Demis Hassabis: Why AGI is Bigger than the Industrial Revolution & Where Are The Bottlenecks in AI

Summary

• Hassabis puts “a very good chance” on AGI within the next 5 years — and insists this isn’t recency-driven: co-founder Shane Legg’s 2010 blog posts extrapolated compute and algorithmic progress to roughly 20 years, “and I think we’re pretty much on track.” His bar hasn’t moved either: a system with all the cognitive capabilities of the human mind, the brain being “the only existence proof we have” that general intelligence is possible.
• Scaling laws haven’t plateaued — returns are “still very substantial, although they’re a bit less than they were” — but he argues against commoditisation at the frontier: the gap among the three or four leading labs is “starting to pull away,” because coding and math tools compound into the next generation and, as “all the juice has been rung out” of existing ideas, advantage shifts to labs that can invent new algorithms.
• He backs DeepMind to be that lab: “about 90% of the breakthroughs that underpin the modern AI industry” came from Google Brain, Google Research, or DeepMind, and the recent surge came from consolidating talent and compute into the biggest models rather than “two or three versions around the company” while “acting almost like a startup.” Open source stays structurally “probably one step back from the absolute frontier” — roughly 6 months to re-implement — with Gemma aimed at small developers and edge.
• The missing pieces before AGI are named, not mysterious: continual learning among the missing capabilities (the brain does it “very elegantly,” probably through things like sleep and reinforcement learning, with consolidation and replay), memory beyond “brute force” long context, long-horizon planning, and consistency — today’s models are “jagged intelligences” that are amazing at certain things posed one way and fail elementary ones posed another.
• Hassabis thinks Isomorphic Labs could have a complete drug-design engine in 5 to 10 years; the real unlock comes when “a dozen or so AI drugs get through the whole process” and regulators can back-test the models — then trials might skip animal testing and climb dose ladders faster. Hassabis sees Isomorphic, headquartered in London, as having the potential to become a trillion-dollar company.
• On safety, he endorses Hawking’s “we must get it right because we might not get another chance,” worries about systems becoming “more agentic, more autonomous… maybe in a year or two’s time,” and wants an atomic-agency-style international body, benchmarks including deception, and a certification “kite mark” — while conceding that perhaps the world’s most consequential technology is arriving amid “a very fragmented international system.”
• The macro frame: AGI is “10 times the industrial revolution at 10 times the speed” — a decade, not a century. AI is “a bit overhyped” today yet “very underappreciated” on a 10-year view; maybe pension funds should buy the big AI companies so everyone owns a piece, alongside sovereign wealth funds; and on energy, he thinks AI will “more than pay for itself” — 30-40% more efficiency from national grids, with fusion among the possible breakthroughs.

Deep dive

1. AGI within five years — on the schedule set in 2010

• The definition hasn’t moved: AGI is a system exhibiting “all the cognitive capabilities the human mind has” — a deliberately high bar, because the brain is “the only existence proof we have… that general intelligence is possible.” On timing he holds a probability distribution, but “a very good chance of it being within the next 5 years.”
• The receipts: Shane Legg’s 2010 blog posts — written when “everyone thought AI basically didn’t work” — extrapolated compute and algorithmic progress to roughly 20 years to AGI. “They’re still there on the internet for people to check… I think we’re pretty much on track.”
• Most domains are ahead of his own expectations — video models, Genie’s interactive world models: “if you’d shown me that 5, 10 years ago, I would have been pretty amazed.” But critical capabilities are still missing: continual learning among them — models don’t learn after training, while the brain does it “very elegantly,” probably through things like sleep and reinforcement learning, with consolidation and memory replay, an analogue he’s “thought for a while” AI may need. Long context windows as memory are “a bit brute force,” and long-horizon hierarchical planning is weak.
• The biggest gap may be consistency: today’s systems are “jagged intelligences” — amazing at certain things when a question is posed one way, failing “quite elementary things” posed slightly differently. Demis agrees with Harry’s example: “Sure, 100%. That’s a disaster.” “A general intelligence shouldn’t be that sort of jagged.”

2. Scaling still pays — and the frontier gap is “starting to pull away”

• The main bottleneck is compute, and not only for scaling: “the cloud is our workbench” — every new algorithmic idea has to be tested at reasonable scale or it won’t hold in the main system, so many researchers with many ideas means a large experimental compute bill.
• On the plateau question, a nuanced no: early generations were “almost like doubling in performance” and that had to slow, but returns from further scaling are “still very substantial, although they’re a bit less than they were.”
• Against commoditization: the three or four leading labs’ gap is starting to pull away because today’s coding and math tools compound into the next generation, and “it’s getting harder and harder to eke out the same gains from just the same ideas” — so “labs that have capability to invent new algorithmic ideas are going to start having bigger advantage” as “all the juice has been rung out” of the current set.
• Versus Yann LeCun: he gives “a 50/50 chance there’s some things maybe missing” — perhaps world models — but bets strongly that foundation models don’t get replaced: “it’s going to get built on top of these foundation models.” The only real question is whether the LLM is the key component or the total system.

