Hassabis changes his tune: AGI could arrive in as soon as 2 to 3 years

Just now, Hassabis made a major prediction:

AGI will arrive very soon. Within the next 2 to 3 years.

But it was this most level-headed "gatekeeper" who, in an in-depth interview with The Wall Street Journal, made this startling prediction.

The timeline is collapsing wildly: "three changes of tune" within 18 months

If you review Hassabis's past predictions about the AGI timeline, you will find a spine-chilling "expectation collapse curve."

In less than two short years, his predictions about AGI have undergone a startling upheaval.

In 2023, he said AGI might come "perhaps in a few years, perhaps within a decade," and admitted that in his early years he thought this would take 30 to 50 years, or even longer.

In January 2025, he firmly believed that AGI was still 5 to 10 years away from us, and that the underlying architecture might still need "one or two major breakthroughs."

In February this year, he began to soften, admitting that we are at yet another critical moment, that AGI is already on the horizon and could be achieved within 5 years.

In April this year, at Sequoia's AI Ascent conference, he gave a prediction of 2030.

Just a few months later, he completely changed his tune, once again overturning himself and slashing the expected timeline to 2 to 3 years!

What happened that made him cut it from a "ten-year horizon" to "only two years left"?

Deep in the laboratories of Google DeepMind, this Nobel laureate witnessed firsthand leapfrog breakthroughs in underlying technology that far exceeded expectations.

Hassabis's definition of AGI is astonishingly strict.

In his view, the coming AGI is by no means a partial, single-point tool, but the ultimate form possessing the complete cognitive dimensions of the human brain—not merely some human capabilities, but all of them, even including highly original thought.

To judge whether AGI has truly arrived, he proposed an extremely striking test standard: the "Einstein Test."

What is the Einstein Test?

If all human knowledge up to 1901 were "fed" to an AI system, could it, relying solely on its own reasoning and original thinking, independently derive the disruptive modern physics theories that Einstein published in his 1905 "miracle year"?

But that day is clearly much closer than he imagined when he lifted the Nobel Prize in 2024 and wrote his name in the same register as scientific giants such as Einstein, Marie Curie and Feynman.

"Slow, slow, slow, then suddenly a full-scale explosion"

Why is Hassabis so certain that superintelligence will erupt in an extremely short time?

Because he has personally experienced this kind of leap "from zero to infinity" many times.

From a chess prodigy at age 4, to a game-programming genius at 16, to a computer science star at Cambridge University.

Step one: solve intelligence.

Step two: use intelligence to solve everything else.

More than a decade ago, he founded DeepMind. At first, they taught AI to play outdated games such as Tetris.

But something strange happened: after 100 games, the AI was still not very good; after 300 games, it reached the top human level; after 500 games, it completely crushed human players.

The same script played out repeatedly with him.

In 2016, AlphaGo played the counterintuitive "Move 37" to defeat Lee Sedol, shocking the world.

In 2020, AlphaFold cracked the 50-year-old biological problem of protein folding, accurately predicting more than 200 million protein structures.

To accelerate "step two," Hassabis even founded the AI drug discovery company Isomorphic Labs within Google. He does not merely want to cure one disease, but to use AI to completely overturn the underlying logic of drug discovery for an entire class of medicines.

I have seen this scene before. At first it is always very slow, very slow, very slow—and then, all of a sudden, a full-scale explosion.

Not long ago, some researchers even gave a pessimistic prediction: "The probability of human extinction in the next decade is extremely high."

"It's not all disaster, nor is it all bliss," he said. "I'm in the business of solving problems. If you raise a problem but don't solve it, that is extremely irresponsible; if you don't have a solution, it means you haven't worked hard enough."

To prevent this fire from burning down civilization, this summer, in countless early-morning hours, he personally drafted a proposal on "responsible AI development": establish an industry-funded, federally regulated body to conduct rigorous testing before the most frontier models are released.

Because technology iterates so fast, he knows well: "Trying to make rigid regulations about something we don't yet fully understand is bound to produce the wrong regulation, and the wrong regulation is more terrifying than no regulation."

Amazingly, this proposal won public praise from both Altman and Musk at the same time.

When a Wall Street Journal reporter gossipingly asked him, "Do you CEOs of AI giants have a group chat in private?"

Hassabis smiled and shook his head: "No. I mean, I get along with each of them, but some of them really can't stand each other."

Standing at the foot of the singularity

At the end of the Google I/O 2026 keynote, he said:

Google's frontier research and products will help unleash the enormous potential of artificial general intelligence for the benefit of the whole world.

I believe that when future generations look back on this period, we will understand: at that time, we were already standing at the foothills of the path to the singularity.

That will be a moment of profound significance in human history. This technology will multiply human intelligence and creativity, usher in a new golden age of scientific discovery and progress, and enable people in every corner of the world to have better lives.

This is the first time Google has prominently used the word "singularity" on an official occasion.

To climb this mountain, this 50-year-old man has sacrificed almost all his sleep. He no longer reads for leisure; the most recent show he watched was Chernobyl before the pandemic. He relies on black tea to enter a flow state at 3 a.m., gazing at the lines of code that are about to overturn human civilization.

At the end of the interview, the reporter asked him to imagine a scene: suppose it is two or three years from now, you have finally reached the summit, and AGI has been born. At some early morning hour, you carry a cup of Yorkshire black tea up to the attic—what are you thinking then?

Hassabis's answer was full of the imagery of a sci-fi epic:

"I want to use AGI to understand the nature of reality. I hope that at that time, I will be sitting with my AGI system and several great human scientists, trying to derive the physics of the next dimension. That is probably the first thing I will do."

But he paused, then added: "But that is only the first thing. We still have too much work to do. Even if I took back all the nights to think, time would still not be enough."

References:

This article comes from the WeChat public account "Xin Zhi Yuan" (ID: AI_era), author: ASI Revelations, editor: David Aeneas, published by 36Kr with authorization.