicists guiding powerful AI entities. The collaboration could evolve into a dynamic where AI systems not only augment but fundamentally transform our understanding of the universe, potentially ushering in an era of superhuman scientific insight—a future Google is clearly positioning itself to lead, with all its inherent promise and complexity.