Demis Hassabis: Future of AI, Simulating Reality, Physics and Video Games | Lex Fridman Podcast #475

DeepMind's V3 video generation model can model liquids, materials, and specular lighting surprisingly well. This suggests it has some notion of intuitive physics, similar to how a human child would understand physical interactions, even if it's not a deep philosophical understanding. This capability challenges the notion that physical understanding can only come from embodied AI systems.

Demis Hassabis envisions a future (5-10 years out) where AI systems can create highly personalized, dynamic open-world games. These games would adapt the story and narrative around player choices, acting as the ultimate 'choose your own adventure.' This would move beyond simple random generation or illusions of choice to truly generative game environments.

AlphaEvolve is a Google DeepMind system that evolves algorithms. It combines Large Language Models (LLMs) with evolutionary computing. LLMs propose possible solutions, and then evolutionary computing searches for novel parts of the solution space, aiming to create new emergent capabilities that traditional evolutionary methods struggled with.

Demis Hassabis's dream project, 'Virtual Cell,' aims to model the full internals of a cell, starting with a yeast cell, to perform in silico experiments. This would dramatically speed up research by allowing predictions and searches in a virtual environment before validation in a wet lab. AlphaFold and AlphaFold 3 are foundational steps toward this goal, tackling static protein structures and dynamic interactions respectively.

AI could help simulate the origin of life by performing a search process through a combinatorial chemical space. By starting with initial chemical 'soup' conditions, AI might generate something resembling a cell, unraveling one of biology's greatest mysteries. Hassabis believes this would reveal a continuum between non-living and living matter, connecting physics, chemistry, and biology.

Beyond brute-force testing of thousands of cognitive tasks, AGI will be identified by 'lighthouse moments' like the 'move 37' in Go. This includes inventing a new, profound conjecture in physics (like Einstein's relativity) or inventing a game as deep and elegant as Go. These novel creative leaps, across multiple domains, would signal a true AGI level system.

Demis Hassabis believes there's still significant room for scaling in pre-training, post-training, and inference time. He anticipates continued scaling due to increasing demand for AI systems in products and the emergence of 'thinking systems' that get smarter with more inference time. Google DeepMind focuses on both maximizing current capabilities and pursuing new breakthroughs for harder problems.

AI is expected to materially help with energy problems, including optimizing data center cooling, grid optimization, plasma containment in fusion reactors (as with Commonwealth Fusion), reactor design, and material design for solar panels, superconductors, and batteries. Hassabis believes fusion and highly efficient solar will be the primary clean energy sources in 20-40 years, enabled by AI.

Leading Gemini's success involved consolidating world-class talent from Google Brain and DeepMind, fostering a competitive research culture, and relentlessly shipping progress. A key challenge in a large company like Google is overcoming bureaucracy to allow this rapid research and shipping culture to flourish while maintaining responsibility across numerous product surfaces.

Hassabis, drawing from his game design background, sees AI product design as simplifying complex technology and designing for future capabilities, not just present ones. He anticipates AI-generated, personalized interfaces, moving beyond current text-box chats to multimodal, collaborative interactions, possibly resembling 'Minority Report' or utilizing smart glasses and neural devices for higher bandwidth.

Hassabis views the 'p(doom)' (probability of human civilization destroying itself due to AI) as definitely non-zero and non-negligible, but also 'hugely uncertain.' He finds attempts to quantify it with a precise number 'ridiculous.' Given the high stakes and uncertainty, his approach is 'cautious optimism,' emphasizing scientific research to define and address risks while pursuing AI's transformative benefits for humanity.

Hassabis believes consciousness is likely a computation, an information processing phenomenon, and that brain phenomena are modelable by classical computers, cordially disagreeing with Roger Penrose's quantum consciousness hypothesis. He suggests that understanding consciousness might come through comparing AI with the human mind, and future neural interfaces could even allow humans to 'feel what it's like to compute on silicon.'