The episode features Demis Hassabis, leader of Google DeepMind and Nobel Prize winner, in his second appearance on the podcast. Demis is recognized as one of the most brilliant minds working on understanding and building intelligence while exploring fundamental mysteries of the universe.

The conversation begins with Demis' Nobel Prize lecture conjecture that any natural pattern can be efficiently modeled by classical learning algorithms, spanning biology, chemistry, physics, cosmology, and neuroscience. This provocative claim stems from AlphaGo and AlphaFold's success in modeling high-dimensional combinatorial spaces.

Discussion explores whether classical computers can model chaotic systems, fluid dynamics, and even consciousness itself, with Demis proposing a new complexity class for learnable natural systems. The conversation touches on P versus NP, the nature of reality as informational, and whether AGI breakthroughs will require new architectures beyond current scaling.

Later topics include video game design, Google's competitive position in AI, the future of human-AI interaction, concerns about job displacement, and the philosophical implications of consciousness and substrate. Demis shares his vision of post-AGI society with energy abundance and discusses the responsibilities of stewarding transformative technology.

Natural Patterns and Classical Computation

Demis' Nobel Prize lecture proposed the conjecture: "Any pattern that can be generated or found in nature can be efficiently discovered and modeled by a classical learning algorithm"

AlphaGo and AlphaFold demonstrate building models of combinatorially high-dimensional spaces that would be intractable through brute force enumeration - "there wouldn't be enough time in the time of the universe" - Demis

Natural systems have structure because they were subject to evolutionary processes that shaped them, making patterns learnable by neural networks through what Demis calls "survival of the stablest"

"If you think about geological times, the shape of mountains being shaped by weathering processes over thousands of years, the orbits of planets, shapes of asteroids - these have all survived processes that acted on them many times" - Demis

Manmade or abstract problems like factorizing large numbers may require quantum computers if there's no pattern to learn, but most natural systems of interest have evolved structure

P Equals NP and Complexity Theory

Demis is working in spare time with colleagues on defining a new complexity class for problems solvable by neural network processes mapped onto natural systems with structure

"I think information is primary. Information is the most fundamental unit of the universe, more fundamental than energy and matter. I think of the universe as a kind of informational system" - Demis

If the universe is informational, then P equals NP becomes a physics question that could help solve fundamental questions about reality

Current AI systems prove classical Turing machines can go far beyond previous expectations - "a lot of people would've thought maybe 10, 20 years ago that was decades away, or maybe you would need quantum machines" - Demis

"We haven't really even scratched the surface yet of what classical systems could do. AGI being built on a neural network system on top of a classical computer would be the ultimate expression of that" - Demis

Veo 3 and Physics Understanding

Veo video generation model can model liquids, materials, and specular lighting surprisingly well by extracting underlying structure from YouTube videos

"I used to write physics engines and graphics engines in my early days in gaming, and I know it's just so painstakingly hard to build programs that can do that" - Demis on fluid dynamics in Veo

Fluid dynamics and Navier-Stokes equations traditionally thought intractable on classical systems, yet Veo models them through passive observation without embodied interaction

"Perhaps there is some kind of lower dimensional manifold that can be learned if we actually fully understood what's going on under the hood. That's maybe true of most of reality" - Demis

Veo 3 challenges the assumption that embodied AI systems are required for understanding physical world - "it seems like you can understand it through passive observation, which is pretty surprising" - Demis

Video Games and Open World Design

Demis envisions AI-generated open world games where systems dynamically create content and narrative around player choices, beyond simple A-B branching or random generation

"Games like Theme Park that I worked on where everybody's game experience would be unique to them because you are kind of co-creating the game" - Demis on early open world design

With vibe coding improvements, Demis hopes to personally create games in spare time, alongside working on physics theory as post-AGI projects

"In the '90s and early 2000s, we weren't just making games, we felt we were creating a new entertainment medium where you as the player were co-creating the story" - Demis

Video games represent fusion of cutting-edge technology with artistic design - "all of the most interesting technical advances were happening in gaming" in the 1990s, including AI, graphics, physics engines, and GPUs

