Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
What this covers
Demis Hassabis discusses the architecture most likely to produce AGI and the safety measures required to govern it. The conversation centers on a specific technical hypothesis: that [large multimodal models](jargon:multimodal) serving as learned priors must be paired with [AlphaZero](jargon:AlphaZero)-style planning and search on top. Hassabis traces this argument through the history of DeepMind's bets on reinforcement learning, deep learning, and search as general, scalable algorithms, and connects it to observations from neuroscience—particularly how memory and imagination both reconstruct from semantic components, suggesting current AI systems lack an analogous mental simulation capacity. The interview also covers the model-search trade-off: richer world models dramatically reduce the search budget needed, as evidenced by AlphaZero versus brute-force approaches and human grandmasters using only hundreds of positions versus millions.
Beyond architecture, Hassabis outlines governance requirements for superhuman systems. These include grounding mechanisms (how language-trained models acquire real-world reference without multimodal input), deception evaluations and benchmarks, narrow specialized AIs to audit system reasoning, and hardened sandboxes both containing the system and securing weights against theft. He argues that scaling is not automatic—new capabilities require hyperparameter adjustment and innovation at each order of magnitude, with non-linear emergence in downstream capabilities despite predictable loss curves. The discussion also addresses where current systems fall short: they cannot yet formulate scientific hypotheses or ask novel questions, lack reliable episodic memory and personalization, and remain insufficiently dependable for real-world deployment, though they are beginning to assist expert search in constrained domains like protein structure.
Hassabis argues that AGI will most plausibly emerge by combining large multimodal models (as learned priors/world models) with AlphaZero-style planning and search on top, while better world models reduce the search needed, and that this must be pursued with 'bold but responsible' safety practices including grounding, evals, and secured weights.
- Large models are necessary but insufficient for AGI; planning/search on top is the missing component
- Better, richer world models allow far less search (AlphaZero vs Deep Blue; human grandmasters)
- Grounding, deception evals, sandboxes, and weight security are required to align superhuman systems
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Human memory is a reconstructive process, not a videotape—it is reassembled from familiar components—and imagination is the same process using the same semantic components recombined in novel ways for purposes like planning; this idea of pulling together world-model parts to simulate something new for planning is probably still missing from current AI systems.
“memory, at least human memory, is a reconstructive process. It’s not a videotape. We sort of put it together back from components that seem familiar to us... That’s what made me think that imagination might be the same thing.”
There is a trade-off between model quality and search: AlphaZero plays beyond world-champion level using only tens of thousands of evaluated positions versus millions for brute-force systems like Deep Blue, while human grandmasters use only a few hundred—because richer, more accurate world models let you pick which nodes to expand, so improving models makes search more efficient and lets you go further.
“AlphaZero and AlphaGo may look at around tens of thousands of possible positions in order to make a decision about what to move next. A human grandmaster or world champion probably only looks at a few hundred moves”
Shane Legg's 2009 thesis proposed that a test for AI is whether you can compress Wikipedia, which is essentially the loss function used to train modern LLMs, and Hassabis's own 2016 paper comparing neuroscience and AI identified attention as a needed capability before transformers existed.
“we had these things called out and we had some early attention papers, but they weren’t as elegant as transformers in the end, neural Turing machines and things like this. Transformers were the nicer and more general architecture of that.”
Games are an efficient AI proving ground partly because reward functions are easy to specify, whereas real-world problems make defining the right objective and reward function in a general yet sufficiently specific way much harder, though even scientific problems usually admit ways to specify the goal.
“How does one define the right objective function, the right reward function, and the right goals? How does one specify them in a general way, but specific enough that one actually points the system in the right direction? For real-world problems, that can be a lot harder.”
Neuroscience has provided directional inspiration for major AI advances such as reinforcement learning combined with deep learning, experience replay, and attention—not as one-to-one algorithmic mappings but as architectural and representational inspiration; the brain is an existence proof that general intelligence is possible.
