Google signals shift from specialized AI tools to autonomous agentic science systems

During Tuesday’s Google I/O keynote, Demis Hassabis, CEO of Google DeepMind, declared we are “standing in the foothills of the singularity.” Yet the real story behind his rhetoric is a strategic rebalancing within Google — and the broader AI industry — away from hyper-specialized scientific tools (like WeatherNext for hurricane prediction or AlphaFold for protein folding) and toward agentic, LLM-based systems that could one day conduct research autonomously. The tension was embodied in a single data point: John Jumper, the Nobel Prize winner for AlphaFold, has reportedly been reassigned to work on AI coding — a move driven by Google's need to catch up to Anthropic and OpenAI in generative coding tools. This shift, reported by the Los Angeles Times, signals that Google is pouring its best talent into building the foundational capabilities (coding, reasoning, and agency) that underpin the next generation of AI scientists.

OpenAI's recent breakthrough reinforces the trend: one of its general-purpose reasoning models (in the vein of GPT-5.5) independently disproved an important mathematics conjecture — arguably the most meaningful contribution generative AI has made to math so far. If general agents can tackle cutting-edge math, scientists believe they could soon do the same in experimental science. Google isn't abandoning specialized tools — last year it released AlphaGenome and AlphaEarth Foundations, and WeatherNext was updated in November. Isomorphic Labs, a Google subsidiary using AlphaFold for drug discovery, just raised $2B. But the center of gravity is shifting. As Pushmeet Kohli, Google Cloud's chief scientist, wrote this week: “We are moving toward AI that doesn’t just facilitate science but begins to do science.” That future, in which AI systems collaborate as peers — or even surpass humans — is arriving faster than the specialized-tool era suggests.