Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
What this covers
Here is my episode with Demis Hassabis, CEO of Google DeepMind. We discuss: * Why scaling is an artform * Adding search, planning, & AlphaZero type training atop LLMs * Making sure rogue nations can't steal weights * The right way to align superhuman AIs and do an intelligence explosion
𝐄𝐏𝐈𝐒𝐎𝐃𝐄 𝐋𝐈𝐍𝐊𝐒 * Transcript: https://www.dwarkeshpatel.com/p/demis-hassabis * Apple Podcasts: https://podcasts.apple.com/us/podcast/demis-hassabis-scaling-superhuman-ais-alphazero-atop/id1516093381?i=1000647410338 * Spotify: https://open.spotify.com/episode/6SWbwjYPs5WevIoCCiSByS?si=nCVFSRr7QGGI_STgbrOBDA * Follow me on Twitter: https://twitter.com/dwarkesh_sp
𝐓𝐈𝐌𝐄𝐒𝐓𝐀𝐌𝐏𝐒 00:00:00 - Nature of intelligence 00:05:56 - RL atop LLMs 00:16:31 - Scaling and alignment 00:24:13 - Timelines and intelligence explosion 00:28:42 - Gemini training 00:35:30 - Governance of superhuman AIs 00:40:42 - Safety, open source, and security of weights 00:47:00 - Multimodal and further progress 00:54:18 - Inside Google DeepMind
Source description (no synthesized summary yet).
Hassabis argues that AGI requires combining large multimodal models as world models with planning/search mechanisms similar to AlphaZero, grounded through multimodal learning and real-world interaction, while emphasizing the need for safety evaluations and responsible scaling before deployment.
- LLMs alone are necessary but insufficient for AGI; they need planning and search on top to explore solution spaces efficiently
- Grounding comes from multimodal data (video, audio, text) and reinforcement learning with human feedback, not language alone
- Safety requires pre-commitment through responsible scaling policies, red-teaming, and evaluations before capability thresholds are crossed
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.
Historical papers from Shane and Hassabis (2009 and 2016) anticipated key developments: Shane's thesis identified compression of Wikipedia as a test for AI (which became the loss function for LLMs), and Hassabis's 2016 neuroscience-AI paper identified attention as a critical capability (which was later crystallized in transformers).
“It's fascinating if you read your old papers or Shane's old papers. In Shane's thesis in 2009, he said 'well, the way we would test for AI is, can you compress Wikipedia?' And that's literally, the loss function for LLMs. Or in your own paper in 2016 before transformers, you were comparing neuroscience and AI and you said attention is what is needed.”
As AI systems become more capable than humans in specialized domains, directly specifying reward functions or goals becomes harder because human raters cannot verify outputs (e.g., a million-line pull request).
“One is that as these models get smarter, they are going to be able to operate in domains where we just can't generate enough human labels, just because we're not smart enough. If it does a million-line pull request, how do we tell it, for example, this is within the constraints of our morality and the end goal we wanted and this isn't?”
There must be a balance between securing AI weights and allowing collaboration with brilliant independent researchers from academia and safety institutes for red-teaming and security validation.
“One has to balance that with allowing for collaboration and speed of progress. Another interesting thing is that of course 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.”
Designing and deploying superhuman AI systems requires evaluations for deceptive capabilities (systems lying about their reasoning), external code exfiltration, and other undesirable behaviors before deployment.
“There are many, many ideas that people have such as more stringent eval systems. 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?”
Human exceptional achievement like Einstein's development of relativity is achieved through intuition built on extremely accurate world models and mental simulations rather than through Monte Carlo tree search; Einstein visualized physical systems, felt what they should be like, and had an intuitive understanding of physics that allowed him to think through possibilities with minimal explicit search.
“I think it's different because our brains are not built for doing Monte Carlo tree search. It's just not the way our organic brains work. I think that people like Einstein, in order to compensate for that, have used their intuition—and maybe we can come to what intuition is—and their knowledge and their experience to build in Einstein's case, extremely accurate models of physics that include mental simulations. If you read about Einstein and how he came up with things, he used to visualize and really feel what these physical systems should be like, not just the mathematics of it. He had a really intuitive feel for what they would be like in reality. That allowed him to think these thoughts that were very outlandish at the time.”
