Sep 12, 2025 · 31m · allin
Inside Google DeepMind: AGI, Robotics, & World Models Explained - Demis Hassabis
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In an in-depth conversation at the All-In Summit, Google DeepMind CEO Sir Demis Hassabis discusses winning the Nobel Prize, the development of world models and robotics, and the core scientific breakthroughs required to achieve Artificial General Intelligence (AGI) over the next decade.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 4.1 Guest teaching 2.6 Guest disagreement 1.0 The hosts pushing back 0.6
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 18:41 Dismissing competitor PhD intelligence claims
Demis directly refutes rival AI leaders who claim current LLMs are PhD intelligences, labeling the claim nonsense and pointing out basic errors in high school math.
Hardest push from the hosts ▶ 20:00 Challenging model progress flatliningThe host directly confronts Demis with reports indicating a performance flatlining and convergence across frontier large language models.
Biggest teaching moment ▶ 26:46 Explaining hybrid deterministic and probabilistic systemsDemis corrects the host's strict binary separation between deterministic and probabilistic models by explaining how hybrid AI systems hardcode physical constraints when training data is scarce.
The host holds their own ▶ 25:59 Drilling into molecular model architecture dynamicsThe host showcases deep technical insight by asking how probabilistic neural networks integrate with deterministic physical and chemical laws during drug discovery.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Recounting the Nobel Prize Win | 1 | 1 | 0 | 0 | The host opens with warm congratulations on the Nobel Prize and asks Demis to recount receiving the news. The tone is purely celebratory and friendly with no pushback or technical drilling. | |
| Genie World Model Showcase and Real-Time Interactivity | 4 | 2 | 0 | 0 | The host guides the conversation through a video showcase of the Genie world model, highlighting how it differs from conventional rendering engines by generating 2D pixels in real time. | |
| How World Models Learn Intuitive Physics vs. Game Engines | 5 | 3 | 0 | 0 | The host articulates how 3D physics engines explicitly program light reflection and laws of motion, comparing it to how Genie inferred intuitive physics from raw video data. Demis validates this perspective while drawing on his 1990s game developer background. | |
| General Robotics: Vision-Language-Action Models and Form Factors | 4 | 2 | 0 | 0 | The host introduces the concept of Vision-Language-Action models and suggests an analogy to an Android operating system layer for general robotics. Demis confirms the vision and elaborates on Google's dual OS and vertical integration strategies. | |
| Humanoid Robots vs. Specialized Form Factors and Hardware Scaling | 4 | 2 | 1 | 1 | The host questions the practicality of humanoid form factors compared to specialized task robots and offers a computing history comparison to 1970s PC DOS. Demis gently notes that in AI hardware development, a decade of progress happens within a single year. | |
| AI in Scientific Discovery and the Definition of Creativity | 3 | 3 | 0 | 0 | The host prompts Demis to define human scientific creativity versus current machine limits. Demis sets benchmarks such as restricting AI knowledge to 1901 to test if it can discover special relativity like Einstein did in 1905. | |
| Timeline to AGI and Missing Technological Breakthroughs | 4 | 4 | 5 | 2 | The host asks Demis to contrast his AGI timeline with competitors like Sam Altman and Dario Amodei. Demis forcefully rejects rival claims that current systems are PhD-level intelligences, calling such claims nonsense and pointing out simple high school math failures. | |
| LLM Progress, Multimodality, and Creative Image Tools | 5 | 2 | 3 | 2 | The host brings up industry reports alleging model performance flatlining and demonstrates personal domain familiarity with early graphics tools like Kai's Power Tools and Bryce. Demis rejects the flatlining premise, explaining internal progress across multimodal tools. | |
| Isomorphic Labs and Revolutionizing Drug Discovery | 6 | 4 | 0 | 0 | The host poses a detailed architectural question asking whether molecular drug discovery requires deterministic physical rules rather than purely probabilistic models. Demis educates the host on hybrid architectures like AlphaFold and AlphaGo. | |
| AI Energy Demand, Model Efficiency, and Climate Solutions | 5 | 3 | 1 | 1 | The host frames a nuanced energy question regarding model distillation and per-token efficiency versus geometric compute demand. Demis clarifies that model serving has gotten 10x-100x more efficient while frontier scaling drives net power demands. |
Statements from this episode (24)
Hassabis: Nobel Committee notifies winners 10 minutes before public announcement
“The way they tell you like 10 minutes before it all goes live.”
Hassabis says billions of people now interact with Google Gemini models
“So, you know, billions of people now interact with Gemini models, whether that's through AI overview, AI mode, or the Gemini app.”
Hassabis: Google DeepMind has 5,000 staff, 80%+ engineers and PhDs
“There's around 5000 people in, in, in my org, in, in Google DeepMind, and, you know, it's predominantly, I guess, 80% plus engineers and PhD researchers.”
