Google I/O: Scientific AI Between Promises and Concrete Realities

At the Google I/O conference, Demis Hassabis, CEO of Google DeepMind, made waves by stating that we are "at the foot of the hills of singularity." This theoretical concept describes a moment when artificial intelligence would surpass that of humans, radically transforming our world. This statement was made during a presentation on scientific AI, illustrated by Google's WeatherNext software, which accurately predicted Hurricane Melissa in Jamaica last year, potentially saving lives.

The contrast is striking between the grand ambitions of AI and the concrete results achieved by tools like WeatherNext. On one side, there are specialized AI systems designed to solve specific scientific problems. On the other, there are large language models (LLMs) that could one day conduct advanced research without human intervention. The latter approach generates significant excitement, particularly around the idea of recursive self-improvement, where AI could become the primary driver of its own development.

Pushmeet Kohli, Chief Scientist at Google Cloud, recently wrote in the journal Daedalus that AI is beginning to do science, rather than just facilitating it. With the emergence of autonomous AI scientists, interest in developing hyper-specialized tools may diminish. Nevertheless, tools like AlphaFold, which revolutionized protein structure prediction, remain widely used, with over three million researchers relying on its predictions last year.

Google continues to develop specialized tools, such as AlphaGenome and AlphaEarth Foundations, but a shift in priorities seems to be emerging. John Jumper, a researcher recognized for AlphaFold, is now working on AI coding, an area where Google is looking to catch up with competitors like Anthropic and OpenAI. It is not surprising that Google is assigning its top minds to the coding problem, as the company recently faced a blow to its reputation due to its coding tools that currently do not compete with those offered by Anthropic and OpenAI.

Agentic research systems show promising potential. This week, OpenAI announced that one of their models had disproved an important mathematical conjecture—perhaps the most significant contribution that generative AI has made to mathematics so far, according to some mathematicians. It is important to note that the model used by OpenAI is not specialized for solving mathematical problems, nor even for research; according to the company, it is a general-purpose reasoning model in the vein of GPT-5.5.

If general agents can make independent contributions to mathematical research, they may soon be able to do the same in science (although the fact that ideas in science must be experimentally verified makes this field more challenging for AI). Google is certainly paying a lot of attention to a future of science driven by agents. The major scientific announcement at I/O was the new Gemini for Science package, which unites several of the company's LLM-based scientific systems under a single brand.

This includes the hypothesis-generating AI Co-Scientist and the algorithm optimizer AlphaEvolve, which are still not publicly available—but as Google now allows any researcher to request access to Gemini for Science, they may soon see broader adoption within the scientific community. Scientists involved in preliminary testing are enthusiastic about their potential: Gary Peltz, a geneticist at Stanford, compared using the Co-Scientist AI to "consulting the oracle of Delphi" in a Nature Medicine article.

Gemini for Science is not incompatible with specialized tools; on the contrary, agentic systems can be designed to call upon such tools when they might be useful. And no agentic system can predict the structure in which a protein will fold without the help of AlphaFold (at least not yet). But the company seems to be shifting its public image—and at least some resources and personnel, like Jumper—away from the specific development of these types of tools. Although only five years have passed since AlphaFold solved the protein folding problem, both the technology and the discourse have quickly evolved beyond that once-revolutionary achievement.

Google has been careful to position this new set of scientific agents as an accelerator for human scientists, rather than a replacement for them—the choice of the name AI Co-Scientist versus AI Scientist, for example, seems quite deliberate. Hassabis uses this same human-centered framework when discussing changes in the landscape of scientific AI. "For the next decade, we should think of AI as this incredible tool to help scientists," Hassabis said in an interview published in the Daedalus issue. "Beyond that period, it’s hard to say for sure, but perhaps these systems will become more like collaborators."

But no one can be an effective scientific collaborator without also being a fully competent scientist. And if Hassabis is close to the truth when he speaks of "the foot of the hills of singularity," then AI scientists could eventually surpass the capabilities of their human counterparts.

In a discussion with journalist Mike Allen during I/O, Hassabis mentioned how he was initially inspired to pursue AI by observing how progress in physics had stagnated since the 1970s; he wondered if the human mind had reached its limits in this field, and whether AI could help overcome that barrier. Superhuman agentic scientists would certainly fit that description. We may never get close, but Google seems to be aiming for that summit.