The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind)
Google DeepMind CEO Demis Hassabis discusses the path to AGI, focusing on the scientific frontiers of the next decade. He covers the importance of solving 'root node' problems like fusion energy, the challenge of 'jagged intelligence' in current models, and the promise of world models and simulations like Genie and SimA. The conversation also explores the balance between scientific rigor and commercial competition, and the profound societal and philosophical questions AGI will force us to confront.
The Next Decade of AI: From Root Problems to AGI
In his annual check-in, Google DeepMind CEO Demis Hassabis reflects on a year that felt like a decade of progress, shifting focus to the fundamental scientific and technological questions that will shape the next phase of AI. The core mission remains tackling “root node” problems—fundamental scientific challenges whose solutions unlock widespread benefits.
Following the success of AlphaFold, the focus is now on other significant challenges:
- Material Science: Pursuing breakthroughs like room temperature superconductors and better batteries.
- Fusion Energy: A deep partnership with Commonwealth Fusion aims to accelerate the development of viable tokamak reactors by using AI to help contain plasma and design materials. Hassabis views fusion as a “holy grail” for clean energy, capable of solving secondary problems like water desalination and creating clean rocket fuel.
- Quantum Computing: Collaborating with Google’s quantum AI team to develop machine learning-based error correction codes.
The State of AI: Jagged Intelligence and Missing Pieces
Hassabis addresses the paradox of current AI systems, which exhibit “jagged intelligence”—achieving PhD-level performance on tasks like International Math Olympiad problems while failing at basic logic or reasoning. He identifies several key components missing for the journey to AGI:
- Consistency: Models must close the gap between their advanced capabilities and their surprising failures on simpler tasks.
- Reasoning and Thinking: Current systems often output the first thing that comes to mind. Hassabis stresses the need for models to incorporate a “thinking” step at inference time to pause, reflect, and double-check their outputs for accuracy and reliability.
- Confidence Scoring: Models need to develop an understanding of what they know and don’t know, similar to AlphaFold’s confidence metrics, to mitigate hallucinations and know when not to answer.
- Continual Learning: A critical missing piece is the ability for models to learn continuously from their interactions in the real world, rather than being static after training.
Scaling, Innovation, and the Commercial Race
Hassabis believes the path to AGI requires both scaling and innovation, stating that Google DeepMind’s effort is split 50/50 between the two. While acknowledging diminishing returns, he asserts that scaling still yields significant improvements and that any perceived “wall” is not a hard stop. He notes that the current challenging landscape favors DeepMind’s research-first culture and its combination of world-class research, engineering, and infrastructure.
He reflects on the intense commercial race, acknowledging that while his original plan was a slower, more deliberate scientific path, the current dynamic has massively accelerated progress by bringing in more resources. He addresses the “AI bubble” by noting that while some areas like startup valuations may be overhyped, the underlying technology is profoundly transformative. Google is well-positioned to weather any market corrections due to its integrated ecosystem.
The Future is Simulated: World Models, Genie, and SimA
A central theme of the conversation is the importance of world models and simulation, a long-standing passion for Hassabis. He argues that language models, while powerful, cannot capture the physical, spatial, and experiential dynamics of the world.
- World Models: These are internal models that learn the causal mechanics and intuitive physics of an environment.
- Genie: A generative interactive model that demonstrates this understanding by creating playable 2D worlds from a single image or text prompt.
- SimA: An agent that can be placed in complex virtual worlds (including commercial games) and be instructed using natural language.
The most exciting development is the synergy between these projects: placing a SimA agent inside a world generated on-the-fly by Genie. This creates a powerful, potentially infinite training loop where an agent can learn by solving an endless stream of automatically generated tasks. A key challenge is ensuring the physics in these simulated worlds are accurate, which is being addressed by creating physics-based benchmarks to test and refine the models.
The Path to AGI and Its Societal Implications
The convergence of projects like Gemini, Genie, and SimA into a single, unified model represents a potential path toward “proto-AGI.” Looking further ahead, Hassabis discusses the profound philosophical questions AGI will force us to confront.
- The Limits of Computation: Hassabis frames his life’s work around the “Turing machine question”: what are the ultimate limits of computation? His working hypothesis is that everything in the universe may be computationally tractable, and he aims to use AI to find that limit. He believes building AGI and comparing it to the human mind will be the ultimate experiment to discover what, if anything, is uniquely human and non-computable.
- Societal Transformation: Drawing parallels to the Industrial Revolution, Hassabis predicts the AI revolution will be “10 times bigger and 10 times faster.” He urges economists, social scientists, and governments to begin designing new economic systems for a post-AGI world, exploring ideas beyond Universal Basic Income (UBI) to address both resource distribution and human purpose.
- Safety and Responsibility: He expresses concern about the lack of international collaboration and institutional preparedness for AGI. The next evolution of AI into more autonomous, agent-based systems within the next few years increases the risks. While hopeful that a major incident can be avoided, he speculates that a “warning shot across the bow” might be necessary to galvanize global cooperation on AI safety standards.