AlphaEvolve and Evolutionary Search

AlphaEvolve combines LLMs proposing solutions with evolutionary computing to find novel regions of search space, representing promising hybrid approach

"You've got the model of the underlying dynamics of the system, and then if you want to discover something new, you need some kind of search process on top" - Demis

Traditional evolutionary computing in the 1990s-2000s could never evolve new emergent properties, only subsets of input properties, but foundation models may overcome this limitation

"Naturally evolution clearly did evolve new capabilities from bacteria to where we are now. So clearly it must be possible with evolutionary systems to generate new patterns and new emergent properties" - Demis

Evolution as search process running over physics substrate for four billion years of computational time generated incredible rich diversity

Virtual Cell and Scientific Discovery

Demis has dreamed of modeling a complete cell for 25 years, discussing with Nobel laureate Paul Nurse since the 1990s about requirements for useful in silico experiments

"Maybe you could 100X speed up experiments by doing most of it in silico, the search in silico, and then you do the validation step in the wet lab. That's the dream" - Demis

AlphaFold provides static 3D protein structures; AlphaFold 3 models pairwise interactions; next steps include modeling whole pathways like TOR pathway involved in cancer, then eventually full yeast cell

Yeast cell chosen as target because it's simplest single-cell organism that's also a full organism, and very well understood scientifically

Modeling different temporal scales in cells requires hierarchical systems or multiple simulated systems that can interact at different temporal dynamics

Goal is to model at protein level rather than atomic level to capture relevant dynamics without over-modeling quantum mechanical aspects

Research Taste and Scientific Creativity

"Taste or judgment is what separates the great scientists from the good scientists. All professional scientists are good technically, but do you have the taste to sniff out what the right direction is" - Demis

"It's harder to come up with a conjecture, a really good conjecture than it is to solve it. We may have systems soon that can solve pretty hard conjectures, but could a system come up with a conjecture worthy of study" - Demis

True creativity requires leaps of imagination like Einstein had with special and general relativity - current systems cannot make such conceptual breakthroughs

Good conjectures split the hypothesis space in two, where learning something useful whether true or false, and are falsifiable with available technologies

"There's no such thing as failure really as long as you are picking experiments and hypotheses that meaningfully split the hypothesis space" - Demis on research methodology

AGI Timeline and Definition

Demis estimates 50% chance of AGI by 2030, defining it as matching all cognitive functions the brain has with consistency across domains, not jagged intelligence

Testing AGI would involve tens of thousands of cognitive tasks humans can do, plus giving world's top experts in each field a month or two to find obvious flaws

"Lighthouse moments" for AGI would include inventing new conjectures like Einstein did with special relativity, or inventing a game as deep and elegant as Go

Back-test proposal: give system knowledge cutoff of 1900 and see if it could come up with special and general relativity like Einstein did

Current systems have missing capabilities in true invention and creativity - "they're good if you give them very specific instructions, but if you give them a very vague high-level instruction that wouldn't work currently" - Demis

Scaling Laws and Research Breakthroughs

Scaling continues across three dimensions: pre-training, post-training, and inference time compute, with room for growth in all areas

"It's kind of 50/50 whether new things are needed or whether the scaling of the existing stuff is gonna be enough. We are pushing both of those as hard as possible" - Demis

About half of Google DeepMind resources dedicated to new blue sky ideas, half to scaling current capabilities to maximum

"I would back us to be the place that does that breakthrough if some new breakthrough is required. I actually quite like it when the terrain gets harder because then it veers more from just engineering to true research" - Demis

"Maybe 80-90% of the breakthroughs that underpins modern AI field today was from originally Google Brain, Google Research, and DeepMind" - Demis on research track record

Energy, Fusion, and Type I Civilization

Demis bets on fusion and solar as primary future energy sources, with solar being "the fusion reactor in the sky" requiring better batteries and transmission

AI systems could help with plasma containment for fusion reactors, reactor design, new solar materials, room temperature superconductors, and optimal batteries within next five years

"If energy is kind of free and renewable and clean, then that solves a whole bunch of other problems. Water access problem goes away because you can just use desalination" - Demis

Free energy enables unlimited rocket fuel by separating seawater into hydrogen and oxygen, combined with self-landing rockets could create "bus service to space"

