The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)

Scaling and Innovation in AGI 00:05

"My bet is you're going to need both to get to AGI."

• Demis Hassabis highlights that the pursuit of Artificial General Intelligence (AGI) requires a balanced focus on both scaling current technologies and fostering innovation. This dual approach is essential for advancing towards achieving AGI.

Understanding Human Uniqueness Through AGI 00:22

"If we build AGI and use that as a simulation of the mind, we will then see what the differences are and potentially what's special about the human mind."

• Hassabis expresses the value of developing AGI as a means to simulate human cognition. By comparing AGI with the actual workings of the human mind, researchers can identify distinctive attributes of human intelligence, such as creativity and emotions.

The Challenge of Turing Machines and Computability 00:35

"Nobody's found anything in the universe that's non-computable, so far."

• The discussion includes insights into the limits of Turing machines and the broader implications for what can be computed. Hassabis emphasizes that current understanding supports the notion that there are no known constraints that prevent computation from addressing complex problems.

Major Developments in AI Technology 01:00

"It feels like we packed in 10 years in one year."

• The past year has seen unprecedented developments in AI technology, with advancements in models such as Gemini 3, which showcases improved multi-modal capabilities. These changes signify rapid progress in the AI landscape.

Key Focus Areas for Future Research 02:41

"The big proof point was AlphaFold."

• AlphaFold marked a significant milestone in AI's ability to address complex biological problems, such as protein folding. This achievement has encouraged further exploration into additional root node problems, including material science and fusion energy.

The Potential of Fusion Energy 03:58

"If we could have modular fusion reactors, this promise of unlimited, renewable, clean energy would obviously transform everything."

• Fusion energy is portrayed as a transformative solution for global energy needs. The development of modular fusion reactors offers the potential for clean, abundant energy, which could have far-reaching positive impacts on climate change and global accessibility to resources.

Addressing Inconsistencies in AI Performance 05:29

"There's something missing still from these systems in terms of their consistency."

• Hassabis points out the paradox of AI systems excelling at high-level tasks, like winning medals at math competitions, while still struggling with basic logic problems. This inconsistency highlights the need for further advancements to achieve true AGI.

The Future of AI with Improved Reasoning 07:21

"We're still in the early days of having this search or thinking on top."

• AI models currently benefit from vast amounts of human knowledge but lack the refined reasoning abilities that would make them fully reliable. Enhancing these capabilities is crucial for the evolution of AI systems towards a level that resembles human reasoning and decision-making.

The Next Step in AI Development: Online Learning 09:06

"One of the things missing from today's systems is the ability to online learn and continually learn."

• Demis Hassabis emphasizes the need for AI systems to go beyond initial training and incorporate online learning capabilities. Unlike humans, current AI systems do not continue to learn once deployed; they rely on pre-trained knowledge. This inability limits the potential for Artificial General Intelligence (AGI), which requires systems that can adapt and learn continuously in real-time.

The Scientific Approach vs. Commercialization 09:40

"If I’d had my way, we would have left AI in the lab for longer and done more things like AlphaFold, maybe cured cancer or something like that."

• Reflecting on the balance between scientific research and commercial pressure, Hassabis notes that rushing to deploy AI technology has both advantages and disadvantages. While there are significant advancements in commercial products like chatbots, he believes a slower, more cautious approach could have led to foundational breakthroughs in science and medicine, paralleling the success of AlphaFold.

Scaling and Diminishing Returns 13:06

"I think we've never really seen any wall, as such."

• Addressing concerns about potential limits in AI development, Hassabis argues that while some believe the field is hitting a wall, there is still significant potential for improvement. He acknowledges the concept of diminishing returns, suggesting that while performance gains may become less drastic, meaningful advancements are still occurring, as evidenced by the capabilities of models like Gemini 3.

Addressing Hallucinations in AI Responses 15:33

"I think we need a confidence score in the same way that AlphaFold does."

