Hassabis: AGI likely within 5 years, with new ideas increasingly decisive

1. Brain-based AGI bar — ▶ 1:34

   DeepMind has defined AGI consistently as a system exhibiting all the cognitive capabilities of the human mind, because the brain is the only existence proof that general intelligence is possible. Hassabis puts a "very good chance" on reaching it within 5 years. His evidence is that co-founder Shane Legg wrote blog posts around 2010 extrapolating compute and algorithmic progress and predicting roughly 20 years, and Hassabis says the field is pretty much on track. The support is a single group's extrapolation, not independent confirmation.

2. Compute as workbench — ▶ 3:06

   Compute matters twice over. It scales models, and it is how researchers test new ideas: "the cloud is our workbench." An algorithmic idea has to be tested at reasonable scale or it may not hold in the main system, so many researchers with many ideas need a lot of compute. On scaling laws, Hassabis rejects the plateau narrative. Early generations delivered gains that were almost like doubling, which could not stay exponential, but returns remain very substantial, if smaller than at the start.

3. Jagged, forgetful systems — ▶ 5:10

   Hassabis says most capabilities are ahead of his expectations, citing video models and the interactive world model Genie. Critical gaps remain. Models do not learn after training, and no lab has yet worked out how to integrate new learning into systems that took months to train. He suggests a brain-like consolidation process, replaying memories as in sleep. He also names long-context windows as brute force, weak long-horizon planning, and inconsistency: "jagged intelligences" that fail elementary tasks when a question is phrased differently.

4. Assembling existing ingredients — ▶ 6:40

   Hassabis claims about 90% of the breakthroughs underpinning the modern AI industry came from Google Brain, Google Research or DeepMind, citing AlphaGo, reinforcement learning and transformers. This is his own assessment, not a verified figure. He attributes the recent acceleration to organizational change rather than new talent: combining people and compute from across the company to build the biggest models rather than two or three parallel versions, and operating with startup-like focus and pace.

5. Frontier gap widens — ▶ 9:21

   Among the three or four leading labs, Hassabis thinks the gap is starting to pull away. Coding and math tools help build the next generation of models, and it is getting harder to eke out the same gains from the same ideas. He infers that labs able to invent new algorithmic ideas will gain a bigger advantage over the next few years. This cuts against the commoditization thesis at the top end, though he offers no quantitative evidence of the divergence.

6. Open models trail frontier — ▶ 10:17

   DeepMind remains committed to open science, especially in applied scientific domains, and cites transformers and AlphaFold as examples. But Hassabis expects open-source models to sit one step behind the frontier, since the community usually needs about 6 months to re-implement new ideas. Gemma, DeepMind's open model suite, targets best-in-class performance for size, aimed at small developers, academics, early startups and edge computing. The pattern resembles today's: frontier models set the benchmark, and open models follow.

7. Building on foundation models — ▶ 11:25

   Hassabis disagrees with Yann LeCun on a few things, but not wholesale. He gives a 50/50 chance that something is still missing, possibly world models. His strong bet is that foundation models will not be replaced, because they keep showing gains from scaling. In his framing the open question is whether an LLM foundation model is the only key component of a future AGI system or one part of a larger total system built on top of it.

8. Drug design before trials — ▶ 13:20

   Hassabis frames drug discovery in two steps. Isomorphic Labs, spun out after AlphaFold, targets the design side: chemistry, compound design and toxicity checks. He expects that engine within the next "5 plus 5 to 10 years" (the transcript wording is garbled). Clinical trials remain the slower constraint. AI might simulate parts of human metabolism and match patients to drugs by genomic makeup. The real shift, he says, comes when about a dozen AI-designed drugs clear the process, giving regulators data to back-test predictions, perhaps eventually skipping animal testing.

9. Standards, audits, international body — ▶ 15:18

   Hassabis worries about misuse by bad actors and about keeping more agentic, autonomous systems within guardrails in a year or two. His proposal is minimum international standards: benchmarks testing for undesirable properties such as deception, and a certification akin to a kite mark. He wants government-backed technical bodies, building on the UK and US AI safety institutes, to audit models, possibly feeding an international body like the atomic agency. He also suggests barring outputs in non-human-readable tokens. He concedes that global coordination is getting worse.

10. Faster, bigger industrial revolution — ▶ 20:31

    Hassabis quantifies AGI as 10 times the industrial revolution at 10 times the speed, unfolding over a decade rather than a century. He accepts that past technologies destroyed old jobs and created higher-paying new ones, but says AGI will be bigger than the internet or mobile, so the same pattern cannot be assumed. Near-term AI is overhyped, he argues, while its 10-year impact is underappreciated. For distribution, he floats pension funds or sovereign wealth funds holding stakes in AI companies, plus redistributing productivity gains.

11. AI paying for energy — ▶ 24:07

    Asked about AI's unprecedented energy demand, Hassabis argues AI will more than pay for itself in energy terms in the medium to long run. His examples: optimizing grids, where he thinks 30-40% more efficiency is probably achievable; better weather and climate modeling; and enabling fusion, new batteries and superconductors, including work with Commonwealth Fusion. These are estimates and aspirations rather than demonstrated results, and his answer addresses the medium to long term rather than the near-term surge in demand.

