Demis Hassabis: We're Three Quarters of the Way to AGI

ELI5/TLDR

Demis Hassabis sits down with Sequoia and says, calmly, that AGI arrives around 2030 — exactly the 20-year arc he predicted in 2010. The interesting parts aren’t the headline. They’re the side claims: drug discovery collapsing from 10 years to weeks; a “virtual cell” on the way; simulation becoming a new branch of science for things economics can’t run controlled experiments on; and a quiet philosophical bet that information, not energy or matter, is the universe’s base layer.

The Full Story

The on-track claim

Hassabis has been saying “AGI by 2030” for years, and he says it again here, in the rapid-fire round, without flinching. The framing is what’s worth noticing — he treats the prediction as a project plan, not a prophecy.

“We thought it would be a 20 year mission. And I think we’re basically exactly on track as a field for that.”

The headline number is “three quarters of the way.” Read against a 2010 start and a 2030 finish, that’s just arithmetic. Hassabis isn’t claiming acceleration. He’s claiming the original schedule held.

Five years ahead, not fifty

The most useful founder lesson in the conversation comes from Elixir Studios, his pre-DeepMind games company. They tried to simulate a million-person country on a late-90s Pentium. It broke them.

“You want to be 5 years ahead of your time, not 50 years ahead.”

He carries the same heuristic into DeepMind’s 2010 founding. Deep learning had just been published. Reinforcement learning was siloed off in another corner. GPUs were getting good. The bet wasn’t that any one ingredient would work — it was that the timing of all three was finally right. He calls his early team “keepers of a secret,” because academia at the time literally rolled their eyes at the phrase “AGI.”

AlphaFold was the real bell

The interviewer asks when biology gets its ChatGPT moment. Hassabis pushes back: it already happened. AlphaFold cracked a 50-year grand challenge — predicting the 3D shape of a protein from its amino-acid sequence. That’s the part you’ve heard.

The part that’s newer: Isomorphic Labs, the DeepMind spinout, is now doing the next leg — designing the molecule that binds to the protein. The pitch is straightforward. Right now, drug discovery averages ten years and costs billions, with most of the time spent in wet labs ruling things out. Move 99% of that exploration into silicon, and the wet lab becomes a validation step rather than a search.

“I think we could reduce drug discovery times instead of taking like an average of 10 years down to months maybe even weeks, and perhaps even days one day.”

If this lands within “next few years” as he claims, personalized medicine — slight variations on a base drug, tuned to one person — becomes routine instead of exotic.

Machine learning as biology’s missing language

Mathematics works for physics because physics is clean. A few variables, hard equations, beautiful answers. Biology isn’t like that. It’s emergent, messy, full of weak signals across too many variables for any human to track.

Hassabis’s claim is that machine learning is to biology what calculus was to physics — the right description language for the problem.

“I think machine learning is the perfect tool to describe those kinds of systems where mathematics hasn’t been able to do that, either because we can’t manage it as top mathematicians because too complex, or the expressive power of maths is not enough.”

DeepMind is building what he calls a “virtual cell” — a learned simulator of cellular biology. Same approach as their weather model, WeatherNext, which is now more accurate and faster than what professional meteorologists use. (No, he’s not trying to control the weather. Just understand it.)

Simulation as a new branch of science

This was the most interesting riff in the conversation. Economics, sociology, the soft sciences — they don’t get to run controlled experiments. You can’t raise rates by half a percent in a thousand parallel universes and watch what happens. You get one shot, and you argue about it for a decade.

But if you can build an accurate enough learned simulator, you can. Sample it ten thousand times. Vary one input. Watch the distribution of outcomes.

“If you could simulate things really accurately, then maybe there’s sort of new sciences to be done where you can rigorously sample from a very accurate simulator.”

He goes one step further — once you have the implicit simulator, you might be able to extract explicit equations from it. A Maxwell’s equations for emergent systems. He hedges. He’s not sure such equations exist. But if they do, this is how you’d find them.

Information as the substrate

Here he goes properly philosophical. Classical 1920s physics treats matter and energy as the primary stuff, with information as something that happens on top. Hassabis flips it.

“I actually think it’s a better way to understand the universe is to think about it as information first.”

Biology, in this view, is matter resisting entropy by processing information. AI is the same trick, in silicon. If information is fundamental, AI isn’t just useful — it’s something closer to direct engagement with the universe’s underlying language.

He pairs this with a Turing point: AlphaFold modeled protein folding, which everyone assumed required quantum-scale computation, and got near-optimal answers on a classical neural network. Things we thought needed quantum computers might just need the right classical model.

Tool first, agency later

Asked when AI stops being a tool, Hassabis gives a careful answer. The agent era is here — systems are getting more autonomous. But agency, consciousness, selfhood — those are a separate frontier, and his preference is to build the tool first, use it to understand our own minds better, and then approach the consciousness question with sharper instruments.

On consciousness itself, he won’t pretend to a definition. Self-awareness, continuity over time, a model of self versus other — necessary but not sufficient. He notes a real asymmetry: we believe each other are conscious partly because we run on the same biological substrate. With silicon, that shortcut never applies.

Key Takeaways

AGI ETA: 2030. Hassabis treats this as on-schedule, not a moving target.
The DeepMind founding bet was timing, not invention — three known ingredients (deep learning, RL, GPUs) ready at once.
AlphaFold already was biology’s ChatGPT moment, just with worse PR.
Drug discovery: the goal is 99% in silico, wet lab as validation only. Targeted “next few years.”
Machine learning is the right description language for emergent systems (biology, climate, cells) the way maths is for physics.
Learned simulators may unlock controlled experiments in economics and other soft sciences.
Information may be more fundamental than matter or energy. If true, AI is closer to physics than to engineering.
Things assumed to need quantum computers may yield to classical neural nets, given the right framing.
His preferred sequence: build the tool, then use the tool to understand consciousness.
Founder rule: aim five years ahead, not fifty. Elixir Studios broke on the latter.

Claude’s Take

This is a polished, low-temperature conversation — Sequoia knows their audience, Hassabis knows his beats. There’s nothing here that will surprise anyone who’s followed him for a year. The 2030 number is unchanged. The AlphaFold-to-Isomorphic story has been told. The “AI for science” framing is core DeepMind doctrine.

What earns the score is the simulation-as-new-science thread and the information-first ontology. Both are real, defensible, non-obvious claims. The simulation one in particular is the kind of thing that, if it works, redraws the line between hard and soft sciences. Imagine economics with replication. That’s a big imagine.

The BS filter: “next few years” is doing a lot of work. He says it about Isomorphic compounds, about virtual cells, about the medical revolution. Hassabis is more careful than most AI executives, but he’s still an AI executive, and the timeline-flex is the standard tool. The AlphaFold track record buys him credibility most can’t afford.

Score: 8. High signal density, novel framings on simulation and information, and a rare moment where someone calmly defends a long-standing prediction without revising it upward to grab a headline.