“Some of our pioneering work we did there were things like experience replay and even the notion of attention, which has become super important. A lot of those original inspirations came from some understanding about how the brain works”
Large models exhibit surprising grounding despite not experiencing the world multimodally, plausibly because RLHF feedback comes from grounded human raters who transmit grounding, and because language may contain more grounding than linguists previously thought—raising philosophical questions not yet scratched.
“I think we get some grounding through the RLHF feedback systems because obviously the human raters are by definition, grounded people. We’re grounded in reality, so our feedback is also grounded.”
Aligning superhuman AIs requires a toolkit including more stringent evaluations and benchmarks for whether a system can deceive or exfiltrate its own code, using narrow specialized AIs to help human scientists analyze what the general system is doing, and hardened cybersecurity sandboxes that both keep the AI in and hackers out so experiments can run more freely.
“I think we don’t have good enough evaluations and benchmarks for things like if the system can deceive you. Can it exfiltrate its own code or do other undesirable behaviors?”
While a pure-RL, no-priors AlphaZero-style path to AGI is theoretically possible, the quickest and most plausible path is to use all existing world knowledge from the Web via scalable algorithms like transformers as a prior, then add planning and search on top; the final AGI will likely include large multimodal models but not rely on them alone.
“by far the quickest way to get to AGI, and the most plausible way, is to use all the knowledge that’s existing in the world right now that we’ve collected from things like the Web.”
Deception is a root-node trait you don't want in AI systems, because if you are confident the system exposes what it actually thinks, you can then use the system itself to explain its own reasoning to you—analogous to how a chess grandmaster can explain post hoc a move you couldn't have found, and possibly provide proofs in mathematical problems.
“I think about things like deception as being quite root node traits that you don’t want. If you’re confident that your system is exposing what it actually thinks, then that opens up possibilities of using the system itself to explain aspects of itself to you.”
New inventions require substantial compute because a new idea is only useful if it can scale, so you must test many things at reasonable scale; some ideas fail at toy scale but work at larger scale—and these are the more valuable ones—so the exploration process itself demands significant compute.
“A new innovation, a new invention, is only useful if it can also scale. So you need quite a lot of compute to do new invention because you’ve got to test many things, at least some reasonable scale... some new ideas may not work at a toy scale but do work at a larger scale. In fact, those are the more valuable ones.”
AI systems are reaching the stage where they can help the best scientists make breakthroughs quicker by triaging the search space and finding solutions like AlphaFold does for protein structure, but they cannot yet create hypotheses or ask the right question—the hardest part of science—because that requires boiling a space down to the critical question and formulating the problem, which systems have no idea how to do; they excel only at searching large combinatorial spaces with a clear objective function.
“They’re not at the level where they can create the hypothesis themselves or ask the right question. As any top scientist will tell you, the hardest part of science is actually asking the right question.”
From the start DeepMind bet on generality and learning, triangulating on reinforcement learning, search, and deep learning as algorithms that scale, are general, and don't require handcrafted human priors—deliberately moving away from the 1990s logic-based expert systems whose hand-coded information proved too rigid; AlphaGo demonstrated these systems were ready to scale and transformers later turbocharged progress.
“we always thought that we bet on generality and learning... That’s why we triangulated on reinforcement learning, and search, and deep learning as three types of algorithms that would scale, be very general, and not require a lot of handcrafted human priors.”
RL agents like AlphaZero and AlphaGo are active learners whose next actions determine which next data or experience they receive, creating a feedback loop; this active learning, including in realistic simulation environments, is essential for understanding how actions affect the world and for robotics.
“these RL agents we’ve always been working on and pioneered, like AlphaZero and AlphaGo, actually are active learners. What they decide to do next affects what next learning piece of data or experience they’re going to get.”
Because the technology is so consequential—bigger than any one company or industry—the governance of superhuman intelligence must be a broad collaboration among civil society, academia, and government, ideally reaching international consensus at a UN level on what to deploy these models for, ensuring they benefit everyone.
“I think it’s much bigger than any one company or even industry in general. I think it has to be a big collaboration with many stakeholders from civil society, academia, government, etc.”