Demis Hassabis's central claim is that the next phase of AI progress will be decided less by who has the most compute and more by who can invent new ideas. He is not saying scaling has stopped. Early model generations produced gains that were almost like doubling in performance. That could not stay exponential, but he says returns on scaling existing systems remain "very substantial," if smaller than at the start. The consequence, in his telling, is that leading labs can no longer extract the same gains from the same recipe. Those able to create new algorithmic ideas, he says, will have a bigger advantage over the next few years, because "all the juice has been rung out of them" for the previous set.

His explanation for why the gap should widen is mechanical. Coding and math tools help build the next generation of models, so a lab that is ahead can accelerate. At the same time, ideas and compute are bound together: Hassabis calls the cloud "our workbench," because a new algorithm must be tested at reasonable scale or it may not hold up in the main system. A reasonable inference is that the two advantages compound. A lab with many researchers needs a lot of compute to test their ideas, and without ideas worth testing, compute buys only the shrinking returns of the old recipe.

Which ideas are missing is where Hassabis is most specific. Current systems do not learn after training, and no lab has yet figured out how to integrate new learning into a model that took months to build. He points to the brain, where memories are replayed and consolidated into existing knowledge, and suggests something similar may be needed. He also criticizes long context windows as brute force, notes weak planning over long horizons, and describes "jagged intelligences" that excel in one phrasing of a question and fail elementary tasks in another. He nonetheless says that in most domains, from video models to the interactive world model Genie, the field is ahead of where he expected.

This creates a tension in his forecast. Hassabis gives a "very good chance" of AGI within 5 years, and says the timeline has held since co-founder Shane Legg extrapolated compute and algorithmic progress around 2010 and predicted roughly 20 years. Yet he also says there is a 50/50 chance something beyond today's foundation models is missing, and that gains from existing ideas are diminishing. His strong bet is that foundation models will be built upon, not replaced. He disagrees with Yann LeCun only on degree: the question, as he frames it, is whether an LLM is the sole key component or one part of a larger system. An extrapolation of algorithmic progress is weakest exactly where progress depends on inventions that have not yet occurred.

The competitive picture he paints has a similar tension. Hassabis claims about 90% of the breakthroughs underpinning the modern AI industry came from Google Brain, Google Research or DeepMind. That is his own assessment, and it serves his argument that DeepMind should be backed to make the next ones. He attributes the recent acceleration to organization rather than new talent: pooling people and compute to build the biggest models rather than two or three versions. Meanwhile he expects open-source models to trail the frontier by about 6 months, the time the community needs to re-implement new ideas. If ideas are the source of advantage and open models close the gap in 6 months, a lead is perishable. He does not reconcile that with his claim that the leaders are pulling away.

The discussion also shows why capability alone does not determine outcomes. In drug discovery, Hassabis expects Isomorphic Labs to build a drug design engine, covering chemistry, compound design and toxicity checks, but clinical trials still take many years. He sees AI helping by simulating parts of human metabolism and matching patients to drugs by genomic makeup. The bigger change, he suggests, would come once about a dozen AI-designed drugs pass the whole process. Regulators would then have data to back-test the models, and a further 10 years might eventually allow skipping steps such as animal testing. The sequencing matters: the technical problem is solved first, and the institutional one follows only after evidence accumulates.

Safety follows the same logic, with a harder problem attached. Hassabis names two worries: misuse of dual-purpose systems by bad actors, and keeping more agentic, autonomous systems within guardrails in a year or two. His prescription is minimum international standards, with benchmarks testing for properties such as deception, a certification akin to a kite mark, and independent audits by technical bodies like the UK and US AI safety institutes. He would go further with an international body modeled on the atomic agency, and he would bar systems from outputting tokens humans cannot read. He also concedes that the world is getting worse at the global coordination this requires. His proposals depend on the thing he says is scarce.

His economic argument rests on the pace of change. He frames AGI as 10 times the industrial revolution at 10 times the speed, over a decade rather than a century. He accepts the historical pattern that old jobs disappear and higher-paying new ones arrive, but says this technology is bigger than the internet or mobile, so the pattern cannot be assumed. He offers no labor-market evidence, and the interviewer's question about whether the same dynamics will hold gets a hedged answer. The compressed timescale is the substantive point: if disruption arrives in a decade, there is less time for new roles and institutions to emerge. His distribution ideas, sovereign wealth funds and pension funds holding stakes in AI companies, are responses to that compression.

His answer on energy shows the same optimistic structure. AI, he argues, will more than pay for itself in the medium to long run. He estimates that national grids could become 30-40% more efficient, points to weather and climate modeling, and expects AI to help reach fusion, batteries and superconductors. These are estimates and aspirations, and the answer is framed over the medium to long term. His view that the next year is overhyped while the 10-year horizon is underappreciated cuts both ways: it tells investors to discount near-term claims, and it tells everyone else that the largest consequences are not yet priced in. What he leaves unanswered is how capability, regulation and governance are meant to stay in step when the first is moving on a decade's timescale and the others are not.