AlphaFold represents a concrete example of how AI for science creates measurable impact on healthcare and biological understanding, serving as template for other scientific applications.
“The good news is that with the popularity of the recent chatbot systems, I think that has woken up many of these other parts of society to the fact that this is coming and what it will be like to interact with these systems. And that's great. It's opened up lots of doors for very good conversations. An example of that was the safety summit the UK hosted a few months ago, which I thought was a big success in getting this international dialogue going.”
Transformers, invented by Google Research and Brain colleagues, enabled scaling to massive amounts of data, fundamentally accelerating modern AI progress and representing a key inflection point.
“Of course then, with the advent of transformers, invented by our colleagues at Google Research and Brain, that was the type of deep learning that allowed us to ingest masses of amounts of information. That has really turbocharged where we are today.”
Public interest in chatbots (like ChatGPT) arrived earlier and more intensely than DeepMind anticipated, creating a different research environment with more venture capital and urgency than if systems had developed more quietly.
“We didn't necessarily anticipate that the general public would be so interested this early in the sequence. If ChatGPT and chatbots hadn't gotten the interest they ended up getting—which I think was quite surprising to everyone that people were ready to use these things even though they were lacking in certain directions, impressive though they are—then we would have produced more specialized systems built off of the main track, like AlphaFold and AlphaGo, our scientific work.”
AI governance should be a broad collaboration involving civil society, academia, government, and private industry, not controlled by individual companies or industries.
“I think that this is so consequential, this technology. 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.”
A critical research challenge is identifying and filling gaps in model training data distributions, including detecting bias, fairness issues, and underrepresented domains, then using synthetic generation to create data for those gaps.
“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. This is important for things like fairness and bias and other stuff.”
Hardened sandboxes and simulations with cybersecurity measures can contain AI systems and enable safer experimentation, preventing both unauthorized AI escape and external hacking.
“I think that there's a lot of promise in creating hardened sandboxes or simulations that are hardened with cybersecurity arrangements around the simulation, both to keep the AI in and to keep hackers out. You could experiment a lot more freely within that sandbox domain.”
Current analysis techniques are not sophisticated enough to mechanistically identify which neurons or components in neural networks are responsible for improvements in specific skills like language and code, requiring development of 'virtual brain analytics' approaches.
“I don't think our analysis techniques are quite sophisticated enough to be able to hone in on that. I think that's actually one of the areas where a lot more research needs to be done, the kind of mechanistic analysis of the representations that these systems build up.”
Robotics remains a valuable research domain despite being data-poor because it forces innovation in sampling efficiency, data efficiency, transfer learning, and sim-to-real transfer that benefits all AI systems.
“We still have that going now because we like the fact that it's a data-poor regime. That pushes us in very interesting research directions that we think are going to be useful anyway: sampling efficiency and data efficiency in general, transfer learning, learning from simulation and transferring that to reality, sim-to-real.”
All frontier AI labs should prioritize cybersecurity of model weights because rogue nation-states and dangerous actors have strong incentives to steal advanced AI systems.
“I think that 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.”
If a system is not properly grounded, it will not be able to achieve goals in the real world properly; grounding is a prerequisite for functional goal achievement in physical domains.
“I think if it's not properly grounded, the system won't be able to achieve those goals properly. In a sense, you have to have some grounding for a system to actually achieve goals in the real world.”
Intelligence likely rests on high-level common algorithmic themes about how the brain processes the world, despite specialized brain regions handling specific tasks.
“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.”
Large language models exhibit unreasonably effective emergent properties including implicit concept formation and grounding without explicit multimodal training, which was surprising to many including scaling hypothesis proponents.
“I think it was pretty surprising to almost everyone, including the people who first worked on the scaling hypotheses, how far it's gone. In a way, I look at the large models today and I think they're almost unreasonably effective for what they are.”
Neuroscience has provided directional inspiration for key AI breakthroughs including reinforcement learning, attention mechanisms, and the general principle that general intelligence is possible, even though the specific implementations in AI systems don't map one-to-one to brain mechanisms.
“In the earlier days of the new wave of AI, neuroscience was providing a lot of interesting directional clues, things like reinforcement learning and combining that with deep learning. Some of our pioneering work we did there were things like experience replay and even the notion of attention, which has become super important.”
