DeepMind CEO: 'AGI' Is Coming Soon, But Here's the Test It Must Pass First - PCMag

In 2018, Elon Musk sent Sam Altman an email that has now become part of the record in the ongoing Musk v. OpenAI trial. Unfortunately, humanity's future is in the hands of Demis, he wrote, referring to Demis Hassabis, the cofounder and CEO of DeepMind. Reuters Seven years later, at Google I/O 2026, Hassabis gave the closing keynote and proposed a test for artificial general intelligence. The test, he said, should measure whether AI can do original physics.

Hassabis called it the Einstein Test. Train a would-be AGI on physics through 1901, he said, then see if it comes up with the same insights and discoveries that Albert Einstein began publishing in 1905. PCMag He was not speaking metaphorically. The foothills of the singularity are real, he told the audience. We are in a moment of immense promise but also enormous responsibility.

The Einstein Test is clever in ways that typical AI benchmarks are not. It is specific and falsifiable. No current system could pass it, and nobody has to take Hassabis's word on that. The cutoff date makes it verifiable. If you cut off training data at 1901 and the model produces special relativity, you have an answer. If it does not, you also have an answer. It also sidesteps the endless debate about what benchmarks actually measure. Passing bar exam variations tells you something about language task performance. Deriving the photoelectric effect from first principles tells you something about whether AI can reason its way to new physical insight.

It is also, by his own framing, something current systems cannot do. That is the point. Hassabis is proposing an unfalsifiable test in the present and using it to define what AGI competition means going forward. If you accept his framing, the question is not which company shipped the most features this quarter. The question is which laboratory is building toward genuine scientific reasoning.

AlphaFold is the credential he stands on. The protein-folding model, released in 2022, has been used by more than three million researchers across 190 countries and correlates with a 40 percent increase in submissions of novel experimental protein structures to public databases. PCMag Hassabis won a Nobel Prize for the underlying work in 2024, the first AI researcher to do so. That history is why the Einstein Test reads as aspiration rather than vapor.

OpenAI, meanwhile, has spent the last two months dismantling the division most directly analogous to what Hassabis is describing. The company's OpenAI for Science unit was dissolved in April, with its leader Kevin Weil departing and the work decentralized into other research teams. EdTech Innovation Hub Sora, OpenAI's video generation product that launched with significant fanfare, was discontinued on April 26. OpenAI Help Center The API will shut down in September. Bill Peebles, who led the Sora project, also departed.

These moves are not equivalent to each other. Sora was a consumer product. The science division was an explicit bet on AI as a tool for discovery. The timing matters. OpenAI gutted the part of its organization most directly comparable to what DeepMind is now positioning itself around, and it did so weeks before Hassabis used a keynote stage to stake that ground as his own.

Google's financial position makes the dynamic harder to ignore. Search advertising generated the majority of Alphabet's $402.8 billion in revenue last year. Reuters That profit engine funds DeepMind the way nothing at OpenAI or Anthropic currently funds their operations. Hassabis cited his version of the Einstein Test at I/O 2026 as the standard. He will have the resources to define it for as long as he can hold the frame.

The Einstein Test is not a product roadmap. It is not a capability claim. It is a definition of success that DeepMind is offering to an industry that has been arguing for years about what AGI actually means. Hassabis is asking the field to accept that scientific reasoning, the kind that produces new physics, is the real benchmark. If the field accepts that framing, DeepMind has the track record and the funding to be the arbiter of who is winning. If the field rejects it, the test becomes a curiosity and the Oppenheimer parallel becomes a punchline.

That is the bet. The foothills comment landed because the room understood what he was saying. They were being asked to accept that climbing matters more than the view from the top. Hassabis is betting that the climb is what Google DeepMind knows how to do.