Hassabis Is Doing Oppenheimer Math at Google I/O

When J. Robert Oppenheimer watched the first atomic bomb test in July 1945, he recalled a line from the Bhagavad Gita: Now I am become Death, the destroyer of worlds. He was a scientist who had done the work and could not escape what the work meant.

Seventy-nine years later, Demis Hassabis stood at a microphone in Mountain View and told several thousand developers that they were standing in the foothills of the singularity. The line drew audible gasps from the room. He was not speaking metaphorically.

When we look back at this time, Hassabis said during the closing of Google I/O 2026 on Tuesday, I think we will realize that we were standing in the foothills of the singularity. (Fast Company)

The comparison is not exact, and Hassabis would likely resist it. Oppenheimer worked on a weapon designed to end a war. Hassabis is trying to build a machine that surpasses human intelligence across every domain — a goal he pegged at a 50 percent likelihood of arriving by 2030, plus or minus one year. But the shape of the discomfort is the same: the people most responsible for making something extraordinary happen are also the ones most acutely aware of what it might cost.

I call myself a cautious optimist, Hassabis told Axios cofounder Mike Allen at the conference. He also told Allen that AI will be roughly 100 times more impactful than the Industrial Revolution. (Business Insider) Both things can be true.

The I/O stage itself was evidence of a shift inside Google. Normally the closing remarks belong to Sundar Pichai, the CEO. This year Pichai gave the final segment to Hassabis — the researcher who cofounded DeepMind in 2010, who was acquired by Google in 2014, who now runs the combined DeepMind operation after its merger with Google Brain in 2023. (Times of India) Sergey Brin, who no longer runs day-to-day operations at the company, materialized onstage beside him as an unscheduled guest. (Axios) The two men placed different bets on the timing: Brin guessed just before 2030. Hassabis guessed just after. Both were smiling.

Hassabis has been building toward this moment for longer than most of his competitors have been in business. He started with games — AlphaGo famous victory over the world champion of Go was a proof of concept for agentic AI, the kind of system that can pursue a goal across many steps and learn from its own mistakes. Even our original Atari work — they were agents, he said at I/O. Maybe we were a bit ahead of our time.

The agents arriving now are different. Hassabis offered a personal example: he has been using AI to build mini video games late at night, projects that would have taken many months of engineering work a few years ago. The compressed timeline is not a demo. It is his actual workflow.

Gemini Diffusion, announced at I/O alongside the broader Gemini 3.5 model family, suggests why that workflow is now possible at all. It generates text at 1,479 tokens per second — roughly ten times the speed of typical autoregressive models — using a diffusion architecture that refines noise into coherent output the same way image generators work. The speed claim is verifiable against the benchmark table on DeepMind model page. Whether the quality holds up in production is a question the research community is already running tests to answer.

Also new: Deep Think mode, which pursues multiple reasoning approaches in parallel before committing to an answer. It is Google answer to the question of whether scaling alone is enough. Hassabis view, shared with Brin, is that both scale and new techniques are required. You need to scale to the maximum the techniques that you know about and exploit them to the limit, he said. And at the same time, you want to spend a bunch of effort on what coming next.

What is coming next, in Hassabis framing, is either the most productive partnership in human history or the last one humanity needs to have. He has said AlphaFold — the protein-folding model that DeepMind released in 2022 — was always the real model for what AI could do for science. Solve disease. Extend life. Reduce suffering. The singularity, for him, is not a horror scenario. It is the moment when those tools become capable enough to matter at scale.

The caveats are not small. AGI timelines from lab leaders have a consistent history of compression — things that were ten to twenty years away become five to ten years away and then, somehow, 2030, plus or minus one year. The singularity framing at a product conference is also a resource allocation argument: the stakes are high enough to justify the investment. That is not cynicism. It is how institutional ambition works.

But the thing about the Oppenheimer parallel that is hardest to dismiss is not the scale of the consequence. It is the quality of the ambivalence. Oppenheimer did not pretend the bomb was not terrifying. Hassabis is not pretending AGI is safe. He is arguing that the people best positioned to make it safe are the ones who understand it best — and that the understanding is not yet complete. That is a more honest position than most of what passes for AI risk discourse from the industry top earners.

The foothills comment landed because it was specific and because the room knew what he meant. They were not standing on a summit. They were not at the bottom. They were somewhere in between, looking up, not entirely sure what they were climbing toward.

That is probably the most accurate description of where Google DeepMind — and the industry — actually is.