Most predictions about AGI come from people who want it to arrive on a particular schedule. Demis Hassabis has been making the same prediction since 2010, and the prediction is just about now.
That's the part I keep coming back to.
When he and Shane Legg started DeepMind, they wrote blog posts extrapolating compute and algorithmic progress. Their estimate was twenty years. Fifteen years in, they are roughly on track. The interesting question is not whether the prediction was right — it's how much of the next five years is just compute, and how much is the stuff that compute alone will not buy you.
I watched Demis on a recent interview and what struck me was not his timeline. It was his honesty about the holes.
The Word That Stuck
Here's the phrase he used that I haven't been able to put down: jagged intelligence.
The systems we have today are amazing at certain things and then catastrophically bad at very basic things, depending on how you ask. Reposition a file in your repo and the agent falls over. Phrase a math problem slightly differently and the model that just solved an Olympiad question fails arithmetic. The intelligence is real, but it has holes in it.
What it actually means
A system that exhibits superhuman capability in narrow strips of capability space, separated by gulfs where it fails at things a child would not. Generality is not just average competence — it is the absence of cliffs. A general intelligence, by Demis's definition, should not have those holes.
Capability is real. The cliffs between the peaks are where every production agent dies.
This is the most useful frame I've come across for what's actually wrong with current models. Benchmarks measure peaks. They do not measure the cliffs between them. And the cliffs are where every production agent dies.
I've felt this acutely at NeoSapien. You build a workflow that looks magical in the demo and then a user does something the model wasn't expecting — moves a directory, renames a variable, asks the same question with one different word — and the whole thing collapses. The capability is there. The robustness is not. That's jaggedness.
Four Holes Demis Is Watching
When pushed on what's actually missing on the path to AGI, he doesn't say "more compute" or "bigger context." He lists four specific things, and they are worth taking seriously because the man has been right about the trajectory for fifteen years.
Continual learning. The models we ship today freeze the moment training ends. They do not integrate new information into their weights the way a brain does. The brain solves this elegantly through something like sleep — replay and consolidation, where the day's experiences get folded into the existing knowledge base. Nobody has cracked the equivalent for a trained foundation model. Every leading lab is working on it. None has shipped it.
Memory. Long context windows are, in his words, "a bit brute force." You just shove everything in. There is almost certainly a more elegant architecture for memory waiting to be invented — something hierarchical, something that knows what to keep and what to compress. We're using vector databases and bigger context as duct tape.
Hierarchical, long-horizon planning. Today's systems can plan a few steps ahead. They cannot plan a year ahead. A human mind can hold a multi-decade goal and decompose it down to today's actions. That decomposition — abstracting at multiple time horizons simultaneously — is something current architectures do not do well.
Consistency. Which brings us back to jagged intelligence. The absence of cliffs. The thing that turns a clever model into something you can trust to operate autonomously in the world.
Demis's four holes — algorithmic and architectural problems, not scaling problems.
Notice that none of these are "make the model bigger." They are all architectural and algorithmic. They are problems that get solved by people inventing new things, not by people writing bigger checks for GPUs. That distinction is going to matter a lot over the next two to three years.
The Last Set of Ideas
His most quietly important claim in the whole interview was this: the labs that have the capability to invent new algorithmic ideas are going to start having a bigger advantage as the last set of ideas have all the juice rung out of them.
Read that twice.
The current generation of frontier models is, broadly, the same idea — transformers, pre-training, RLHF, longer context, more chain-of-thought — applied at increasing scale with diminishing but still real returns. Those returns are slowing. Not stopping. Slowing. And once the marginal gain from another order of magnitude of compute starts to compress, the labs that can invent the next thing pull away from the labs that can only scale the current thing.
This is why the picture in the AI industry is shifting. Twelve months ago the meta-conversation was about who had the most GPUs. Today it's quietly about who has the deepest research bench. Demis would obviously say DeepMind. He would also note, accurately, that something like 90% of the breakthroughs underpinning the modern AI industry came out of Google Brain, Google Research, or DeepMind. Transformers. AlphaGo. Reinforcement learning at scale. AlphaFold. The history is hard to argue with.
If new ideas matter more than new chips for the next phase, this is a re-ranking event. And it's already happening. I switched my own primary research model to Gemini Deep Research about six months ago. It wasn't ideology. It was just better.
The juice is being rung out of the current paradigm. The next phase belongs to labs that can invent.
Ten Times the Industrial Revolution at Ten Times the Speed
This is the line of his that should be tattooed on every economist and policymaker working on this transition.
I sometimes quantify the coming of AGI as ten times the industrial revolution at ten times the speed. — Demis Hassabis
The industrial revolution unfolded over roughly a century. Child mortality fell from 40% to single digits. Modern medicine emerged. Cities reorganized. Whole categories of labor disappeared and were replaced by ones nobody could have anticipated. There was massive disruption and there was massive flourishing, often in the same decade.
Now compress that into a decade. Multiply the magnitude by ten.