Beyond the risk of deployed models being misused, rogue actors and nation-states have strong incentives to steal model weights and fine-tune them for harmful ends, so all frontier labs must use best-in-class cybersecurity, double layers of protection, hardened sandboxes, and eventually air gaps and secure hardware solutions to secure the weights.
“I think that’s key and I think all frontier labs should be doing that because otherwise for rogue nation-states and other dangerous actors, there would obviously be a lot of incentive for them to steal things like the weights.”
Scaling laws don't work by magic; at each new scale you must adjust hyperparameters and add innovations rather than repeat the same recipe, because extrapolating predictions several orders of magnitude can fail, capabilities can appear as step functions, and intermediate data points are needed to keep the scaling law true—making scaling an art form, with about one order of magnitude the maximum jump between eras.
“One order of magnitude is probably about the maximum that you want to do between each era.”
LLMs must keep improving as reliable world models—a necessary but insufficient component of AGI—and AlphaZero-like planning mechanisms should be built on top to make concrete plans, chain reasoning, and use search to explore massive possibility spaces, a capability currently missing from large models.
“That’s clearly a necessary, but probably insufficient component of an AGI system. On top of that, we’re working on things like AlphaZero-like planning mechanisms on top that make use of that model in order to make concrete plans to achieve certain goals in the world.”
Current analysis techniques are not sophisticated enough to localize where in a neural network shared improvements occur; the field needs better mechanistic analysis of learned representations, analogous to fMRI or single-cell recording in real brains, where computational neuroscience methods can be applied.
“I sometimes like to call it virtual brain analytics. In a way, it’s a bit like doing fMRI, or single-cell recording from a real brain. What are the analogous analysis techniques for these artificial minds?”
The strong scaling hypothesis has gone surprisingly far—large models are almost unreasonably effective and exhibit emergent properties like implicit concepts and abstractions that five years ago would have seemed to require an additional algorithmic breakthrough—but whether scaling hits an asymptote or brick wall remains an empirical question that should simply be tested while also doubling down on new invention.
“I look at the large models today and I think they’re almost unreasonably effective for what they are. I think it’s pretty surprising some of the properties that emerge.”
Core metrics like training loss can be predicted from much smaller models, but downstream capabilities such as MMLU or math performance don't necessarily follow linearly from loss—there are non-linear effects, so predictable loss does not translate into predictable real capabilities.
“You can often predict the core metrics like training loss or something like that, but then it doesn’t actually translate into MMLU, or math, or some other actual capability that you care about. They’re not necessarily linear all the time.”
Robotics is harder because it is a data-poor regime, but treating any input—an action, a word, an image pixel—as just a token (as in Gato) lets large models transfer to robotics, and true multimodality produces cross-modal benefit where understanding video improves language, so general multimodal systems will ultimately be more capable than single-modality ones despite being harder to train initially.
“just treat tokens like Gato as any type of token. The token could be an action, it could be a word, it could be part of an image, a pixel, or whatever it is. That’s what I think true multimodality is.”
Powerful current chatbots haven't yet automated large sections of the economy because we're at the beginning of the era; broader use cases need new capabilities like planning, search, personalization, and episodic memory—not just long context windows but actually remembering conversations from 100 conversations ago—and they aren't yet reliable enough to do science.
“for more general use cases we still need new capabilities, things like planning and search but also things like personalization and episodic memory. That’s not just long context windows, but actually remembering what we spoke about 100 conversations ago.”
Because intelligence is so broadly applicable, there must be high-level common algorithmic themes in how the brain processes the world, even though specialized brain regions handle specific tasks; underlying shared principles underpin general intelligence.
“I think that suggests there must be high-level common algorithmic themes around how the brain processes the world around us. Of course, there are specialized parts of the brain that do specific things, but I think there are probably some underlying principles that underpin all of that.”
