Demis Hassabis's AGI timeline is "like 2030 or something like this." His estimate for how long a true deep tech company takes to mature is about 10 years. Put those two numbers side by side and you get the most practical thing he said during a live conversation with Y Combinator's Garry Tan: anyone starting a deep tech venture today should assume AGI shows up in the middle of it.
"It's not bad necessarily," he said, "but you have to take that into account."
The warning carries weight because of who's giving it. Hassabis co-founded DeepMind in 2010, when investors and academics told him AI was a niche idea "we tried in the '90s and we know doesn't work." His teams went on to build AlphaGo and AlphaFold, and the second earned him the 2024 Nobel Prize in Chemistry. He now runs Google DeepMind, which ships Gemini into more than a dozen products with over a billion users each. Right now, founders are watching engineers claim 500x to 1,000x productivity gains and asking which businesses survive the next model release. Hassabis offered something close to a map: where the current paradigm is solid, where it still breaks, and which problems a foundation model upgrade won't flatten.
The current recipe stays, but one or two big ideas may be missing
Hassabis is confident that large-scale pre-training, reinforcement learning from human feedback and chain-of-thought reasoning will be part of the final AGI architecture. "I can't see a world in which we'll sort of realize in a couple of years this was a dead end," he said.
What's missing is a short list: continual learning, long-term reasoning, some aspects of memory, and consistency across tasks. The open question is whether scaling plus incremental innovation closes those gaps, or whether one or two genuine breakthroughs are still required. He puts the odds at roughly 50/50 and caps the number of missing ideas at two. Google DeepMind funds both bets.
That split is more candid than the usual lab messaging. It also frames the rest of the conversation, because nearly every limitation he described traces back to that short list.
Memory still runs on duct tape
Hassabis's PhD examined how the hippocampus folds new knowledge into what the brain already knows, largely during sleep, when REM cycles replay important episodes. DeepMind borrowed that idea early. Its 2013 Atari system, DQN, mastered games partly through experience replay, rerunning successful trajectories many times.
Today's approach is cruder. When Tan described the "dream cycles" and stitched-together memory systems developers build around stateless models, Hassabis agreed they amount to duct tape: "shove it all in the context window." A context window works like working memory. Humans hold about seven digits; models hold a million or 10 million tokens, but they store everything, including the unimportant and the wrong. Even with perfect storage, retrieving the one fact relevant to the current decision has a real cost.
The ceiling also arrives faster than it seems. Naively tokenize live video and a million tokens covers about 20 minutes. An assistant that understands a month or two of your life needs something smarter than a bigger buffer.
This is where memory connects to agents. Without continual learning, agents handle pieces of tasks well but can't adapt to the specific setting you drop them into. Hassabis called that "the missing piece" for agents that are truly fire-and-forget.
AlphaGo's ideas are returning inside today's models
Asked whether reinforcement learning is underrated, Hassabis said it "potentially" is. DeepMind described its work as agent research from day one: AlphaGo, and later AlphaStar for StarCraft, were systems that set plans and pursued goals, confined to games to keep the problems tractable. The past few years have been about generalizing that to world models and language. He sees today's thinking modes as "aspects of what was sort of pioneered with AlphaGo coming back now," and his team is revisiting old techniques such as Monte Carlo tree search at scale.
Reasoning is where the gaps show. Hassabis likes to play chess against Gemini because the thinking traces are easy to audit. He has watched the model consider a move, recognize it as a blunder, fail to find anything better, and play the blunder anyway. "You just shouldn't be seeing that happening in a very precise reasoning system."
He calls the result jagged intelligence: gold-medal performance at the International Mathematical Olympiad alongside elementary errors when a question is phrased a certain way. Current thinking methods are "fairly brute force," models often overthink into loops, and he suspects the missing ingredient is a form of introspection about their own reasoning. Monitoring a chain of thought and interjecting midway is one direction he flagged. He also noted the fix might be just one or two tweaks.
Today's frontier becomes tomorrow's edge model within a year
The more encouraging trend runs in the other direction. Tan cited Gemini's Flash models as roughly 95% as capable as the frontier at a tenth of the price, and Hassabis described distillation as one of DeepMind's core strengths, a technique Google researchers helped invent. The motive is practical: Search, Maps, YouTube and the Gemini app all need fast, cheap, low-latency serving.
He hasn't seen an information-density limit yet. DeepMind's working assumption is that a frontier model's capability reaches "really tiny, almost edge models" six months to a year after launch. Gemma 4, the open-weight family, hit about 40 million downloads in two and a half weeks.