The honest read is that "this time is different" is the most dangerous sentence in technology — and also, occasionally, true. Every previous wave was met by labor critics saying it would destroy work, and they were wrong, because the new technology unlocked categories of jobs that did not previously exist. Demis's caveat is the right one: the historical pattern is clear, but the speed and magnitude of this transition do not have a precedent. Pretending they do is a comfortable lie.
The thing that worries me is not whether new jobs emerge. They will. The thing that worries me is the velocity gradient — the gap between how fast capability moves and how fast institutions, retraining systems, social safety nets, and human meaning-making can move. The industrial revolution had a hundred years for societies to adapt. We have ten.
The velocity gradient: capability moves fast, institutions move slow, the gap is where the pain lives.
The Drug Discovery Two-Step
Demis gets emotional when he talks about science. His mother has multiple sclerosis. The thing he wants AGI to do most is shorten the painful gap between "we know how to fix this" and "the patient gets the drug."
His framing of how AI fixes drug discovery is unusually honest about the bottleneck:
Step one is the scientific design problem. Protein folding, compound design, toxicity prediction, binding affinity. AlphaFold was the opening move. Isomorphic Labs is the follow-through — a full drug design engine, end to end, ready in the next five to ten years. This is the part AI clearly accelerates. He's confident on this.
Step two is the regulatory problem. Clinical trials still take a decade. AI helps at the margin — simulating metabolism, stratifying patients to genomic profiles — but the trial structure itself is the bottleneck. The honest path is for a dozen AI-designed drugs to make it through the existing system, generate the back-test data, and only then can regulators trust the predictions enough to skip steps. Maybe animal testing becomes optional. Maybe dose escalation accelerates.
This is a ten-plus year arc just to compress the next ten-plus year arc. It's the most realistic version of "AI will cure diseases" that I've heard a serious person give.
And it's a useful pattern for thinking about every other industry AI is about to touch: the technical problem usually solves first, the institutional problem solves second and slower, and the gap between them is where most of the disappointment lives.
Technical solves first. Institutional solves second and slower. This pattern repeats in every industry AI touches.
The Question Nobody Is Asking
The bit that quietly haunted me — and has been haunting me since the interview — was when he was asked what nobody is talking about.
His answer was not technical. It was not economic. It was philosophical.
Assume we get the technical problem right. Assume we get the economic problem right. Both are very hard. But then there is a third problem that neither alignment researchers nor AI economists are equipped to handle: what is meaning, what is purpose, what does it mean to be human in a world where you are no longer the smartest thing in the room?
This is the question that I think about more than anyone in tech wants to admit they think about. It's the reason I keep ending up reading Vedanta and Kashmir Shaivism instead of more capability papers. The technical and economic transitions will be brutal. The philosophical transition is the one that will quietly determine whether the brutality was worth it.
"We need some great new philosophers to help us navigate that." From the man running the lab most likely to actually deliver AGI. Not "we need better policy." Not "we need more compute." We need philosophers. He is exactly right, and almost nobody is staffing for this.
The reason this matters: the people building AGI are some of the most thoughtful technologists alive, and even they are admitting they cannot answer the questions their work is forcing on the species. If the builders are saying "we need philosophers," everyone else needs to be listening.
London, on Purpose
One small thing I loved. He was asked why he hasn't moved DeepMind to the Valley.
His answer: London produces some of the deepest research talent in the world, the competition for it is structurally less fierce than it is in the Bay, and being a few thousand miles away from the maelstrom is conducive to thinking about twenty-year missions instead of next-quarter narratives.
He acknowledges the cost. You miss the gossip, the network, the vibe. You're not plugged in. But for deep tech — for the kind of work that requires thinking originally rather than reacting to whatever is trending — being slightly outside the cyclone is a feature, not a bug.
I think about this a lot. The Valley optimizes for velocity within an existing paradigm. Originality often requires being far enough away that you stop hearing the consensus.
What I'm Taking From This
A few things, in no particular order.
First, the timeline is real. When the man who has been right about the trajectory for fifteen years says five years, you can argue with the calibration but you cannot dismiss the bet. Plan accordingly.
Second, the next phase is algorithmic, not just computational. The labs that can invent new ideas — continual learning, memory architectures, hierarchical planning, consistency — are about to pull away from the labs that can only scale the current ones. This re-orders the industry.
Third, jagged intelligence is a better diagnostic than benchmark performance. If you're building anything on top of these models, the question is not "how good is it on average" — it is "how cliffs-y is it." Where are the holes. What inputs collapse the system. That is the only question that matters in production.
Fourth, the philosophical work is the actual bottleneck. Technical alignment is hard. Economic redistribution is harder. Figuring out what humans are for in a world of synthetic minds is the hardest problem of all, and it's not on most roadmaps. It should be.
And fifth, the fact that one of the most powerful people in technology spends his time worrying about his mother's MS, the right kind of regulation, and what philosophers we need — that is the kind of person you want building this. The interview was a reminder that the AGI transition will be shaped by a very small number of human beings making a very large number of decisions under speed and uncertainty. Some of those people are unserious. Demis is serious.
Five years from now, we'll know if he was right.
The harder bet is whether we'll have figured out what to do with the answer.