🎙️ All-In Podcast: Google DeepMind CEO Demis Hassabis on AI, Creativity, and a Golden Age of Science
PODCAST INFORMATION
All-In Podcast
Google DeepMind CEO Demis Hassabis on AI, Creativity, and a Golden Age of Science
Hosts: Chamath Palihapitiya, Jason Calacanis, David Sacks, David Friedberg
Guest: Demis Hassabis (CEO of Google DeepMind, Nobel Prize winner, knighted by King Charles)
Episode Duration: Approximately 31 minutes
🎧 Listen here.
HOOK
The CEO of Google DeepMind reveals how AI is poised to unlock a new golden era of scientific discovery, fundamentally transforming everything from drug development to creative industries in ways that will make the next decade look like a renaissance of human innovation.
ONE-SENTENCE TAKEAWAY
Demis Hassabis envisions a future where AI will not only transform creative industries and robotics but will fundamentally accelerate scientific discovery to solve humanity's greatest challenges, with AGI achievable within the next decade.
SUMMARY
This episode of the All-In Podcast features a profound conversation with Demis Hassabis, CEO of Google DeepMind and recent Nobel Prize winner, as he shares insights about the current state and future of artificial intelligence. The interview begins with Hassabis recounting his surreal experience winning the Nobel Prize, describing the moment he received the call from Sweden and the tradition of signing the Nobel book alongside legendary scientists like Einstein and Marie Curie.
Hassabis explains DeepMind's role within Alphabet as "the engine room" of AI development, with approximately 5,000 people (mostly engineers and PhD researchers) working on models like Gemini that are integrated across Google products. He emphasizes their mission to build cutting-edge AI systems and immediately deploy them to billions of users through Google's various platforms.
A big portion of the discussion focuses on the Genie World Model, which Hassabis demonstrates as creating interactive 3D environments from text prompts without traditional rendering engines. He explains how this model "reverse engineers intuitive physics" by learning from videos, representing a step toward AI systems that understand the physical world. The demonstration shows users controlling fully generated environments in real-time, with the AI creating consistent worlds as users explore different areas.
The conversation then turns to robotics, with Hassabis discussing the potential for both specialized and humanoid robots. He suggests that while specialized robots will have industrial applications, humanoid form factors may be important for general use since our world is designed for humans. He predicts a "wow moment" in robotics in the coming years but notes that both algorithms and hardware need further development before mass deployment.
Hassabis elaborates on his primary motivation for pursuing AI: accelerating scientific discovery. He highlights AlphaFold's success in protein structure prediction and mentions applications in material design, fusion reactor control, weather prediction, and mathematics. However, he notes that current AI lacks "true creativity" - the ability to make intuitive leaps and develop new hypotheses that characterize great scientific breakthroughs. He suggests this ability to create new conjectures or theories would be a key test for achieving AGI.
Regarding AGI timelines, Hassabis pushes back on claims that AGI is just a few years away, suggesting it's more likely 5-10 years. He identifies missing components including true reasoning capabilities, consistency across domains, and continual learning abilities. He criticizes the notion that current AI systems have "PhD-level intelligence," noting their inconsistencies and limitations.
The discussion covers the democratization of creativity through tools like Nano Banana, which allows users to create sophisticated images without technical expertise. Hassabis envisions a future of entertainment involving co-creation between human visionaries and AI tools, where professional creatives will be supercharged while everyday users gain access to powerful creative capabilities.
Hassabis discusses Isomorphic Labs, his spinout company focused on revolutionizing drug discovery using AI, with the goal of reducing drug discovery timelines from years to weeks or days. He explains the hybrid approach they're taking, combining machine learning with known principles of chemistry and physics.
On energy concerns related to AI, Hassabis acknowledges the issue but notes that model efficiencies are improving dramatically (10-100x over two years). He believes AI will ultimately contribute more to solving energy and climate challenges than it consumes in the coming decade.
The interview concludes with Hassabis's vision for the next 10 years: achieving AGI and ushering in a "new golden era of science" and "renaissance" that will benefit fields from energy to human health. He emphasizes that this potential is what drives his work and the mission at DeepMind.