DeepMind's Genie generates one to two minutes of consistent interaction
“It's not perfect yet, but it can generate a consistent minute or two of interaction as you as the user in many, many different worlds.”
Hassabis: DeepMind's Genie was trained on video and game engine synthetic data
“It was trained off a video and some synthetic data from game engines.”
Hassabis: AGI requires physical world understanding, not just text and math
“For an AI to be truly general, to build AGI, we feel that the AGI system needs to understand the world around us and the physical world around us, not just the abstract world of languages or mathematics.”
Google DeepMind pursues Android-style OS and hardware-integrated robotics
“That's certainly one strategy we're pursuing is a kind of Android play, if you like, a crossroad as a kind of robotics, almost an OS layer, cross robotics. But there's also some quite interesting things about vertically integrating our latest models with speci…”
Hassabis: Humanoid form factors are crucial for general robotics in human environments
“On the other hand, for general use or personal use robotics and just interacting with the ordinary world the humanoid form factor could be pretty important because, of course, we've designed the physical world around us to be for humans. And so steps, doorways…”
Hassabis predicts a major breakthrough in general robotics within two years
“I think in the next couple of years, there'll be a sort of real wow moment with robotics, but I think the algorithms need a bit more development. The general purpose models that these robotics models are built on still need to be better and more reliable and b…”
Hassabis: Present-day AI lacks true creativity to formulate new theories
“AI today, I would say doesn't have true creativity in the sense that it can't come up with a new conjecture yet or a new hypothesis. It can maybe prove something that you give it but it's not able to come up with a sort of new idea or new theory itself.”
Hassabis proposes testing AGI by asking it to discover special relativity
“A good test for it would be something like give one of these modern AI systems, a knowledge cutoff of 19 oh one and see if it can come up with special relativity like Einstein did in 19 oh five, right? If it's able to do that, then I think we're onto something…”
Hassabis dismisses competitor claims of PhD-level AI intelligence as nonsense
“So you often hear some of our competitors talk about you know, these modern systems that we have today are PhD intelligences. I think that's a nonsense. They're not PhD intelligences. They have some capabilities that are PhD level but they're not in general ca…”
Hassabis predicts true AGI is five to ten years away
“So I think that we are maybe, you know, I would say sort of five to 10 years away from having an AGI system that's capable of doing those things.”
Hassabis predicts achieving AGI requires one or two fundamental new breakthroughs
“And so, a lot of these, I think, core capabilities are still missing, and maybe scaling will get us there, but I feel, if I was to bet, I think there are probably one or two missing breakthroughs that are still required and will come over the next five, five o…”
Hassabis denies rumors of flatlining LLM progress at Google DeepMind
“No, I mean, we're not seeing that internally, and we're still seeing a huge rate of progress.”
Director Darren Aronofsky is making films using DeepMind's Veo AI
“People like the director, Darren Aronofsky, who's a good friend of mine, an amazing director. And he's been making, and his team is making films using VO and some of our other tools.”
Hassabis predicts new entertainment formats featuring audience co-creation
“I actually foresee a world, and I think a lot about this, having started in the games industry as a game designer and programmer, is that in the nineties is that, you know, I think the future of entertainment, this is what we're seeing is the beginning of the …”
Hassabis predicts AI could shorten drug discovery to days within a decade
“I think we could reduce down drug discovery from taking years, sometimes a decade to do, down to maybe weeks, or even days over the next 10 years.”
Hassabis expects Isomorphic Labs to enter pre-clinical drug phase in 2026
“We're building up the platform right now, and it's, ah, we have great partnerships with Eli Lilly, I think you had, ah, the CEO speaking earlier, and Novartis, which are fantastic, and our own internal drug programs, and I think we'll be entering sort of pre-c…”
DeepMind will likely build hybrid AI models for the next five years
“Actually we, for the moment, and I think probably for the next five years or so, we're building what maybe you could call hybrid models.”
Hassabis: AlphaFold is a hybrid model incorporating physics and chemistry rules
“So AlphaFold itself is a hybrid model where you have the learning component, this probabilistic component you're talking about, which is, you know, based on neural networks and transformers and things, and that's learning from the data you give it you know, an…”
Hassabis: End-to-end AI learning from data is always preferable to hybrid systems
“So it's always better if you can do end-to-end learning and directly predict the thing that you're after from the data that you, you're given.”
Hassabis says AI model serving efficiency improved 10x-100x over two years
“Over the last two years the model efficiencies are like 10 X, you know, even a hundred X better for the same performance.”
Hassabis predicts AI energy benefits will outweigh consumption within a decade
“I think AI systems will give back a lot more to energy and climate change and these kind of things than they take in terms of efficiency of grid systems and electrical systems, material design, new types of properties, new energy sources. I think AI will help …”