Demis Hassabis, CEO of Google DeepMind, declared from the Google I/O stage that humanity is "standing in the foothills of the singularity." The phrase, uttered as a video showed how WeatherNext anticipated Hurricane Melissa's landfall in Jamaica and potentially saved lives, reveals a growing tension inside the company: should Google keep investing in hyper-specialized AI tools for science, or bet everything on autonomous systems that will one day do science without human intervention?

The Big Picture

Hassabis's speech reflects a schism running through the entire industry. On one side, tools like WeatherNext or AlphaFold solve concrete problems with millimeter precision. On the other, agentic systems based on large language models (LLMs) promise to execute complete research projects, from formulating hypotheses to publishing results. This second vision fuels current enthusiasm, including the concept of recursive self-improvement: machines designing smarter machines in an accelerating cycle.

But concrete facts show a more complex reality. Just last week, Pushmeet Kohli, chief scientist at Google Cloud, published in the journal Daedalus that "we are moving toward AI that doesn't just facilitate science but begins to do science." Yet Google has not abandoned its specialized tools: AlphaGenome and AlphaEarth Foundations were released in summer 2025, and the latest WeatherNext version came out in November. Over three million researchers have used AlphaFold to predict protein structures, and Isomorphic Labs, Google's drug-discovery subsidiary, just closed a $2 billion Series B round.