3. DeepMind’s comeback: 90% of the breakthroughs, run like a startup

• His claim to the invention moat: “about 90% of the breakthroughs that underpin the modern AI industry” — AlphaGo, reinforcement learning, transformers — came from Google Brain, Google Research, or DeepMind. The recent acceleration Harry noticed came from organizational consolidation: pooling talent and compute into the biggest models instead of “two or three versions around the company,” and “acting almost like a startup.” “I would back us to make those breakthroughs in the future if there are any missing ones.”
• Open models will sit “probably one step back from the absolute frontier” — it takes the community about 6 months to re-implement frontier ideas — while Google keeps open-sourcing science (transformers, AlphaFold) and pushes Gemma to be “best in class for their sizes” for small developers, academics, and edge computing.

4. Isomorphic: solve drug design first, then convert the regulator

• The stakes are personal — his mother has multiple sclerosis — and Harry’s worry is the pipeline, not the science: even a discovered cure takes a decade of trials to reach her. Hassabis’s answer is sequenced: Isomorphic Labs (spun out after AlphaFold, “doing extremely well”) is solving the rest of drug discovery — chemistry, compound design, toxicity — and he thinks the whole drug-design engine will be ready in the next 5 to 10 years.
• Trials themselves could get AI help: simulating parts of human metabolism and stratifying patients so each gets the drug suited to their genomic makeup. But “the real revolution” comes when “a dozen or so AI drugs get through the whole process” and regulators can back-test the models’ predictions — then “maybe skip out some steps,” like animal testing, and climb dose ladders faster, perhaps ten further years out.
• The ambition is stated without hedging: “I want to literally cure cancer. I know people say that’s the cliche” — but Isomorphic is a general-purpose platform, targeting neurodegeneration, cardiovascular, immunology, and cancer first, “eventually applicable to every disease area.”

5. Safety needs an atomic agency — with deception among the benchmarks

• He endorses Hawking’s line — “we must get it right because we might not get another chance” — and names two distinct worries: misuse by bad actors repurposing dual-use systems, and the technical problem of keeping models “on the guardrails” as they become “more agentic, more autonomous… maybe in a year or two’s time.”
• His magic-wand design: an international body “similar to the atomic agency,” fed by AI safety institutes (the UK’s, set up under Sunak, is “doing great work”), running benchmarks for undesirable traits — including deception, since a deceptive system “could be getting around other safeguards” — plus a certification “kite mark” companies can safely build on. One proposed safeguard is avoiding systems outputting tokens “that are not human readable.”
• The uncomfortable timing, in his own words: “the most consequential maybe technology the world’s ever seen at the same time as a very fragmented sort of international system.”

6. “10 times the industrial revolution at 10 times the speed”

• Harry relays that Mark Andreessen called him “a Marxist” for raising labor displacement; Hassabis splits the difference — history does replace old jobs with new, higher-paying ones, and “you have to be very careful to say this time is different” — but he does think the scale is different: “10 times the industrial revolution at 10 times the speed,” unfolding over a decade instead of a century. Pre-industrial child mortality was 40%; you’d never wish the revolution away, “but ideally this time around we mitigate some of the downsides a bit better.”
• The hype dichotomy, held simultaneously: “literally today and in the next year things are a bit overhyped in AI… But it’s still very underappreciated how revolutionary this is going to be in the time scale of about 10 years.” Beyond the economic questions, he worries about the philosophical questions: assuming the technical and economic parts get solved, “what is meaning, what is purpose… what does it mean to be human? I think we need some great new philosophers.”
• On concentration of wealth: maybe pension funds should buy into the big AI companies, and maybe every country should have a sovereign wealth fund “so that everyone has a piece,” alongside deliberate redistribution of narrow productivity gains — and possibly breakthroughs in fusion, superconductors, and batteries.
• On the energy crisis AI creates: he thinks AI “will in the medium to long run more than pay for itself” — “30-40% more efficiency out of our national grids,” world-best weather modeling, and breakthrough technologies including fusion (DeepMind works with Commonwealth Fusion), which would mean “effectively unlimited rocket fuel” from seawater.

7. London as structural edge — and Europe’s trillion-dollar gap

• Why he never moved: the UK has “three or four of the top 10 universities in the world,” less competition for that talent, and distance from the Valley — less plugged into “the gossip and the latest trends and vibes” but “very conducive to thinking deeply,” which suits a 20-year deep-tech mission.
• Europe’s trillion-dollar company: “Not yet.” Daniel (likely Daniel Ek) might get there with Spotify or Helsing — “I’m going to try and do that with Isomorphic, which is headquartered here.” The fix: unlock pension funds for growth-stage investing — the billion-dollar rounds were missing when he raised sub-£1m rounds for DeepMind and are “still kind of missing today” — plus maybe the EU Inc innovation.