Mechanistic understanding of AI system representations through analysis and interpretability—making systems less alien and more transparent—is a key safety requirement for responsible scaling.
“the analysis stuff we talked about earlier, where we can analyze and understand what the concepts are that this system is building and what the representations are. So maybe then they're not so alien to us and we can actually keep track of the kind of knowledge that it's building.”
Real-world applications and products drive research progress by providing feedback on whether systems work in practice, preventing deviation from concrete impact and keeping research grounded in reality.
“This is where things like the sandbox simulations are important. I would hope we're experimenting in a safe, secure environment when something very unexpected happens. Because most real-world challenges require that. So it kind of keeps you honest and pushes you to keep nudging and steering your research directions to make sure they're on the right path. So I think it's fantastic. Of course, the world benefits from that.”
Games have been used as a proving ground for AI research partly because research is more efficient in that domain, but more importantly because reward functions are extremely easy to specify in games (winning or improving score), whereas real-world systems face the challenge of defining the right objective function and reward function in a general but sufficiently specific way to guide learning.
“Of course that's why we pioneered, and what DeepMind is sort of famous for, using games as a proving ground. That's partly because it's efficient to research in that domain. The other reason is, obviously, it's extremely easy to specify a reward function. Winning the game or improving the score, something like that is built into most games. So that is one of the challenges of real-world systems. 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?”
Deception is a 'root node trait' that enables cascading failures; if a system is non-deceptive and transparently exposes what it actually thinks, this opens possibilities for the system to explain itself and its reasoning to humans.
“That's why 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.”
Open-sourcing foundational AI systems creates risks that bad actors will repurpose them for harmful ends, and proponents of open-source AI have not provided compelling answers to how these risks would be mitigated.
“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? We have to answer that question. I don't know what the answer is to that, but I haven't heard a compelling, clear answer to that from proponents of just open sourcing everything.”
Large language models need planning and search mechanisms on top of improved world models—similar to AlphaZero—to explore solution spaces efficiently and achieve AGI capabilities, rather than relying on LLMs alone.
“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.”
Reinforcement learning with synthetic data—through simulation, self-play between systems, and multimodal data generation—can overcome data bottlenecks and provide the scalable learning signal needed for advanced AI systems.
“I'm very optimistic about that. First of all, there's still a lot more data that can be used, especially if one views multimodal and video and these kinds of things.”
Large multimodal models can treat different modalities as tokens (words, pixels, actions, etc.), enabling transfer across domains; training such systems is harder than pure language models, but the resulting multimodality produces better performance across all modalities as they learn from each other's structure.
“They can learn in the general domain, language domain and other things, and then 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. To begin with, it's harder to train a system like that than a straightforward language system. But going back to our early conversation on transfer learning, you start seeing that with a true multimodal system, the other modalities benefit some different modalities. You get better at language because you now understand a little bit about video. So I do think it's harder to get going, but ultimately we'll have a more general, more capable system like that.”
Systems will show asymmetric progress across domains, with math and code becoming superhuman while other capabilities remain below human performance, because math and code have clear specification and objective functions while scientific creativity requires asking the right question.
“I think that we're making great progress with math and things like theorem proving and coding. But it's still interesting if one looks at creativity in general, and scientific endeavor in general. I think we're getting to the stage where our systems could help the best human scientists make their breakthroughs quicker, almost triage the search space in some ways. Perhaps find a solution like AlphaFold does with a protein structure. 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.”
Memory is reconstructive, not reproductive—humans rebuild memories from familiar semantic components rather than retrieving exact recordings—which provides insight into how imagination uses similar mechanisms.
“We sort of put it together back from components that seem familiar to us, the ensemble. That's what made me think that imagination might be the same thing.”
Improving world models enables more efficient search because accurate models allow systems to identify and prioritize the most promising solution branches, reducing the total search space that must be explored.
“If you improve the models, then I think your search can be more efficient and therefore you can get further with your search.”
Scaling laws don't automatically hold across orders of magnitude; new capabilities can emerge as step functions rather than smooth curves, requiring intermediate training runs to verify predictions and adjust hyperparameters.