"For the first time in human history, we wouldn't be resource constrained. It's not zero sum. This radical abundance era where there's plenty of resources to go around" - Demis

Demis would not be surprised if humanity reaches Type I Kardashev scale civilization within 100 years given energy breakthroughs

Google DeepMind's Competitive Position

Google went from "losing" to winning in LLM products within a year through incredible team led by Koray, Jeff Dean, and Oriol, plus research culture combining Google Brain and DeepMind

"Relentless progress along with relentless shipping of that progress" has been key to success in unbelievably competitive space

Google DeepMind operates as startup within large company, fighting to cut away bureaucracy while maintaining responsible practices for billions of users

"There's very few places in the world you can do incredible world-class research on one hand and then plug it in and improve billions of people's lives the next day" - Demis

Demis brings product design skills from 1990s game development, combining cutting-edge technology with user experience - "I love the combination of cutting-edge research and then being applied in a product"

Programming Jobs and AI Displacement

Programming and math turned out easier for AI than expected because synthetic data can be created and verified, contrary to earlier assumptions about "harder skills"

"People who embrace these technologies become almost at one with them will become sort of superhumanly productive. The great programmers will be even better, 10X even what they are today" - Demis

Top programmers will have advantages in specifying architecture, guiding coding assistants, and checking code quality for foreseeable future

"Anytime where there's a lot of disruption and change, we've had this many times in human history with the internet, mobile, industrial revolution. It's gonna be one of those eras" - Demis

Impact will be "probably 10 times the impact the industrial revolution had but 10 times faster as well. Instead of a hundred years, it takes 10 years" - Demis

Demis encourages top economists and philosophers to think about societal effects and solutions like universal basic provision where increased productivity gets distributed

Consciousness and Substrate Dependence

Demis disagrees with Roger Penrose on quantum consciousness, noting no convincing mechanisms found for quantum behavior in brain despite good neuroscientist collaborations

"My betting is there is that it's mostly just classical computing going on in the brain, which suggests that all the phenomena are modelable or mimicable by a classical computer" - Demis

"One of the best definitions I like of consciousness is it's the way information feels when we process it" - Demis

We assume others are conscious for two reasons: same behavior and same substrate. With AI on silicon, we can't rely on substrate similarity even if behavior matches

Neural interfaces like Neuralink might eventually let us feel what it's like to compute on silicon, potentially bridging the consciousness gap between substrates

"We've never had to confront that before" - the challenge of radical empathy with different substrate, similar to empathizing with plants or other life forms

P Doom and Existential Risk

"I don't have a p doom number. The reason I don't is because I think it would imply a level of precision that is not there" - Demis on probability of doom

"It's definitely non-zero and it's probably non-negligible. So that in itself is pretty sobering. My view is it's just hugely uncertain" - Demis

Given uncertainty but huge stakes both ways, "the only rational, sensible approach is to proceed with cautious optimism" - Demis

Risks operate over different timescales: bad actors using technology for harmful ends (near-term) versus autonomous AGI control problems (longer-term), both equally important

"How does one restrict bad actors access to these powerful systems but enable access at the same time to good actors to maximally build on top. It's a pretty tricky problem" - Demis

AI could help with early warning on bad actor use cases in bio or nuclear domains, but only if the AI itself is reliable - "it's a sort of interlocking problem"

Von Neumann, Manhattan Project, and Collaboration

John von Neumann foresaw that computers would be more impactful than nuclear weapons, predicting "10 times at least of the industrial revolution" impact - Demis agrees

Von Neumann talked about learning machines as "grown rather than programmed" in the 1950s, presaging modern AI approaches

The maniac by Benjamin Labatut explores the double-edged sword of discovery through von Neumann's life and work

"We need to approach it with whatever you wanna call it, a spiritual dimension or humanist dimension. This idea of a soul, what makes us human, the spark that we have" - Demis

Demis hopes for CERN-like collaborative model for final AGI steps rather than Manhattan Project-style weapons race between states

"At least on the scientific level, it's important for the researchers to keep in touch and keep close to each other" despite difficult geopolitical climate - Demis