Demis Hassabis on AGI Timelines and AI Safety Risks

TL;DR

Google DeepMind CEO Demis Hassabis puts roughly a 50% chance on reaching AGI within the next 5 to 10 years, calling DeepMind's original 20-year mission dead on track at year 15. He argues today's systems are not AGI because their performance is inconsistent, solving gold-medal olympiad math while tripping on high school problems or counting letters in a word.

Transcript

It's a very intense time [music] in the field. We obviously want all of the brilliant things these AI systems can do. Come up with new cures for diseases, new energy sources, incredible things for humanity. [music] That's the promise of AI. But also, there are worries. If the first AI systems are built with the wrong value systems, or they're built... Read More

Key Insights

• AGI as DeepMind defines it is a system that exhibits all the cognitive capabilities humans have, a definition co-founder and chief scientist Shane Legg helped establish around the early 2000s. The human mind is the reference because it is the only existence proof that general intelligence is possible.
• Hassabis puts roughly a 50% chance on AGI arriving in the next 5 to 10 years, noting some peers predict 2 to 3 years and others somewhat longer. He describes DeepMind's original 20-year mission, now about 15 years in, as pretty much dead on track.
• The strongest evidence current systems are not AGI is inconsistency of responses. AlphaProof solves international math olympiad problems to gold medal standard, yet the same class of systems still trips up on high school maths or counting the number of letters in a word.
• Missing capabilities in today's LLMs and chat bots include reasoning, planning, and memory. Hassabis says they cannot do true invention or true creativity, and cannot hypothesize new scientific theories, which makes them extremely useful and impressive but full of holes.
• Hassabis expects AGI's arrival to look incremental rather than as a step function, because the physical world still operates under physical laws. Factories, robots, and other real-world systems mean the impact of digital intelligence may take another decade plus to materialize.
• The hard takeoff scenario holds that self-improving AGI systems could code future versions of themselves so fast that a few days' lead becomes a chasm. Hassabis treats this as an unknown rather than a given, since less accelerating, more incremental paths would make near-simultaneous arrival matter little.
• AI systems carry an imprint of the values, norms, and culture of the designers who built them, which is why Hassabis frames geopolitical and international questions as central rather than peripheral to how the first powerful systems get built.
• Hassabis names two risks he worries about most: bad actors, whether individuals or rogue nations, repurposing general purpose AI technology for harmful ends, and the technical risk that safety guard rails around increasingly powerful, agentic systems can be circumvented.
• Regulation needs to be international to work, in Hassabis's view, because these are digital systems and restricting them in one area does not improve overall safety. He also calls for nimble regulation that adapts as research knowledge improves, alongside societal debate on governance and institution building.

Questions & Answers

Q: When does Demis Hassabis think AGI will arrive?

Hassabis puts roughly a 50% chance on having what DeepMind defines as AGI within the next 5 to 10 years. He frames this against DeepMind's founding mission, which was set up as a 20-year plan to solve intelligence and then use intelligence to solve everything else. About 15 years in, he says the company is pretty much dead on track by its own estimate. He notes that some of his peers predict 2 or 3 years and others say a little more, and that the spread in predictions largely follows from disagreement over how AGI is defined in the first place.

Q: What is DeepMind's definition of AGI?

DeepMind defines AGI as a system that has the ability to exhibit all the cognitive capabilities humans have. Hassabis credits co-founder and chief scientist Shane Legg with helping define the term around the early 2000s, and says the company has been consistent about this definition from the very beginning. The reference to the human mind matters because the human mind is the only existence proof we have, perhaps in the entire universe, that general intelligence is possible. Under this definition, claiming general intelligence requires showing that a system generalizes across all these domains, not just excelling in some of them.

Q: Why does Hassabis say current AI systems are not yet AGI?

His main reason is the consistency of responses. Systems can solve international math olympiad problems to gold medal standard, as DeepMind's AlphaProof system does, yet still trip up on high school maths or even counting the number of letters in a word. That level of difference in performance across the board is not what you would expect from a genuinely general system, so it shows the systems are not fully generalizing yet. He also points to missing capabilities in reasoning, planning, and memory, and says today's models cannot do true invention, true creativity, or hypothesize new scientific theories.

Q: Will AGI arrive as a sudden phase shift or gradually?

