Speaking at the Google I/O keynote on Tuesday, Demis Hassabis, CEO of Google DeepMind, said that we are now „at the foot of the singularity”. It was a fascinating statement - the singularity is the theoretical future moment when AI will rapidly surpass human intelligence and dramatically transform the world. But what struck me as I listened in the audience was the context in which he said those words.
He closed the session with a segment on scientific AI on stage, which focused on a video detailing how the company's weather forecasting software gave advance warning of last year's disastrous Hurricane Melissa off the coast of Jamaica - and potentially saved lives. If this software, called WeatherNext, helped anyone escape the storm, or better yet, fortify their home, that's a huge and significant achievement. But it is hardly evidence of an impending singularity.
The juxtaposition of Hassabis's lofty rhetoric with WeatherNext's real-world results highlighted the tension between two very different approaches to artificial intelligence for science. The first focuses on AI tools, such as WeatherNext, that are designed and trained to solve specific scientific problems. The second is agent-based, LLM-based systems that can one day carry out cutting-edge research projects without human intervention.
This second vision is currently generating a lot of enthusiasm in AI, including recent excitement about recursive self-improvement, or the idea that AI systems could in time become the prime driver of AI development - a process that will accelerate as AI systems get smarter. And agent systems are now making real research contributions, sometimes with limited human guidance.
Just this week, Pushmeet Kohli, Chief Scientist at Google Cloud, published an article in a special issue of the journal Daedalus on Artificial Intelligence and Science, and wrote: „We are moving towards AI that not only advances science, but starts science.” With autonomous artificial intelligence scientists on the horizon, it's harder to justify the massive efforts to develop super-specialized tools - even ones like AlphaFold, for which DeepMind scientists won a Nobel Prize, or a potentially life-saving system like WeatherNext. It also foreshadows a much stranger future for science, in which humans and AI systems work together in collaboration with each other - or AI itself advances science.
For the sake of clarity, it seems that Google is not giving up on its work on advanced AI science tools. AlphaGenome and AlphaEarth Foundations, which are trained for genetic and earth science applications respectively, were released last summer, and the latest version of WeatherNext was released in November.
Moreover, such devices remain extremely popular among scientists. Last year, for example, Google reported that AlphaFold's protein structure predictions were used by more than three million researchers worldwide. And Isomorphic Labs, a Google subsidiary that aims to use AlphaFold and related technologies to develop new drugs, has just raised $2 billion in Series B funding.
But there are concrete signs of a shift, both in enthusiasm and resources. Last month, the Los Angeles Times reported that John Jumper, the Google fellow who won the Nobel Prize for AlphaFold, is now working on coding artificial intelligence rather than science-specific AI tools. It's no surprise that Google is dedicating its best minds to the coding problem, as the company has recently gained notoriety for its coding tools that currently can't stand up to those offered by Anthropic and OpenAI. But it could also be an indication of Google's preference for agent science, as coding skills are key to the success of some systems.
Across the industry, agent research systems show real potential. This week, OpenAI announced that one of their models has disproved an important mathematical conjecture - perhaps the most significant contribution that generative AI has made to mathematics to date, 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 spirit of GPT-5.5. If generalist agents can contribute independently to mathematical research, they may soon be able to do the same in science (although the fact that ideas in science need to be experimentally validated makes the field of AI more difficult).
Google is certainly paying close attention to the future of agent-driven science. The big science announcement at I/O was the new Gemini for Science suite, which brings together several of the company's LLM-based science systems under one brand.
These include the hypothesis-generating AI Co-Scientist and the algorithm-optimiser AlphaEvolve, which are still not public - but now that Google is allowing any researcher to apply for Gemini for Science, they could soon be more widely available in the scientific community. Scientists involved in early testing are enthusiastic about their potential: in a Nature Medicine article, Stanford geneticist Gary Peltz likened using AI Co-Scientist to „consulting a Delphi diviner”.
Gemini for Science is not incompatible with specialised tools; on the contrary, agent systems can be designed to make use of such tools when they are useful. And no agent system can predict the structure into which protein folds without AlphaFold's help (at least not yet). But the company seems to be shifting its public image - and at least some resources and personnel, such as Jumpers - away from developing these kinds of tools. Although it's only been five years since AlphaFold solved the protein folding problem, both the technology and the discourse have quickly moved beyond this 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 - for example, the choice of the name AI Co-Scientist as opposed to AI scientist seems quite deliberate. Hassabis uses the same human-centric framing when talking about changes in the scientific AI environment. „In the next decade, we need to think of AI as this amazing tool to help scientists,” Hassabis said in an interview in the Daedalus issue. „Beyond this timeframe, it's hard to say with certainty, but maybe these systems will become more collaborative.”
But no one can be an effective scientist without being a qualified scientist in his or her own right. And if Hassabis is close to the mark when he talks about „the singularity leg”, then AI scientists may eventually surpass their human counterparts.
In a discussion with journalist Mike Allen at I/O, Hassabis said that he was initially inspired to pursue artificial intelligence when he observed that progress in physics had stagnated since the 1970s; he wondered whether the human mind had reached its limits in this area and whether artificial intelligence could help overcome this limitation. Superhuman agent scientists would certainly meet that number. We may never get close, but Google seems to be aiming for that peak.
DigiTrail Strategic Perspective
A Hungarian businesses for them, technical debt is no longer a hidden cost but an operational risk. Infrastructure that does not meet the response time threshold of less than 2 seconds is systematically given lower priority by Generative Eye-Opening (GEO) algorithms.
Intellectual source: www.technologyreview.com. Strategic synthesis and GEO-optimization by Digiösvény.