The promise of artificial intelligence in the laboratory has long been framed as a quest for better tools: systems that can predict a protein’s structure or map the path of a hurricane with unprecedented accuracy. However, recent developments suggest that the industry is pivoting toward a more radical ambition. During Tuesday’s Google I/O keynote, Demis Hassabis, the CEO of Google DeepMind, positioned this evolution in grand terms, declaring that we are currently “standing in the foothills of the singularity.” While the term often evokes science-fiction scenarios of machine intelligence outpacing human control, the immediate shift is more practical: a move away from building specialized, one-off software toward creating autonomous, agentic systems capable of conducting research independently.
This shift is detailed in the report from MIT Technology Review, which notes a distinct tension between the legacy of "narrow" AI and the rise of general-purpose scientific agents. The headlines following Google’s presentation emphasized the life-saving potential of the company’s weather prediction software, WeatherNext, which provided an advance alert about Hurricane Melissa’s catastrophic landfall in Jamaica last year. Yet, while such tools are undeniably impactful, they represent the "tool-based" paradigm—software designed to solve a singular, predefined problem. The newer, agentic approach aims for something far more expansive: AI that functions not just as a calculator, but as an investigator.
This transition is not merely theoretical. Pushmeet Kohli, Google Cloud’s chief scientist, recently articulated this change in a special issue of the journal Daedalus, stating that we are moving toward AI that does not just facilitate science but begins to do science itself. This philosophy underpins the new Gemini for Science package, which integrates hypothesis-generating systems like AI Co-Scientist and algorithm-optimizing tools like AlphaEvolve. The appeal of such systems is significant; Gary Peltz, a geneticist at Stanford University, has already likened the experience of using the AI Co-Scientist to “consulting the oracle of Delphi.”
Despite the excitement, there are clear limitations to consider. The transition to agentic AI faces a fundamental hurdle that specialized tools do not: the requirement of physical verification. While an AI can hypothesize or optimize code, scientific progress is ultimately tethered to the empirical world. Furthermore, the strategic realignment of talent suggests that Google is hedging its bets. Reports indicate that John Jumper, the Nobel Prize-winning scientist behind AlphaFold, has shifted his focus toward AI coding. While this move addresses the company’s need to compete with firms like OpenAI in the coding space, it also highlights the critical importance of general-purpose programming abilities for the next generation of autonomous research agents.
The data suggests that the specialized "tool" era is far from over, even as the focus shifts. Last year, Google reported that over three million researchers worldwide utilized protein structure predictions from AlphaFold. Additionally, Isomorphic Labs, a Google subsidiary focused on drug discovery, recently secured a $2 billion Series B funding round, proving that there remains immense market appetite for systems that solve specific, high-value biological problems. These tools continue to receive updates, such as the release of AlphaGenome and AlphaEarth Foundations last summer and the latest iteration of WeatherNext in November.
The trajectory of this technology will be defined by whether these systems remain human-centric "co-scientists" or evolve into autonomous peers. Hassabis has suggested that for the next decade, AI will likely remain an "amazing tool" for human researchers. Whether these systems can eventually overcome the stagnation in scientific fields like physics—a goal that originally inspired Hassabis to enter the field—remains the ultimate, if distant, test. The next reading of the adoption rates for the Gemini for Science package, which is now open for researcher applications, will provide an early signal as to whether the scientific community is ready to move from using AI as a calculator to trusting it as a collaborator.