Small models matter for more than cost. A model at 90% of the frontier can win back more than the lost 10% through faster iteration, especially when a human is collaborating with it. On-device models also keep personal data private: Hassabis sketched phones, glasses and home robots processing audio and video locally, delegating to cloud models only when needed.
That logic drives the open-weight strategy. Edge models are "vulnerable anyway" once deployed on devices, so Google decided they might as well be fully open. Hassabis added that he wants a strong Western open-source stack, since Chinese models currently lead in open weights.
Don't expect this to make inference free, though. Jevons' paradox means swarms of agents will absorb whatever compute exists, and even if fusion or better batteries drive energy costs toward zero, chip fabrication stays a bottleneck "for the next few decades." Efficiency keeps mattering.
Agents work, but nobody has shipped the vibe-coded hit
Hassabis agreed with Tan that agents are just getting started, and that an active system that solves problems on its own is the only path to AGI. But he was specific about what hasn't happened yet. He can now prototype Theme Park, the game that took him six months at 17, in half an hour. He sees people running dozens of agents for 40 hours without output that justifies the input. And no kid has used these tools to build a game that sells 10 million copies.
"That should be possible given the effort that's gone in. So something's still somehow missing."
He expects the full value to arrive in six to 12 months, and he expects it to appear first as individuals operating at 1,000x who ship best-selling products, with automation following later.
The same gap shows up in creativity. AlphaGo's famous move 37 was brilliant, but Hassabis wants a system that can invent Go itself from a brief: a game you can learn in five minutes, that takes lifetimes to master, that's aesthetically beautiful and finishes in an afternoon. Today's systems can't do that. He left open an uncomfortable possibility for the labs: perhaps the models already can, and what's missing is a creative person fluent enough with the tools to be "almost at one with" them.
The AlphaFold pattern has three conditions
The most reusable framework Hassabis offered is his lesson from the Alpha projects, which he says he should write up "when I have five minutes spare." A problem is ripe when it has:
- A massive combinatorial search space, large enough that brute force and special-case algorithms fail (Go moves and protein configurations both exceed the number of atoms in the universe).
- A clear objective function you can hill-climb, such as winning the game or minimizing a protein's free energy.
- Enough real data, or a simulator that generates in-distribution synthetic data.
Drug discovery fits. Somewhere, physics permits a compound that treats a disease without side effects; the only question is finding it efficiently. AlphaFold is now used by more than 3 million researchers, and Isomorphic Labs, DeepMind's spinout, is building out the adjacent chemistry.
The next target is a virtual cell, a simulation accurate enough that you can perturb it and trust the outputs. Hassabis estimates it's about 10 years away. DeepMind is starting with the nucleus because it's relatively self-contained, which illustrates the general trick: pick a slice of complexity whose inputs and outputs you can approximate. Data is the other constraint. If someone could image a living cell at nanometer resolution without killing it, the problem would become a vision task DeepMind knows how to solve. No such technique exists yet, so progress may come from new hardware or from better learned simulators.
Real discovery means posing the question, not just answering it
Hassabis says we're "close" to AI that does scientific reasoning, citing DeepMind's co-scientist system, but he hasn't yet seen a genuine major discovery from any lab. Going beyond known bounds requires something past pattern matching and extrapolation, closer to analogical reasoning.
His bar is high. Solving a Millennium Prize problem might be a couple of years out. Inventing a new set of problems that top mathematicians consider worth a lifetime of work is "another level harder." To measure progress he proposes an Einstein test: train a system on knowledge up to 1901 and see whether it produces what Einstein did in 1905, special relativity included. He suggested the labs should simply run that test repeatedly.
Build the tool an AGI would want to call
The founder advice follows from everything above. Anything built purely on model capability gets swamped by the next update. Hassabis points instead to teams that combine machine learning with another deep technical field, especially one involving physical atoms: materials, medicine, hard science. Ideally the founders are expert in both, or the team covers both.
His architecture view explains why that holds even after AGI. He expects general systems such as Gemini or Claude to call specialized systems like AlphaFold as tools. Training protein folding into Gemini would degrade its language abilities, so the specialists stay separate. A company that owns one of those specialist systems, or the factory or financial infrastructure around it, becomes something an AGI uses.
He also rejected the idea that ambitious work costs more effort. Deep problems and shallow ones are "just differently difficult," and life is short enough that you might as well spend it on something that wouldn't happen without you.
The test for any plan, then, is concrete: imagine the world where a general system arrives five years in, and ask whether what you've built becomes more valuable or redundant.