• The issue of "hallucinations," where AI provides incorrect but confident responses, is highlighted as a challenge. Hassabis proposes that future AI systems could benefit from incorporating a confidence score that reflects the reliability of their outputs. This feature would help the models better recognize uncertainty and adjust their responses accordingly.

The Importance of World Models 17:48

"World models and simulations have been my longest-standing passion."

• Demis Hassabis emphasizes the significance of world models, which manage to encapsulate the causative mechanics of the world, unlike language models that focus primarily on linguistic understanding.

• While language models have shown an unexpected richness in understanding world dynamics through words, they still fall short in capturing spatial awareness and physical interactions that are often challenging to articulate.

• Hassabis points out that many experiences and concepts, particularly sensory perceptions like smell or motor skills, cannot simply be described in language; they require direct experience.

• To develop robotics or a universal assistant that seamlessly integrates into daily life, a profound understanding of the world is essential, highlighting the core need for effective world models.

Realistic Simulations and Their Potential 20:14

"If you can generate a realistic world, it shows that your system understands the mechanics of the world."

• The ability to generate realistic environments is a crucial indicator of a model's understanding of world dynamics. Hassabis mentions ongoing projects like Genie and Veo that illustrate impressive advancements in generating complex worlds.

• These models are not only promising for applications in gaming but also hold potential for advancing robotics and assisting scientific exploration by simulating systems efficiently.

• Hassabis argues that there is a vast opportunity for world models to be utilized in scientific fields, such as simulating complex biological or physical systems using learned dynamics from raw data.

Simulated Agents and Their Applications 21:22

"We can drop agents into simulated environments and allow them to explore with curiosity being their main motivator."

• The introduction of agents into simulated worlds presents an exciting avenue for exploration and innovation. Projects like SIMA, featuring agents that can interact within virtual environments, enable agents to carry out tasks and learn dynamically.

• Hassabis envisions a future where agents can navigate and learn from environments created in real-time by other AI systems, leading to a cycle of training and adaptation that could yield endless learning opportunities.

• This interactive setup could significantly enhance the versatility of simulated agents, making them useful not only in gaming contexts but also for practical applications in robotics.

Testing and Ensuring Realism in Simulations 23:26

"Hallucinations may be interesting, but we want to eliminate inaccuracies in physics."

• Realism in simulations is pivotal, and there is a focus on minimizing inaccuracies that could arise from models hallucinating incorrect physics. Hassabis highlights the challenges in ensuring models capture the laws of physics accurately.

• Current models are approximations and, although they may visually appear realistic to the untrained eye, they still lack the precision necessary for reliable applications, particularly in robotics.

• Hassabis advocates for creating a physics benchmark through the use of accurate game engines that can simulate fundamental physical experiments, thus allowing for rigorous testing of the models' understanding of real-world physics.

Understanding Life and Consciousness Through AI Simulations 26:46

"Accurate simulations will be an unbelievable boon to science."

• Demis Hassabis emphasizes the necessity of AI tools to grasp the origins of life and consciousness.

• He argues that simulations can provide insights that are statistically significant by allowing researchers to run numerous controlled experiments with varying initial conditions.

• This approach can be powerful for addressing complex questions that are challenging to analyze in the real world.

The Importance of Safety in Simulations 27:33

"You can run simulations in pretty safe sandboxes."

• Hassabis acknowledges the need for caution when conducting simulations, particularly due to the emergent properties of models that can yield unexpected results.

• He points out that simulations can be monitored continuously, allowing scientists to scrutinize all activities and data throughout the experimentation phase.

• AI tools are seen as essential for managing the complexity of these simulations and identifying any significant anomalies or concerns automatically.

The Impact of AI Bubbles on the Ecosystem 28:50

"There are parts of the AI ecosystem that are probably in bubbles."

• Hassabis discusses the ongoing debate about the existence of an AI bubble, noting that although certain aspects may reflect unsustainable valuations, other parts of the AI industry have solid foundations.