“Scaling laws also don't just work by magic. You still need to scale up the hyperparameters, and various innovations are going in all the time with each new scale. It's not just about repeating the same recipe at each new scale.”
Gemini's multimodal capabilities represent the beginning of understanding how systems can interact with environment through camera, phone, glasses, voice, and potentially touch sensors, expanding beyond text-only chatbots.
“I think we're just at the beginning of actually understanding how exciting that might be to interact with a full multimodal model system. It'll be quite different from what we're used to today with the chatbots. I think 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.”
The most important step going forward is to move the AI field away from 'move fast and break things' culture toward cautious optimism using the scientific method to understand and mitigate risks before deployment.
“So I want us to move away, as a field, from a sort of 'move fast and break things' attitude which has maybe served the Valley very well in the past and obviously created important innovations. I think in this case we want to be bold with the positive things that it can do and make sure we advance things like medicine and science whilst being as responsible and thoughtful as possible with mitigating the risks.”
The brain constructs world models through imagination or mental simulation—creating rich visual-spatial simulations used for planning—a capability that remains largely unresolved in understanding how the brain works.
“As for going forward, I think there's still a lot of interesting things to be resolved around planning. How does the brain construct the right world models? I studied how the brain does imagination, or you can think of it as mental simulation. How do we create very rich visual spatial simulations of the world in order for us to plan better?”
Responsible scaling policies that commit in advance to safety checks and capability thresholds—empirical governance frameworks—provide a more principled approach than reactive responding to crises.
“That's why it seems like the responsible scaling policies are something that are a very good empirical way to pre-commit to these kinds of things. Yes, exactly.”
Gemini required only roughly the same amount of compute as GPT-4, or slightly more, making it approximately as compute-efficient as OpenAI's latest system.
“I think that actually Gemini 1 used roughly the same amount of compute, maybe slightly more, than what was rumored for GPT-4. I don't know exactly what was used but I think it was in the same ballpark.”
AGI systems could accelerate AI research by helping design and improve future versions of themselves through coding and mathematical reasoning, though this creates additional safety considerations.
“I think that's potentially possible. I think it partly depends on what we, as a society, decide to use the first nascent AGI systems or proto-AGI systems for. Even the current LLMs seem to be pretty good at coding and we have systems like AlphaCode.”
Narrowly specialized AI systems can assist human scientists in analyzing and explaining what more general systems are doing, serving as interpretability tools rather than replacements.
“There are also ideas of using AI, not general learning ones but maybe narrow AIs that are specialized for a domain, to help us as the human scientists to analyze and summarize what the more general system is doing. So there's narrow AI tools.”
Current systems cannot reliably perform scientific research without improvements to factuality and grounding, but once fixed, AI could become the world's best research assistant for scientists and clinicians.
“They're not reliable yet enough to do things like science with them. But I think one day, once we fix factuality and grounding and other things, I think they could end up becoming the world's best research assistant for you as a scientist or as a clinician.”
The broader AI research field is experiencing 'chaos' with excessive venture capital and intense competition, which risks moving away from responsible, scientific, and thoughtful approaches to development.
“It's a little bit more chaotic because there's so many more things going on, and there's so much VC money going into it, and everyone's sort of almost losing their minds over it.”
Superhuman AI systems could explain their reasoning just as chess grandmasters can explain their moves even if the human listener cannot independently discover those moves, through post-hoc explanation and mathematical proof.
“If I were to play a game of chess against Garry Kasparov, which I've played in the past, Magnus Carlsen, or the amazing chess players of all time, I wouldn't be able to come up with a move that they could. But they could explain to me why they came up with that move and I could understand it post hoc, right?”
Reinforcement learning and deep reinforcement learning are examples of research ideas that were underrated by the field for years despite early promise, then came back into focus at scale; similar to how scaling was underrated before recent successes, there are likely current research directions that are underrated but will be important when combined with new advances.
“Nowadays, people are paying attention to the RL stuff that DeepMind did many years before. What are the early research directions, or something that was done way back in the past, that you think will be a big deal but people just haven't been paying attention to it? There was a time where people weren't paying attention to scaling. What's the thing now that is totally underrated? Well, I think that the history of the last couple of decades has been things coming in and out of fashion, right? A while ago, maybe five-plus years ago, we were pioneering with AlphaGo and before that DQN. It was the first system that worked on Atari, our first big system really more than ten years ago now, that scaled up Q-learning and reinforcement learning techniques and combined that with deep learning to create deep reinforcement learning.”
