The Man Who Won a Nobel Prize for AI Just Said AGI Is Four Years Away. Here’s Everything Demis Hassabis Said This Week.

The Man who won a Nobe Prize

On stage at Google I/O, then in three separate interviews, the Google DeepMind CEO laid out his most specific and urgent views on AGI, curing disease, the singularity, and what humans should actually be doing right now.

Two days ago, Demis Hassabis sat down with The Rundown AI for an exclusive video interview, the one you were linked to and said interesting things about AGI and how AI is progressing.

He had just closed Google I/O 2026 from the main stage in Mountain View with seven words that became the most-discussed line of the entire three-hour conference: “We’re at the foothills of the singularity.”

Then he did three more interviews in the hours and days that followed. With Axios, with Fast Company, with Semafor. Each one added more detail. Taken together, it’s the clearest picture Hassabis has ever given publicly of where he thinks we are, where we’re heading, and what he’s actually worried about.

This is a full breakdown of everything he said, across all of those conversations.

Who is Demis Hassabis and why does his opinion carry weight here?

Before getting into the substance, this part is worth stating clearly for anyone who isn’t already deep in the AI world.

Hassabis is not a commentator or an investor making predictions from the sidelines. He co-founded DeepMind in 2010 with Shane Legg and Mustafa Suleyman, after completing a PhD in cognitive neuroscience at University College London. Google acquired DeepMind in 2014, and in 2023, Hassabis merged DeepMind with Google Brain to create Google DeepMind, the combined research division he now leads.

In 2016, his team’s AlphaGo became the first AI system to defeat the world champion at Go, a game long considered AI-proof because of its astronomical complexity.

In 2020, AlphaFold 2 solved the protein folding problem, a challenge that had stumped biologists for 50 years.

In October 2024, Hassabis shared the Nobel Prize in Chemistry for that work, the first time the Nobel committee had honored AI research so directly. He’s also the person who originally built the early Atari-playing agents that became the conceptual foundation for modern reinforcement learning.

When he says something about AGI timelines, he is the researcher most responsible for the pace at which we’re getting there. His opinion is therefore, at minimum, unusually informed.

On AGI: “2030, plus or minus a year”

This is the headline from The Rundown interview, and Hassabis was more specific about it than he has been before.

He said AGI is on track for around 2030, with 2029 now looking possible. That’s not a dramatic departure from what he’s said previously, but the way he framed it has shifted. He described a growing confidence that the industry has found the right technical path. Not certainty, he was careful about that, but directional conviction that wasn’t there two or three years ago.

The specific quote from his Axios conversation:

“We can see agents really happening now and imagine what they will be in another year, and how useful they’ll be.”

He also said that the wave of AI agents we’re experiencing right now which include Claude Code, Codex, Gemini Spark, autonomous workflows all should not be seen as the destination. In his words:

“You can imagine the agentic era in this next year is a little bit like a practice run.”

A practice run for what? For systems that are significantly more capable. The agents we have now are early, limited, and prone to the exact kinds of failures documented in production, the database deletions, the fabricated policies, the supply chain attacks. That’s his point. We are building the infrastructure and the intuitions for running AI systems in the world. The systems themselves are going to keep getting stronger.

What’s still missing between now and AGI

Hassabis named four specific unsolved problems that stand between current AI and what he’d call genuine AGI.

World physics: Current models don’t have a robust, generalized understanding of how the physical world works like how objects move, how forces interact, how cause and effect play out in three-dimensional space. Language models reason about physics in text. They don’t have the embedded, intuitive grasp of physical dynamics that even a toddler develops from playing with objects. This is why robotics remains genuinely hard even as language and coding capabilities have exploded.

Memory: Modern AI systems don’t retain meaningful memory across sessions unless you explicitly engineer mechanisms to give them that memory. Every conversation begins fresh. Every agent session, unless someone builds a harness to preserve context, starts from zero. Hassabis has talked about this before in other contexts, the Karpathy-style knowledge base workflow is essentially a band-aid on this exact problem. But at the model architecture level, truly persistent, selective, associative memory, the kind humans have, isn’t solved.