If red-teaming or external testing discovers a dangerous latent capability such as helping a layperson build a pandemic-class bioweapon, mitigations could include a different constitution, additional guardrails, more RLHF, or removing training data, but the first requirement is detecting it ahead of time via the right evaluations, and the capability must be fixed before general deployment.
“that might require a different kind of constitution perhaps, or different guardrails, or more RLHF to avoid that. Or you could remove some training data, depending on what the problem is.”
Shane Legg always thought in terms of compute curves comparing roughly to the brain's number of neurons and synapses, and we are now in that regime with roughly the right order of magnitude of synapse-equivalent compute.
“we’re actually in that kind of regime now with roughly the right order of magnitude of number of synapses in the brain and the sort of compute that we have.”
As systems like Gemini ingest video and audiovisual data alongside text, the system correlates these modalities together, which constitutes a form of proper grounding and will let systems better understand real-world physics; ungrounded systems cannot achieve real-world goals properly.
“As we start ingesting things like video and audiovisual data as well as text data, then the system starts correlating those things together. I think that is a form of proper grounding.”
Many ideas from earlier deep reinforcement learning work like DQN and AlphaGo are coming back into fashion and should be recombined with the new advances in large multimodal models, with significant potential in merging older and newer ideas.
“I do actually think a lot of those ideas need to come back in again and, as we talked about earlier, combine them with the new advances in large models and large multimodal models”
There is large scope to overcome data bottlenecks through more multimodal/video data, synthetic data via realistic simulation, and self-play where systems interact and learn from each other's mistakes, as proven with AlphaGo and AlphaZero, though building general world data via self-play is more complicated.
“It worked very well for us with AlphaGo and AlphaZero where we got the systems to play against each other and actually learn from each other’s mistakes and build up a knowledge base that way.”
It is potentially possible that an AGI speeds up further AI research, since current LLMs are good at coding and systems like AlphaCode and theorem provers exist; combining and improving these could let systems help design future versions of themselves, but the safety implications must be considered.
“I could imagine these systems being quite good at designing and helping us build future versions of themselves, but we also have to think about the safety implications of that of course.”
Securing weights must be balanced against collaboration and progress speed, and against the need to expose systems to independent academic and government red teamers like the UK and US AI Safety Institutes (though not necessarily the weights themselves), with access granted only to those who need it.
“you want brilliant independent researchers from academia or things like the UK AI Safety Institute and the US one to be able to red team these systems. So one has to expose them to a certain extent, although that’s not necessarily the weights.”
Improving large models in a specific domain like coding can asymmetrically improve their general reasoning, indicating cross-domain transfer, mirroring how humans specialize through practice while using general learning systems.
“when these large models improve at coding, that can actually improve their general reasoning. So there is evidence of some transfer although we would like a lot more evidence of that.”
Although DeepMind strongly supports open source and open science (publishing AlphaFold, transformers, AlphaGo, GraphCast), for general-purpose foundational technology there is no compelling answer to how one stops bad actors—individuals up to rogue states—from taking open-sourced systems and repurposing them for harmful ends, so some balance is required.
“when it comes to the general-purpose foundational technology, I think the question I would have for open source proponents is, how does one stop bad actors, individuals or up to rogue states, taking those same open source systems and repurposing them for harmful ends?”
Given the transformative power and huge uncertainty of AI, the only sensible approach is cautious optimism using the scientific method to gain foresight before consequences hit, avoiding live A/B testing in the world because unintended consequences may be severe—deliberately moving the field away from Silicon Valley's 'move fast and break things' attitude toward being 'bold and responsible.'
“the only sensible approach when you have huge uncertainty is to be cautiously optimistic and use the scientific method to try and have as much foresight and understanding about what’s coming down the line”
Contrary to perceptions that DeepMind is less compute-efficient, Gemini 1 used roughly the same amount of compute, maybe slightly more, than what was rumored for GPT-4, and DeepMind uses compute efficiently across both scaling and new invention, with Google expected to have by far the most compute of any research lab this year.
“I think that actually Gemini 1 used roughly the same amount of compute, maybe slightly more, than what was rumored for GPT-4.”
