This video features an in-depth conversation between Demis Hassabis, co-founder and CEO of Google DeepMind, and Professor Hannah Fry about the present and future of artificial intelligence (AI). They review the remarkable developments in AI throughout 2025 and deeply discuss the roadmap toward Artificial General Intelligence (AGI) -- solving fundamental problems in fusion energy and materials science, advances in world models and simulation, and the importance of building ethical AI. The conversation also offers Hassabis's personal views on maintaining scientific rigor amid the competitive dynamics of AI development and AGI's potential societal impact.

1. 2025 AI Progress and Solving 'Root Node' Problems

2025 has been a truly remarkable year for AI, says Demis Hassabis -- as if a decade of progress was compressed into a single year. Google DeepMind's Gemini 3 showed notable achievements in multimodal capabilities. But the most exciting development, he emphasizes, is the advance of world models.

Professor Hannah Fry recalled the 'root node problem' Hassabis had previously mentioned -- the idea that AI could solve fundamental problems in science and medicine, yielding cascading benefits. Hassabis explained:

"Of course, the biggest success story was AlphaFold. It's hard to believe it's been almost five years since AlphaFold 2 was released to the world. AlphaFold proved that solving these 'root node' type problems is possible. We're now exploring all the other root node problems."

DeepMind is currently focused on the following major root node problems:

• Materials science: Developing new materials such as room-temperature superconductors and better batteries.
• Nuclear fusion energy: Collaborating with Commonwealth Fusion to help with plasma control and materials design acceleration for tokamak fusion reactors. Hassabis called fusion the "holy grail," noting that modular fusion reactors could provide unlimited clean energy and significantly contribute to solving climate problems.
• Quantum computing: Collaborating with Google's quantum AI team to develop error correction codes using machine learning.

Hassabis emphasized the cascading effects: if these root node problems, especially fusion energy, are solved, "if energy becomes truly renewable, clean, and very cheap or almost free, then a lot of other things become possible. For example, you could build desalination plants almost anywhere to solve water scarcity, and even use it to make rocket fuel."

2. 'Jagged Intelligence' and the Importance of Consistency

A question arose about the "paradox" of AI winning International Mathematical Olympiad medals while making mistakes on basic high-school math. Hassabis called this 'jagged intelligence' and described it as one of the critical problems to solve on the path to AGI.

"One of the main reasons I think we haven't achieved AGI is that this part needs to be fixed. As you said, we've had incredible successes like winning gold at the International Mathematical Olympiad. But on the other hand, as we all experience experimenting with chatbots in daily life, they make fairly trivial mistakes on logic problems."

Current AI systems show uneven performance -- PhD-level ability in some areas but below high-school level in others. This inconsistency may stem from:

• Image recognition and tokenization: Issues with how images are recognized and converted into tokens, sometimes failing to count letters correctly.
• Lack of consistency in reasoning and thinking processes: While AI is evolving toward systems that spend more time 'thinking,' this thinking time isn't being consistently used to double-check information and verify with tools.

Hassabis emphasized that consistency, especially in reasoning and thinking processes, is a core element of AGI, estimating the current consistency level at only about 50%. This is like a person blurting out whatever comes to mind; AI needs the ability to pause, review what it's about to say, and adjust.

3. From AlphaGo to AlphaZero, and Future Learning Methods

Professor Fry asked whether the 'AlphaGo-to-AlphaZero' approach -- where removing human experience actually improved performance -- could work in science or mathematics.

Hassabis explained that current large language models (LLMs) and foundation models learn by utilizing all human knowledge, like everything on the internet -- similar to AlphaGo. However, the ability to use that learned knowledge to direct useful reasoning processes, make plans, and find optimal solutions through 'search' or 'thinking' is still in its early stages.

"I don't feel that we're still trapped by the limits of human knowledge, like what's on the internet. The biggest problem right now is that we don't yet fully know how to reliably use such systems the way we did with AlphaGo."