Consistency: AI systems behave differently across identical or near-identical inputs. Ask the same question twice and you may get different answers. Deploy the same agent on two machines and you may get different decisions. That inconsistency which is baked into the stochastic nature of how these models work creates serious problems for systems that need to be reliable and predictable.

Continual learning: Once a model is trained, it is largely frozen. It doesn’t keep learning from its daily interactions the way humans do. You can fine-tune it, you can add information through context, but the underlying model weights don’t update continuously as it operates. True AGI, in Hassabis’s framing, would need to keep learning, to accumulate experience and update its understanding in real time, the way an intelligent agent in the world naturally would.

On the singularity, what he actually meant

The phrase “foothills of the singularity” got the most attention from Google I/O, and Hassabis addressed directly why he chose it in his Semafor interview.

“We debated it back and forth,” he said. “I was closing, and I wanted to be authentic about what I’m thinking with AGI.”

When pressed on what he meant by the singularity specifically, he clarified: “The singularity, at least my interpretation of that word and that term, means the era that we’re in.”

He’s not predicting a sudden, dramatic inflection point where machines become infinitely capable overnight. He’s describing a trajectory, one where we are clearly at the beginning of something whose end state will be unrecognizable from the starting point.

In his Axios conversation, he added the political dimension explicitly. He chose provocative language deliberately. “This is partly why I use some of the terms I used, yeah, which were a little bit provocative.”

The goal wasn’t to create hype. It was to force urgency into conversations happening outside the tech industry, among economists, policymakers, and government officials who Hassabis thinks are still not treating this seriously enough.

His framing on *Anthropic’s Project Glasswing* is relevant here. When Glasswing revealed that a model had autonomously found thousands of previously unknown vulnerabilities in major operating systems, including a 27-year-old bug in OpenBSD that had survived decades of human review, Hassabis pointed to it as exactly the kind of shock the broader world needs to register.

He called it “probably a good warning shot across the bow.” The lesson he wants people to take: these systems are already doing things that most of the world doesn’t realize are possible, and the capabilities are still climbing.

On recursive self-improvement: the next threshold

This came up specifically in the Axios interview and it’s the part of the conversation that carries the most long-term weight.

Recursive self-improvement is the scenario where an AI system becomes capable of meaningfully accelerating its own development ,of being a significant contributor to building the next generation of AI. Not just writing code efficiently, but making the research decisions, identifying the architectural improvements, running the experiments that make future models better.

All the leading labs are quite focused on that. There’ll be clear gains in terms of speed of your research. But there are also risks with that type of system.

He was careful to draw a line between where we are now and that threshold. “I think what we’re seeing is soft self-improvement, in the sense of these coding agents are making engineers much more productive.” That’s a real and significant thing, using AI to accelerate AI research, which is already happening at every major lab.

But it’s not the same as systems that improve their own weights autonomously. We’re not there yet. The pace of development is accelerating, but it’s still primarily humans, assisted by AI tools, making the core research decisions.

if recursive self-improvement happens at any meaningful scale, the gap between the 2029 prediction and something much sooner narrows dramatically. Hassabis is aware of this, which is probably why his timeline includes the hedge “or even sooner.”

On disease: “hundreds of diseases-that’s the aim”

Isomorphic Labs, the biotech company Hassabis founded and spun out of DeepMind in 2021, is built on the same AI-first approach that produced AlphaFold.

He described its mission as not being about any one particular drug or one particular disease. The goal is to build an AI system capable of transforming the drug discovery process itself, making it, in his words, “1,000 times more efficient.”

In The Rundown interview, he specified where the timelines are hardening: oncology and immunology first.

These are the fields where the AI-assisted pipelines are most advanced, where Isomorphic Labs has active preclinical work. The broader goal — “an engine that could help cure any disease” is further out. But the near-term goal, starting with cancer and immune system diseases, has concrete timelines attached to it.