Hassabis says this is debated and not clear to him, but his guess is that it looks more like an incremental shift than a step function. His reasoning is that even with such a system in hand, the physical world still operates under physical laws. Factories, robots, and other real-world infrastructure set their own pace, so it will take a while for the effects of this digital intelligence to really impact many real-world things, possibly another decade plus. He acknowledges there are other theories where the impact could come faster.

Q: What is the hard takeoff scenario in AI?

The hard takeoff scenario is the idea that AGI systems become able to self-improve, coding future versions of themselves, and do so extremely fast. Under that assumption, what starts as a slight lead of a few days could suddenly become a chasm, which is the reasoning behind Eric Schmidt's claim that whoever gets AGI even 10 minutes before the next party could never be caught. Hassabis treats this as one of many unknowns rather than settled fact. He points out there are many other ways it could go where progress is more incremental and self-improvement does not accelerate that way, in which case arriving around the same time would not make much difference.

Q: What AI risks does Demis Hassabis worry about most?

He names two. The first is bad actors, whether individuals or rogue nations, repurposing general purpose AI technology for harmful ends. The second is the technical risk of AI itself: as systems get more powerful and more agentic, making sure the guard rails around them are safe and cannot be circumvented. Both interact with the question of what the first systems built by humanity will be like, since commercial imperatives, national imperatives, and safety concerns all bear on who is in the lead and where those projects sit. He says today's systems are impressive but flawed, and do not pose any sort of existential risk yet.

Q: Does Hassabis still support AI regulation?

Yes. He says he has been consistent on this and continues to believe that smart regulation around these increasingly powerful systems will be important, even as geopolitical overlays have to be taken into account and the world looks very different from five years ago. His key qualification is that regulation needs to be international, which looks hard given how the world is currently working. Because these are digital systems that will affect everyone, restricting them in one area does not really help the overall safety of the systems being built. He also calls for nimble regulation that moves as research knowledge improves.

Q: What is DeepMind doing about AI safety and controllability?

Hassabis says DeepMind is increasingly putting resources into security and areas like cyber, as well as research into the controllability and understanding of these systems, a field sometimes called mechanistic interpretability. He describes these as subbranches of AI receiving investment, and says even more needs to happen. Beyond the technical work, he argues for broader societal debate about institution building: how governance should work, and how to reach international agreement on at least some basic principles covering how these systems are used, deployed, and built. He is optimistic the technical unknowns can be overcome, but thinks the geopolitical questions may prove trickier.

Summary & Key Takeaways

• Demis Hassabis, CEO of Google DeepMind, Nobel Prize winner and knight, sat down with Wired to discuss AGI, the future of work, and how Google plans to compete in the age of AI. He founded DeepMind on the idea of solving intelligence and then using intelligence to solve everything else, a mission he estimates is about 15 years into a 20-year timeline.

• Hassabis defines AGI as a system able to exhibit all the cognitive capabilities humans have, a definition his co-founder and chief scientist Shane Legg helped establish around the early 2000s. The human mind matters as a reference because it is the only existence proof we have, perhaps in the universe, that general intelligence is possible at all.

• He identifies missing capabilities in current systems: reasoning, planning, and memory. Today's models cannot do true invention, true creativity, or hypothesize new scientific theories. The clearest evidence they are not yet generalizing is inconsistency, with gold-medal-standard performance on international math olympiad problems from AlphaProof alongside failures on high school maths.

• On arrival, Hassabis expects an incremental transition rather than a step function, partly because the physical world still runs on physical laws. Factories, robots, and other real-world systems mean the effects of digital intelligence may take another decade plus to land, though he acknowledges competing theories where it could come faster.

• He treats the hard takeoff scenario, where a few days' lead becomes a chasm through self-improvement, as one of many unknowns rather than a certainty. Because systems carry an imprint of their designers' values, norms, and culture, he considers international questions important and continues to argue for smart, nimble, internationally coordinated regulation.

• Two risks concern him most: bad actors, whether individuals or rogue nations, repurposing general purpose AI for harmful ends, and the technical risk that guard rails around increasingly powerful and agentic systems can be circumvented. He says today's systems pose no existential risk, but resources are going into security, cyber, controllability, and mechanistic interpretability.