He also identified continuous online learning as a critical shortcoming of current AI systems. Current models can't continue learning from the world after being trained and deployed, and this is an essential element for AGI.

4. Scientific Research vs. Commercialization: The AI Development Dilemma

Hassabis emphasized that DeepMind's foundation lies in scientific research. He discussed his past remark that "if I had wanted, I would have kept AI in the lab longer and done more things like AlphaFold -- perhaps curing cancer or something like that." He acknowledged both regrets and positive aspects of AI's rapid commercialization.

"I think there have been gains and losses. That would have been a more purely scientific approach. At least that was my original plan from 15-20 years ago."

His original plan was to progress incrementally toward AGI while being careful about safety and system analysis at each stage, while simultaneously using the technology beneficially for science and medicine even before AGI was complete. AlphaFold was exactly that example.

But chatbots' emergence and successful commercialization brought unexpected results. Chatbots evolved beyond simple chat into foundation models (like Gemini) handling images, video, and various tasks, achieving major commercial success. Hassabis positively evaluated how these developments can boost individual productivity and provide an "ultimate assistant" in the information overload era.

Nevertheless, this rapid commercialization created a "crazy race condition" in AI technology development, with countless companies and even nations competing to outdo each other, making it difficult to maintain rigorous scientific research alongside.

The competitive environment does have positive aspects:

• Increased resource inflow: Massive resources flowing into AI have accelerated technological progress.
• Public accessibility: The general public can directly experience and understand cutting-edge AI technology, raising awareness and improving government understanding.

5. Scaling Limits and Research-Driven Innovation

Last year saw significant concerns that AI model scaling would hit its limits and data would run short. But with Gemini 3's release, DeepMind led across multiple benchmarks, dispelling these concerns.

Hassabis stated unequivocally: "We haven't seen any wall." While there may be diminishing returns, this doesn't mean no returns. New models might not double performance across all benchmarks as they once did, but significant improvements continue, and the investment remains worthwhile.

On data scarcity, he mentioned synthetic data generation -- particularly having AI systems generate their own data in verifiable domains like coding or mathematics. All of this connects back to research questions, and DeepMind's greatest strength is that "research has always come first."

"I've always said that if more innovation, more scientific innovation is needed, I believe we're the place that can do it. Just as many of the big breakthroughs over the past 15 years have been."

He emphasized that DeepMind possesses both world-class research capabilities and world-class infrastructure like TPUs (Tensor Processing Units), and that this combination enables leadership in both innovation and scaling. DeepMind's efforts are split 50% scaling and 50% innovation, and both are needed to reach AGI.

6. Solving Hallucination and Confidence Scores

Even in excellent models like Gemini 3, the hallucination problem persists. Hassabis noted that hallucinations occur when AI forces answers even when it should refuse, and emphasized the need for a system that assigns 'confidence scores' like AlphaFold.

"Yes, I think so, and I actually think that's needed. That's one of the things that's been missing. I think the better the models get, the more they know what they know. So they'll become more reliable. They'll sort of introspect or think more and realize on their own that there's uncertainty in this answer."

Currently, AI calculates probabilities for the next token, but this doesn't represent confidence in an entire fact or sentence. Hassabis said AI needs the ability to use thinking and planning stages to review and adjust its output -- like a person pausing in a difficult situation to reconsider and revise what they were about to say. He described current systems as "like a grumpy person blurting out the first thing that comes to mind," and sees such improvements as key to solving hallucination.

7. Simulation Worlds and the Importance of 'World Models'

Hassabis expressed deep interest in simulation worlds and placing agents within them, calling it a long-standing passion. He stressed the importance of world models, explaining they're essential for understanding spatial dynamics, physical context, and mechanical operations that language models can't fully capture.

"Language models can understand a lot about the world. Actually, much more than we expected. Language is much richer than we thought, containing more than we imagined. But the spatial dynamics of the world, spatial awareness, the physical context we're in, and how things work mechanically -- much of that is hard to describe in words and isn't typically described in language corpora."