His position at Semafor was even more direct. Don’t think of Isomorphic as a drug company. Think of it as an AI company applying its capabilities to medicine. The advantage isn’t proprietary clinical data it’s the models themselves, pushed harder and further than they’ve been pushed in the biomedical space before. In his words: “pushing the models a lot harder.”

The broader context: Isomorphic has existing partnerships with Eli Lilly and Novartis. AlphaFold’s protein structure predictions have already been used by more than 2 million researchers globally. The pipeline from AlphaFold to drug candidates to clinical trials is still long, but it’s no longer theoretical.

what humans should do with themselves

This is where Hassabis’s thinking is probably most useful for people who aren’t building AI systems but who are trying to figure out how to navigate a world in which AI is becoming more capable very fast.

In The Rundown interview, he talked about what he can’t wait to see, the things students and the next generation will build with advanced AI that we can’t anticipate. He was specific about which human qualities will become more valuable as AI gets stronger: taste, original thinking, and emotional connection.

Not intelligence in the narrow, test-score sense. Not raw technical knowledge that can be retrieved from a database.

• Taste, the cultivated, hard-to-articulate judgment about what’s worth doing and how to do it well.
• Original thinking, the capacity to formulate questions that nobody else has thought to ask, rather than answering questions that have already been posed.
• Emotional connection, the specifically human quality of understanding and relating to other people in ways that can’t be reduced to pattern matching.

He made a related point earlier this year at the India AI Impact Summit, when he said that the difference between good scientists and great scientists is taste, the intuition about which problems matter and which hypotheses are worth pursuing. “I think that’s what separates the great scientists from the good scientists,” he said. “And I think that’s good. I think that’s what separates the great scientists from the good scientists.”

He added: “That’s probably the hardest thing in science, and will probably be the hardest thing for machines to be able to mimic.”

He was also direct about what he’d do personally after AGI, assuming he’s still around and involved. He said he’d want to turn to understanding the nature of reality using AI to help answer the deepest questions in physics and consciousness. These were the questions that drove him to AI research in the first place as a kid, and they haven’t changed.

The idea, as he put it in an earlier interview, is that AI could be “the ultimate tool for science”, not replacing scientists, but enabling them to tackle problems that are currently too complex for unaided human cognition.

On safety: “this is a good moment to strike while the iron is hot”

Hassabis was notably direct in the Axios interview about the gap between his private sense of urgency and the political reality.

He said that his economist friends are “I” That the conversation about AI’s society-reshaping potential remains largely confined to tech circles. And that it needs to change pointedly, soon, while a potential AI executive order that would mandate testing before new models are released is still under active discussion.

“I think [safety] needs to be accelerated,” he said. “This is a good moment to kind of strike while the iron is hot.”

He also said he’s currently in discussions with leaders at other top AI labs about possible safety measures, the specific content of which he declined to share publicly.

The concern he articulated about Glasswing is worth restating here. What the Anthropic incident demonstrated, in his reading, is not just that AI can find cybersecurity vulnerabilities. It’s that most governments, companies, and institutions discovered this capability cold. They weren’t prepared. That unpreparedness, multiplied across every domain where AI will develop similarly unexpected capabilities over the next four years, is what he’s trying to generate urgency about.

On Google and the competitive race

A few lines from his Fast Company interview round out the picture. On the narrative about Google falling behind and then ahead and then behind again: “These are ever-changing because the technology is fluid. We’re still very early in the AI race.”

On whether being at the frontier of both research and commercial deployment is internally contradictory: “It’s complicated, because you’ve also got the most voracious competition in tech history going on. I won’t pretend that it’s easy. But I think we get that balance right better than anyone else.”

On the deep roots of current agentic AI in DeepMind’s earliest work: “AlphaGo was an agent. Even our original Atari work… they were agents. Maybe we were a bit ahead of our time.”

And on the broader philosophy that drove the 2023 merger of Google Brain and DeepMind: “If we don’t disrupt ourselves, someone else will.”

The full interview is at the YouTube link: youtube.com/watch?v=4tVCHeAv0D4. Worth watching in full.