Both founders of DeepMind prioritized safety considerations from the beginning in 2010 when AGI was not widely discussed, because they planned for eventual success and understood the transformative implications.
“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. In 2010, no one was thinking about AI let alone AGI. But we already knew that if we could make progress with these systems and these ideas, the technology created would be unbelievably transformative.”
DeepMind's approach was grounded from inception (2010) in three types of algorithms: reinforcement learning, search, and deep learning—algorithms chosen for their scalability, generality, and minimal requirement for handcrafted human priors, contrasting with earlier AI approaches in the 1990s that relied on hand-coded knowledge.
“We thought that actually, and I know you've interviewed my colleague Shane. He always thought in terms of compute curves and comparing it roughly to the brain, how many neurons and synapses there are very loosely. Interestingly, 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. But I think more fundamentally, we always thought that we bet on generality and learning. So those were always at the core of any technique we would use. 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.”
DeepMind's timeline for AGI-like systems is roughly aligned with their original 20-year estimate from 2010, suggesting they would not be surprised by AGI-level systems within the next decade.
“When we started DeepMind back in 2010, we thought of it as a 20-year project. And I think we're on track actually, which is kind of amazing for 20-year projects because usually they're always 20 years away.”
DeepMind uses significant compute not just for scaling but for testing innovations and new architectures, as new techniques only prove valuable if they scale, and some innovations only work at large scales.
“I think we're very efficient with our compute and we use our compute for many things. One is not just the scaling but, going back to earlier, more innovations and ideas. A new innovation, a new invention, is only useful if it can also scale.”
The merger of Brain and DeepMind created synergies through resource pooling, idea sharing, and deeper engineering collaboration than project-by-project collaboration, accelerating frontier model development.
“Of course it's been challenging to do, like any big integration coming together. You're talking about two world-class organizations with long, storied histories of inventing many important things from deep reinforcement learning to transformers. So it's very exciting to actually pool all of that together and collaborate much more closely.”
During Gemini development, there were no single big surprises but rather ongoing learning about managing large-scale training systems from organizational and tracking perspectives, and continued refinement of understanding the mapping between metrics being optimized and final downstream capabilities desired.
“I wouldn't say there was one big surprise. It was very interesting trying to train things at that size and learning about all sorts of things from an organizational standpoint, like how to babysit such a system and to track it. There's also things like getting a better understanding of the metrics you're optimizing versus the final capabilities that you want. I would say that's still not a perfectly understood mapping, but it's an interesting one that we're getting better and better at.”
Weight access management will become more important and restrictive as systems become more powerful and general, though currently it's still early enough that broader access is possible.
“Right now, I think we're still in the early days of those kinds of systems being at risk. As these systems become more powerful and more general and more capable, I think one has to look at the access question.”
DeepMind has internally maintained safety councils and checks on scaling (implicit policies) for years, and is now planning to publish these approaches publicly through blog posts and technical papers.
“We have those implicitly internally in various safety councils that people like Shane chair and so on. But it's time for us to talk about that more publicly I think.”
Google DeepMind has world-class cybersecurity defenses through Google's firewall and cloud protection, with additional DeepMind-specific protections, though vigilance must be maintained and further measures like air gaps are planned.
“First of all, there's two parts. One is security, one is open source, which maybe we can discuss. The security is super key just as normal cybersecurity type things. I think we're lucky at Google DeepMind. We're behind Google's firewall and cloud protection which I think is best in class in the world corporately.”
Current large language model systems are not dangerous in Hassabis's opinion, but in a few years they may have potential danger as capabilities increase; careful empirical monitoring is needed before that window arrives.
“The systems that are around today are not dangerous, in my opinion, but in a few years they might have potential.”
Google and Alphabet take AI safety and responsibility questions very seriously as part of organizational culture, ensuring DeepMind's safety priorities are supported by parent company values.
“This is one of the reasons we joined forces with Google back in 2014. I think the entirety of Google and Alphabet, not just Brain and DeepMind, takes these questions of responsibility very seriously.”