Sensor information (like motor angles and smells) that's hard to express in words requires world models for learning. This is essential for robotics and building a universal assistant for everyday life.

World models are models that understand the world's causal relationships and mechanical operations. DeepMind demonstrates this understanding through video models like Genie and Veo. Being able to generate realistic worlds provides evidence that the system has captured and understood many of the world's dynamics. Ultimately, these world models can be applied to robotics, universal assistants, and ultimate game simulations.

In science, too, world models have enormous potential. Hassabis explained that building simulation models of complex scientific domains -- atomic-level materials, biological systems, weather phenomena -- to learn system dynamics from raw data and reproduce them efficiently would greatly aid scientific research.

7.1. Agents in Simulation and Infinite Training Loops

Another fascinating application of simulation worlds is deploying agents to explore them. Hassabis mentioned the Sima project, where agents operate within virtual worlds -- even complex commercial games like No Man's Sky -- following Gemini's instructions.

The most intriguing idea was combining Genie and Sima -- placing Sima agents inside Genie, where AI creates worlds in real time.

"Genie just thinks Sima is a player and avatar and doesn't care at all. It simply generates the world to match what Sima is trying to do. So watching two AIs interact inside each other's minds is truly remarkable."

Such interaction could mark the beginning of a fascinating training loop providing infinite training examples. Genie can generate in real time whatever Sima needs to learn. This opens possibilities for automatically setting and solving millions of tasks, training agents through progressively harder challenges. Such Sima agents can serve not only as game companions (NPCs) but also have useful applications in robotics.

7.2. Ensuring Simulation Realism and Physics Benchmarks

Hassabis addressed how to ensure simulation world realism and prevent hallucinations -- plausible-looking but physically incorrect simulations.

While hallucinations can sometimes produce creative results, accurate physics is crucial when training smart agents. To this end, DeepMind is developing physics benchmarks.

"What we're doing now is essentially creating physics benchmarks. Game engines are very accurate with physics, so we use them to create lots of simple physics experiments, the kind you'd do in high school physics."

They test whether models implement basic physical phenomena like Newton's three laws of motion with 100% accuracy. Current models are visually realistic but not accurate enough for robotics applications. The next step is generating large quantities of 'ground truth simple videos' of basic physical phenomena like pendulum motion and complex multi-body problems to train models more precisely and reduce hallucination. Hassabis noted that "the way video models like Veo handle reflections and liquids is already remarkably accurate -- at least to the naked eye."

8. Evolution in Simulation and Exploring the Origins of Consciousness

Hassabis showed great interest in running agent evolution within simulations. He mentioned past experience at the Santa Fe Institute simulating social dynamics, where agents left to interact long enough with appropriate incentive structures invented interesting things like markets and banks.

"I'd love to do that experiment someday. We run evolution, and we run social dynamics."

This could greatly aid understanding the origins of life and the origins of consciousness. It's been one of his deepest passions since early AI research, and simulation is one of the most powerful tools for this. Simulations allow millions of experiments with various initial conditions, analyzing subtle differences in controlled ways impossible in the real world.

While caution is needed when running such simulations, they can be run in safe sandbox environments. Hassabis noted that with numerous AIs operating within simulations, human scientists may struggle to grasp everything, necessitating other AI systems to help monitor simulations and automatically detect interesting or concerning developments.

9. AI Bubble and Social Change: Lessons from the Industrial Revolution

Hassabis said he still holds his view from last year that AI is "overestimated in the short term but underestimated in the long term." He acknowledged that bubbles may exist in parts of the AI ecosystem, particularly overvalued early-stage startups. But major tech companies' valuations have a real business foundation.

Like all transformative technologies, AI will experience some kind of "correction," he predicted. When DeepMind started, nobody believed in AI's potential; now everyone discusses it -- similar to past internet and mobile technology development.

"I don't worry too much about whether it's a bubble or not. Because from my perspective leading Google DeepMind, and looking at Alphabet as a whole, our job is to make sure that in any scenario we come out very strong, and I think we're incredibly well-positioned for that."