INSIGHTS
- AI is currently missing "true creativity" - the ability to make intuitive leaps and develop new hypotheses that characterize great scientific breakthroughs; this represents a key threshold for achieving AGI.
- The future of entertainment will likely involve co-creation between human visionaries and AI tools, democratizing creativity while still valuing professional expertise and artistic vision.
- Robotics is approaching a significant breakthrough moment but needs further development in both algorithms and hardware before mass deployment; we may be in an equivalent of the 1970s computing era for robotics.
- AI model efficiencies are improving dramatically (10-100x over two years), which may help address energy concerns even as demand for AI capabilities grows.
- Hybrid AI systems that combine learning components with deterministic rules are currently necessary for scientific applications but may eventually be replaced by end-to-end learning systems as AI advances.
- The distinction between current AI systems and AGI includes consistency across domains, reasoning capabilities, and continual learning abilities that are still missing from today's models.
- AI will ultimately contribute more to solving energy and climate challenges than it consumes in the coming decade, potentially revolutionizing how we approach these global issues.
- The ability of AI systems to understand intuitive physics and the physical world through video training represents a significant step toward more general intelligence.
- Drug discovery could be transformed from a process taking years or decades to one taking weeks or days through AI acceleration, potentially revolutionizing medicine.
- The integration of AI across all Google products demonstrates how rapidly AI is moving from research labs to everyday applications used by billions of people.
FRAMEWORKS & MODELS
Genie World Model
This revolutionary system creates interactive 3D environments from text prompts without traditional rendering engines. The model works by "reverse engineering intuitive physics" through training on millions of videos and synthetic data from game engines. Key components include:
- Text-to-world generation that creates fully interactive environments
- Real-time pixel generation that only creates parts of the world as users explore them
- Consistent world physics that maintains object properties and user modifications
- Natural language integration allowing users to add elements to scenes dynamically
- Potential applications in gaming, simulation, and as a step toward AI systems that understand physical reality
Gemini Models
Google's multimodal AI systems represent DeepMind's approach to building more general intelligence. These models are designed to:
- Accept any type of input (images, audio, video, text)
- Generate any type of output across modalities
- Be integrated across Google's product ecosystem
- Scale efficiently to serve billions of users
- Serve as the foundation for more specialized applications like robotics and scientific research
Hybrid AI Architecture
This approach combines learning components with deterministic rules for scientific applications. The framework includes:
- Neural network components that learn from available data
- Deterministic rules based on known principles of physics, chemistry, or other domains
- Sophisticated methods for integrating these different types of systems
- A strategy to eventually "upstream" deterministic knowledge into the learning components
- Applications in systems like AlphaFold, which combines learning with known chemical constraints
Vision Language Action Models
These systems represent the bridge between AI understanding and physical action. The model framework includes:
- Visual perception of the environment through cameras or other sensors
- Natural language understanding of human instructions
- Translation of abstract instructions into specific motor movements
- Real-world understanding that allows robots to navigate physical spaces safely
- Applications in robotics, smart glasses, and other physical AI systems
Scientific Discovery Acceleration Framework
This represents DeepMind's approach to applying AI to scientific challenges. The framework includes:
- Identification of intractable scientific problems that could benefit from AI acceleration
- Development of specialized AI systems for specific domains (like AlphaFold for protein folding)
- Integration of domain knowledge with machine learning capabilities
- Validation of AI-generated insights through traditional scientific methods
- Extension of successful approaches to adjacent problems (like Isomorphic Labs extending AlphaFold to drug discovery)
QUOTES
"We're the engine room of the whole of Google and the whole of Alphabet. So Gemini, our main model that we're building, but also many of the other models that we also build, the video models and interactive world models, we plug them in all across Google now." - Demis Hassabis, describing DeepMind's central role within Alphabet and how their AI models are integrated across Google's products.
"All of these videos, all these interactive worlds that you're seeing, so you're seeing someone actually can control the video. It's not a static video. It's just being generated by a text prompt. And then people are able to control the 3D environment using the arrow keys and the spacebar. So everything you're seeing here is being fully all these pixels are being generated on the fly." - Demis Hassabis, explaining the revolutionary Genie World Model and how it creates interactive environments without traditional rendering engines.