• He stresses that new transformative technologies often experience overreactions and overcorrections within market sentiments, similar to the internet and mobile technology.

• Regardless of the market dynamics, he believes that DeepMind's positioning is strong, enabling continuous advancements toward artificial general intelligence (AGI).

Building Responsible AI Systems 31:59

"We've seen what happens with some systems that were overly sycophantic."

• Hassabis reflects on the significance of developing AI that enhances user engagement without creating harmful echo chambers.

• He advocates for a balanced approach where AI systems support users while gently challenging incorrect beliefs, thereby promoting critical thinking.

• Gemini, the AI model being developed, is designed to maintain a scientific approach and users' personal preferences while adhering to a core authentic personality.

Advancements Toward AGI 34:42

"I think that's very exciting."

• Hassabis highlights the recent launch of the Nano Banana Pro system as a significant step toward AGI, enabling comprehensive understanding and generation related to images.

• He notes the capability of this system to semantically comprehend images, indicating profound advancements in image processing and rendering.

• The progression in world models like Genie and SIMA further illustrates the ongoing research and development efforts aimed at achieving a general-purpose AI system.

Convergence Towards Proto-AGI 36:04

"We need to converge all of those different projects into one big model."

• Demis Hassabis emphasizes the need for integrating various AI projects into a unified model that could represent a candidate for proto-AGI. He acknowledges that while current endeavors are interconnected, they ultimately require convergence to realize advanced intelligence capabilities.

Lessons from the Industrial Revolution 36:19

"There's a lot we can learn from the Industrial Revolution to mitigate disruptions expected as AGI comes."

• Hassabis believes that studying the Industrial Revolution can provide insights into managing societal changes induced by AGI. He notes that the Industrial Revolution originated from economic motives, specifically the textile industry, which transitioned into advancements such as sewing machines and mainframe computers.

• He highlights how it created significant societal shifts, resulting in improvements like reduced child mortality and modern medicine, but also introduced challenges requiring adaptations, such as the formation of labor unions.

Economic Reconfiguration Post-AGI 38:43

"I think society, in general, needs to spend more time thinking about how to restructure in a post-AGI world."

• The discussion addresses the necessity for rethinking economic systems as AGI arrives, with new models required to support a transformed workforce. Hassabis hints at the potential introduction of universal basic income and direct democracy-like systems where community voting could dictate local resource allocations.

• He identifies the need for proactive solutions to ensure equitable distribution of benefits amid profound societal changes due to AGI.

Institutional and Collaborative Challenges 41:28

"I am worried about the lack of international collaboration on addressing AGI-related challenges."

• Hassabis expresses concern about insufficient collaboration and preparation among global institutions as AGI development accelerates. He notes that the fragmented nature of current institutions does not provide the necessary influence or structure required to address upcoming challenges effectively.

• The geopolitical tensions complicate efforts for cooperation, especially evident in areas like climate change where agreements have been difficult to reach.

Anticipating Disruptions and Ethical Considerations 44:13

"It may be that institutions aren't currently adequate to deal with the complexities of AGI."

• Hassabis acknowledges the possibility of unforeseen incidents arising from AGI advancement. He maintains that most labs strive for responsible development while recognizing that rogue actors could pose risks.

• He considers the implications of a society with abundant resources, suggesting that the transition to a post-scarcity world raises philosophical questions about purpose and fulfillment, given that many derive their sense of purpose from work and family responsibilities.

The Uniqueness of the Human Mind and Its Limits 45:12

"Maybe it's creativity, maybe it's emotions, or consciousness."

• Demis Hassabis discusses what may be special about the human mind, proposing that aspects like creativity, emotions, and consciousness could set humans apart from machines.

• He raises the question of what can be computed and references the limitations of a Turing machine, which has been a central focus of his life and work.

• Hassabis believes that exploring the boundaries of what a Turing machine can do has been essential to their work at DeepMind, including advancements in complex tasks like protein folding.

