Lex Fridman Podcast

Demis Hassabis on Lex Fridman: How AlphaFold Changed Biology

Hassabis conjectures any pattern shaped by nature is learnable by classical systems; alphafold solved protein folding while veo models fluid dynamics passively.

LFLex FridmanhostDHDemis Hassabisguest

Jul 22, 2025·2h 28m·Watch on YouTube ↗

🙋 FREQUENTLY ASKED QUESTIONS

Direct answers grounded in the episode transcript. Tap any timestamp to verify against the source.

1. 1

   What is Veo 3 learning about intuitive physics?

   Veo 3 suggests a video model can learn enough world dynamics to make short clips coherent. Hassabis describes its understanding as limited but real in a practical sense: if it can predict next frames coherently, that is a form of understanding. It is not a deep philosophical or anthropomorphic understanding, and he explicitly says he does not think the system has that. What impresses him is that it can model physics behavior, lighting, materials, and liquids well enough to produce roughly eight seconds of consistent video that is hard to spot as wrong at a glance. He compares this to a child's intuitive physics rather than a PhD-level grasp of equations. Later he adds that passive observation may be enough to learn parts of physical reality, challenging the idea that robotics or embodied action is required.

   ▸ 15:25 in transcript
2. 2

   What is AlphaEvolve and how does it use LLMs?

   AlphaEvolve is presented as an example of hybrid AI: foundation models propose possibilities, then search methods explore the space. Hassabis says LLMs can suggest possible solutions, while evolutionary computing searches for novel regions of the search space. He treats this as a promising direction because the foundation model supplies a learned basis, and other computational methods can be layered on top. Evolutionary methods are one option, but he also names Monte Carlo tree search and other reasoning or search algorithms. The larger point is creativity: once a model captures what is already known, a search process can move beyond known examples. He connects that to scientific discovery and medicine, where an objective function can guide the system toward useful new regions rather than searching randomly through an impossibly large space.

   ▸ 31:28 in transcript
3. 3

   What is Google DeepMind's Virtual Cell project?

   Virtual Cell is Hassabis's long-running dream of modeling the inside of a living cell well enough to run useful experiments in silico. He says the challenge is breaking a grand goal into interim steps that are useful on their own, with AlphaFold as one of the components. The dream is to do most of the search on a virtual cell, then use the wet lab mainly for validation, potentially speeding experiments by 100X. He would probably start with a yeast cell because it is a single-cell organism, effectively a full organism in one cell, and it is well understood. The path starts with AlphaFold's static protein structures, moves through AlphaFold3's modeling of protein, RNA, and DNA interactions, then toward whole pathways such as the TOR pathway, and eventually a whole-cell model.

   ▸ 42:10 in transcript
4. 4

   What is Demis Hassabis's 2030 AGI timeline?

   Demis Hassabis gives AGI a roughly 50 percent chance within five years, meaning by 2030. His bar for AGI is high: it must match the brain's cognitive functions, because human minds are broadly general and created modern civilization. He says today's systems are jagged, strong in some areas but flawed in others, so a true AGI would need consistent intelligence across the board. It would also need missing capabilities like genuine invention and creativity. For testing, he suggests tens of thousands of cognitive tasks that humans can do, plus giving a few hundred top experts a month or two to look for obvious flaws. He also looks for lighthouse moments: inventing a new physics hypothesis like Einstein did, or creating a game as deep, elegant, and beautiful as Go.

   ▸ 52:33 in transcript
5. 5

   What is Demis Hassabis's p(doom)?

   Hassabis refuses to give a precise p(doom) number because he thinks the number would imply precision that does not exist. He says he does not know how people are arriving at their p(doom) estimates, and calls the notion a little ridiculous. His actual position is that the risk is definitely non-zero and probably non-negligible, but hugely uncertain. The uncertainty comes from not knowing what the technologies will do, how quickly they will take off, how controllable they will be, and whether the hardest problems are easier or harder than expected. He frames the situation as high stakes on both sides: AI could help solve disease, energy, scarcity, and enable human flourishing, while p(doom) scenarios remain real. His preferred stance is cautious optimism plus much more research to define and address the risks.

   ▸ 1:58:14 in transcript

Answers are AI-generated from the transcript and may contain errors. Tap a question to verify against the source.