DeepMind was started in 2010 as a 20-year project and is on track, so AGI-like systems within the next decade would not be surprising, though Hassabis declines to give specific numbers due to many unknowns and human ingenuity producing surprises.
“when we started DeepMind back in 2010, we thought of it as a 20-year project. And I think we’re on track actually... So I wouldn’t be surprised if we had AGI-like systems within the next decade.”
Human brains do not do Monte Carlo tree search; geniuses like Einstein compensate by building extremely accurate intuitive world models with mental simulations—visualizing and feeling physical systems rather than just the math—which is analogous to a model getting you to a high-quality leaf node so only a little search is needed.
“our brains are not built for doing Monte Carlo tree search. It’s just not the way our organic brains work.”
Rather than waiting for AGI, building domain-specific solutions delivers benefits now and battle-tests research ideas, because real-world applications give direct feedback that prevents internal metrics from deviating from real-world impact and reveals whether systems truly scale or need more data/sample efficiency—keeping research honest.
“it battle tests your ideas. You don’t want to be in a research bunker where you theoretically are pushing things forward, but then actually your internal metrics start deviating from real-world things that people would care about”
The science of data curation and analysis—identifying gaps and holes in the data distribution and ensuring it represents the distribution to be learned—is still nascent; tricks like overweighting or replaying parts of data, or targeting synthetic generation at identified gaps, can address fairness, bias, and coverage.
“I think we’re still in the nascent stage of this, of data curation and data analysis and actually analyzing the holes that you have in your data distribution.”
Interacting with full multimodal model systems will be quite different from today's chatbots; the next versions over the coming 12-18 months will gain contextual understanding of the environment via cameras, phones, or glasses, becoming more fluid by sampling from video, using voice, and eventually touch and robotic sensors.
“The next versions of this over the next year, 18 months, we’ll maybe have some contextual understanding of the environment around you through a camera or a phone or some glasses.”
AGI progress was already priced into Hassabis's world model technologically, but the general public's intense interest this early—prompted by ChatGPT and chatbots being adopted despite their limitations—was surprising; had that not happened, DeepMind would have produced more specialized off-main-track systems like AlphaFold and AlphaGo, and the public would have engaged later.
“we didn’t necessarily anticipate that the general public would be so interested this early in the sequence.”
Today's systems are not dangerous, but if in a few years a system in a safe sandbox displays a new unexpected capability or does something explicitly told not to do and lies about it, one would ideally pause and get to the bottom of why before continuing.
“We explicitly told the system we didn’t want it but then it did and it lied about it. These are the kinds of things where one would want to then dig in carefully... Then you would ideally pause and really get to the bottom of why it was doing those things before one continued.”
Google DeepMind splits roughly half its effort on scaling and half on inventing new architectures and algorithms, betting that reaching AGI requires pushing both as hard as possible because a new invention is only useful if it can also scale.
“You can think of half our effort as having to do with scaling and half our efforts having to do with inventing the next architectures and the next algorithms that will be needed”
Top human players use a much richer, more accurate world model of Go or chess than AlphaGo, while brute-force engines have no real model beyond game heuristics, which is why humans achieve world-class decisions with only a few hundred move evaluations.
“AlphaGo has quite a decent model but the top human players have a much richer, much more accurate model of Go or chess. That allows them to make world-class decisions on a very small amount of search.”
Aligning systems smarter than humans has been a forefront concern since before DeepMind's founding because they planned for success, anticipating 20 years ago that the technology would be unbelievably transformative with both positive (science, AlphaFold, health) and negative consequences, requiring understandable and controllable systems.
“This is something Shane and I, and many others here, have had at the forefront of our minds since before we started DeepMind because we planned for success.”
Google DeepMind already has internal checks and safety councils and will start publishing blog posts and technical papers along the lines of responsible scaling laws over the course of the year, making its internal pre-commitments public.
“We have those implicitly internally in various safety councils that people like Shane chair... it’s time for us to talk about that more publicly I think. So we’ll be doing that throughout the course of the year.”