"AI to accelerate scientific discovery and help with things like human health is the reason I spent my whole career on AI and I think if we build AGI in the right way it will be the ultimate tool for science." - Demis Hassabis, articulating his primary motivation for pursuing AI development and his vision for AGI's ultimate purpose.
"I think we're just scratching the surface of what AI will be able to do and there are some things that are missing. So 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 new hypothesis." - Demis Hassabis, identifying a key limitation of current AI systems and what separates them from true AGI.
"I think that will usher in a new golden era of science. So, a kind of new renaissance. And I think we'll see the benefits of that right across from from energy to to human health." - Demis Hassabis, describing his vision for the impact of achieving AGI within the next decade.
"They're not PhD intelligences. They have some capabilities that are PhD level but they're not in general capable and that's what exactly what general intelligence should be of performing across the board at the PhD level." - Demis Hassabis, pushing back on claims that current AI systems have human-level intelligence and highlighting their inconsistencies.
"I actually foresee a world where there's a bit of co-creation. I still think that you'll have the top creative visionaries. They will be creating these compelling experiences and dynamic story lines and they'll be of higher quality even if they're using the same tools than the everyday person can do." - Demis Hassabis, describing his vision for the future of creative industries and human-AI collaboration.
HABITS
Focus on Practical Applications
Prioritize building AI systems that solve real-world problems, particularly in scientific discovery. This involves identifying intractable challenges across disciplines and developing specialized AI approaches to address them, as demonstrated by AlphaFold's solution to protein folding.
Balance Research and Deployment
Maintain a dual focus on cutting-edge research and immediate practical application. DeepMind's approach involves developing breakthrough AI systems while simultaneously integrating them into products used by billions of people, creating a feedback loop that improves both research and applications.
Interdisciplinary Integration
Combine insights from multiple fields including neuroscience, psychology, computer science, and various scientific domains. This habit enables the development of more sophisticated AI systems by drawing on diverse knowledge bases and approaches.
Long-Term Vision with Incremental Progress
Keep the ultimate goal of AGI in sight while solving intermediate problems. This involves building systems that address specific challenges while contributing to the broader development of more general intelligence capabilities.
Collaboration with Domain Experts
Work closely with specialists in various fields to guide AI development and applications. This includes collaborating with scientists for research applications, artists for creative tools, and industry experts for practical implementations.
Iterative Improvement
Build on previous breakthroughs to tackle increasingly complex problems. This approach is evident in the progression from AlphaGo to AlphaZero to more general systems, each incorporating lessons from the previous iteration.
Efficiency Optimization
Continuously improve model efficiency to address scalability and energy concerns. This involves developing techniques like distillation and other optimization methods that dramatically improve performance while reducing resource requirements.
REFERENCES
Nobel Prize
Awarded to Demis Hassabis for AlphaFold's breakthrough in protein structure prediction, representing one of the highest recognitions in science and highlighting the impact of AI on fundamental research.
AlphaFold
DeepMind's revolutionary system for predicting protein structures with unprecedented accuracy, solving a 50-year grand challenge in biology and opening new possibilities for drug discovery and understanding of biological processes.
Isomorphic Labs
Spinout company founded by Hassabis focused on revolutionizing drug discovery using AI, building on AlphaFold's breakthrough to accelerate the development of new medicines and potentially reduce discovery timelines from years to weeks or days.
Genie World Model
DeepMind's system for creating interactive 3D environments from text prompts without traditional rendering engines, demonstrating AI's ability to learn intuitive physics from video data.
Gemini Models
Google's multimodal AI systems being integrated across products, representing DeepMind's approach to building more general intelligence through systems that can process and generate any type of content.
Nano Banana
Google's image generation tool noted for its consistency and instruction-following capabilities, exemplifying the democratization of creative tools through AI.
VO (Video Generation)
Google's text-to-video generation system that represents advances in AI's ability to create dynamic visual content, being used by professional filmmakers and creators.
AlphaGo and AlphaZero
DeepMind's game-playing systems that defeated world champions and invented new strategies, demonstrating early breakthroughs in reinforcement learning and representing steps toward more general AI capabilities.
Hybrid AI Architecture
The approach of combining learning components with deterministic rules for scientific applications, used in systems like AlphaFold and representing an important step toward more capable AI systems.
Crepi il lupo! 🐺