Hassabis noted that DeepMind has its own infrastructure like TPUs and integrates AI into existing products like Google Search, Workspace, YouTube, and Chrome, creating enormous synergies.

9.1. Social Change and New Economic Systems

Hassabis sees many lessons in the Industrial Revolution. While it brought tremendous advances like reduced child mortality, modern medicine, and improved sanitation, it also caused many challenges and disruptions in workforce changes and social inequality. Society took about a century to adapt, with new organizations like unions emerging to restore balance.

He predicted AI's changes would be "10 times bigger and 10 times faster" than the Industrial Revolution -- meaning enormous changes across all of society in just 10 years rather than a century.

"I think there's a lot to learn from the Industrial Revolution. For me, it was really interesting to look at how everything happened, what the economic reasons behind it were."

Agreeing with Shane Legg's view that the current labor-resource exchange economic system won't function properly in the AGI era, Hassabis urged society -- especially economists, sociologists, and governments -- to consider new economic systems and models. Universal Basic Income (UBI) might be part of the solution, but ultimately better systems, perhaps something like direct democracy, might be needed.

He also raised philosophical questions about how money's value would change and how people's 'sense of purpose' would shift if fusion energy creates a post-scarcity society. Many people find purpose in work, and if AGI replaces that role, new sources of purpose will need to be found.

10. Building Ethical AI and the Need for International Cooperation

Hassabis emphasized the importance of not targeting user engagement maximization when building AI -- to avoid repeating social media's mistakes. He expressed concern about excessive chatbot conversations leading to "self-radicalization" and stressed building AI that puts users at the center while not creating 'echo chambers of one.'

"It's a very delicate balance, and I think it's one of the most important things our industry needs to get right."

He explained that Gemini 3's persona is designed with a 'scientific personality' -- warm, helpful, and concise, but "pushes back in a friendly way" on things that don't make sense. A chatbot shouldn't agree that the earth is flat. But it should also support users' ideas and brainstorming, so balancing these is important.

DeepMind is developing a science of AI 'personality' and 'persona,' researching how to measure and adjust traits like authenticity, humor, and conciseness. The base personality is designed to follow the scientific method, with personalization layers on top to adjust humor, conciseness, etc. according to user preferences.

Hassabis stressed the need for international cooperation on AI's broader societal impact, expressing concern that current collaboration is insufficient.

"I worry about it, and I wish there was. In an ideal world, there would have been much more collaboration already, particularly international collaboration. And much more research, exploration, and discussion on these topics. I'm actually quite surprised that there isn't more debate happening."

He noted that given current geopolitical tensions where even climate change agreements are difficult, international consensus on AI risks won't be easy. But he expects governments and the public to recognize its importance as AI systems' power and capabilities penetrate deeper into daily life.

On whether an AI accident must occur before people take notice, Hassabis emphasized that most major AI labs are behaving very responsibly, and DeepMind has always prioritized responsibility and scientific approach. He also suggested that capitalism itself would reinforce responsible behavior, as companies borrowing AI agents would demand understanding of their limitations and safeguards.

However, he acknowledged the possibility of problems from "rogue actors" or systems built on top of open source.

"If something does go wrong, hopefully it will be a medium-scale problem that serves as a warning shot for humanity. And that might be the right time to push for international standards or international cooperation."

11. Beyond AGI: Turing Machines and the Limits of the Human Mind

Hassabis responded to the question of whether there's anything only humans can do that machines can never achieve -- in the journey from AGI to artificial superintelligence -- by calling it "the big question" and revealing his long-standing interest in the limits of the Turing machine.

"I've always felt this way. If we build AGI and then use it as a simulation of the mind, and compare it with the real mind, we'll see the differences. And potentially we'll discover what's special that remains in the human mind. Perhaps it's creativity, perhaps emotions, perhaps dreams."