Computational Potential and Quantum Considerations 46:12

"I'm not sure what the limit is; maybe there isn't one."

• Hassabis expresses uncertainty about the limits of computation, suggesting that we might need quantum computing to unlock new possibilities.

• He introduces the idea that if consciousness involves quantum effects, classical machines may never replicate it; however, if everything is computationally tractable, Turing machines could potentially model everything in the universe.

Exploring the Nature of Reality 48:10

"In the end, it's all information, and we're information-processing systems."

• He posits that our experience of reality could ultimately be distilled into information-processing, aligning with his view that biology operates as an information system.

• Hassabis believes that if information is the most fundamental unit of the universe, then sensory experiences are essentially interpretations of this information.

The Role of Simulation in Understanding Limits 48:56

"If you can simulate it, then, in some sense, you've understood it."

• Hassabis asserts that simulations are crucial for exploring the limits of computation and understanding various phenomena.

• He highlights that their work aims to discover the boundaries of what can be achieved through simulation, which plays a significant role in gaining insights into complex systems.

Personal Reflections on the Emotional Weight of Progress 49:13

"It's unbelievably exciting, and I'm doing everything I ever dreamed of."

• Hassabis shares the emotional complexity of being at the forefront of AI and scientific discovery, acknowledging both the exhilaration and the weight of responsibility associated with it.

• He reflects on the excitement of making groundbreaking discoveries regularly, while also feeling the enormity of their implications for humanity.

The Intersection of AI and Creativity 51:04

"On the one hand, they've got these amazing tools that speed up prototyping ideas, but on the other hand, is it replacing certain creative skills?"

• He discusses the dual nature of AI's impact on creative fields, where advancements in technology can enhance creativity but may also displace certain skills.

• Hassabis respects the creative arts and recognizes the trade-offs that come with integrating powerful technology into these domains.

Anticipating the Future of AI Systems 53:22

"Right now, the systems are passive; in the next couple of years, we'll start seeing some really impressive, reliable ones."

• He anticipates a shift from current passive AI systems to more autonomous, agent-based ones, which could provide significant capabilities as assistants.

• However, he expresses concern over the risks associated with these capable systems and highlights their team's focus on cyber defense in preparation for potential challenges posed by advanced AI.

The Mission to Steward AGI Safely 54:34

"My mission has always been to help the world steward AGI safely over the line for all of humanity."

• Demis Hassabis expresses a strong commitment to the safe development of Artificial General Intelligence (AGI), emphasizing the importance of collaboration in this critical endeavor.

• He acknowledges the significant work that remains, which not only involves reaching AGI but also addressing subsequent challenges related to superintelligence and its economic implications.

• Hassabis indicates that there is a broader societal context that he hopes to contribute to once AGI is achieved, viewing this as a key part of his life's mission.

Future Collaboration and Personal Break 55:03

"I think it's going to require collaboration, and I'm quite a collaborative person."

• In discussing the future, Hassabis highlights the necessity of collaboration in advancing AGI safely, indicating his willingness to work with others in the field.

• He humorously alludes to his need for a break after this monumental task, suggesting that once his core mission is complete, he looks forward to a well-earned sabbatical.

• This desire for a personal respite underlines the intensity of his work and the dedication he has towards his mission, reinforcing the balance between ambition and self-care.

Podcast Conclusion and Future Engagement 55:31

"Be sure to subscribe so you will be among the first to hear about our return in 2026."

• Professor Hannah Fry wraps up the podcast series, inviting listeners to subscribe for updates on the future of the program, signaling an anticipation for continued discussions around advanced topics like AGI.

• She encourages the audience to explore the extensive library of past episodes, showcasing the breadth of topics covered, highlighting advancements in technology from driverless cars to drug discovery.

• This emphasis on revisiting previous discussions underlines the importance of continuous learning and engagement in the rapidly evolving field of AI and technology.