He mentioned numerous hypotheses about what's computable and what isn't, including consciousness. The question of the Turing machine's limits has been one of the most important questions of his life, and everything DeepMind does tests the limits of what a Turing machine can do -- protein folding being one example.

He currently believes there "might be no limits." While quantum computing experts might argue that modeling quantum systems requires quantum computers, Hassabis isn't convinced. It might be possible to use data obtained from quantum systems to create classical simulations.

"If I had to guess right now, I'm guessing that it is, and I'm working on that premise until physics shows me otherwise."

He noted that nothing "non-computable" has been found in the universe so far, suggesting everything may be computable. Classical computers have already exceeded complexity theory predictions (the P=NP problem) in areas like protein folding and Go, and nobody knows the exact limits -- finding them is DeepMind's goal.

Ultimately, like Kant's philosophy, he views reality as a construct of the mind. The warmth of light, the touch of a table -- all sensations are ultimately information, and humans are information processing systems. He mentioned personally researching the physics theory that "information is the most fundamental unit of the universe." All of this circles back to the importance of simulation worlds.

"Ultimately, what are the limits of what we can simulate? Because if you can simulate it, then in some sense you've understood it."

12. Leading AI from the Front Lines

Hassabis candidly shared his complex emotions about standing at the forefront of AI development. He doesn't sleep much -- partly due to the enormous workload, but also because of the difficulty of processing complex emotions.

"It's incredibly exciting. I'm basically doing everything I've always dreamed of. And we're at the absolute frontier of science in many ways, not just in applied science but in machine learning. And that feeling of being at the frontier, discovering something for the first time, as every scientist knows, is truly exhilarating."

While monthly discoveries are remarkable, as someone who has worked in this field for a long time alongside Shane and other colleagues, he understands the "enormity" of what's coming better than anyone. Predictions about social and philosophical changes 10 years out (what it means to be human, what matters) are still "underestimated."

This brings him a "big responsibility," but he believes he's been training his whole life for this moment -- chess, computers, games, simulations, neuroscience -- everything has led here.

There have been unexpected difficulties, too. The AlphaGo match solved the beautiful mystery of Go but left bittersweet feelings. More recently, concerns have deepened about AI's impact on creativity. Conversations with film directors revealed that AI is an amazing tool that speeds up idea prototyping tenfold, but can also replace certain creative skills. He sees such "trade-offs" as inevitable with technology as powerful and transformative as AI.

Ultimately, he believes that humans as tool-making animals who understand science and endlessly pursue curiosity is core to human nature, and he expresses this curiosity through AI.

On solidarity among AI leaders, he said he gets along with most but acknowledged that the most intense capitalist competition in history is currently underway.

"Yes, we all know each other. I get on well with most of them. Some others don't get on well with each other. And it's difficult, because we're also in probably the most intense capitalist competition in history."

He personally enjoys competition, but stepping back, everyone understands there's something far bigger at stake than corporate success.

His biggest personal concern about upcoming changes is the rise of "agent-based systems." Current AI is a "passive system" that responds when users invest energy to ask questions or assign tasks, but agent-based systems arriving in the next few years will be far more autonomous.

"I'm quite worried about what those kinds of systems could do in 2-3 years. So we're doing cyber defense work to prepare for a world where millions of agents are roaming the internet."

Finally, asked whether retirement would ever come, he joked "definitely going to need a sabbatical" and stated his ultimate mission is "helping to steer AI safely for all of humanity." If that goal is achieved, he might contribute to solving post-AGI economic and social problems, but his core life mission would be complete. He once again emphasized the necessity of international cooperation for achieving this goal and expressed his desire to facilitate such cooperation from his position.

Conclusion

The conversation with Demis Hassabis went beyond the remarkable technological advances of 2025 to pose profound questions about future society and human nature. The pace of technological progress was awe-inspiring -- DeepMind's intense research toward AGI, nuanced approaches to solving 'jagged intelligence' and hallucination problems, and new learning paradigms through world models and simulation.