Person
1 episodes · 25 quoted claims
Demis Hassabis says proteins fold in milliseconds in the body, a problem DeepMind's AlphaFold has now solved computationally.
“You know, proteins fold in milliseconds in our bodies, so somehow physics solves this problem that we've now also solved computationally.”
Hassabis says Google's Veo video model appears to learn how materials behave just from watching videos, whereas he found hand-writing physics engines painstakingly hard.
“yet somehow these systems are, you know, reverse engineering from just watching YouTube videos. So presumably what's happening is it's extracting some underlying structure around how these materials behave.”
Hassabis describes Veo 3 as generating about eight seconds of consistent video that is hard to fault at a glance.
“Whatever it is, eight seconds of consistent video that, by eye, at least, you know, at a glance, is quite hard to distinguish what the issues are.”
Hassabis worked on the game Theme Park, an open-world simulation where each player's experience was unique.
“So games like Theme Park that I worked on, where everybody's game experience would be unique to them, right?”
Hassabis says he worked on the early stages of the game Black & White, which contained what he considers still probably the best learning AI in a game.
“Black and White, uh, was the game that I worked on. Early stages of that that had the still probably the best AI learning AI in it.”
Hassabis grew up with a Spectrum and then a Commodore Amiga 500, which he calls his favourite computer and where he learned programming.
“I had a Spectrum and then a a Commodore, Amiga 500, which was my my favorite computer ever. And that's why I learned all my programming.”
Hassabis describes AlphaEvolve as combining large language models that propose solutions with evolutionary computing that searches for novel ones.
“So LLMs are kind of proposing some possible solutions, and then you do. You use evolutionary computing on top to to to find some novel part of the of the search space.”
Hassabis says AlphaGo used Monte Carlo tree search, which found the novel strategy known as Move 37 in Go.
“AlphaGo, we just used Monte Carlo tree search, right? And that's what found Move 37, the new, uh, kind of never seen before strategy in Go.”
Hassabis says DeepMind's AlphaProof system earned a silver-medal level result on maths Olympiad problems the previous year.
“Alpha proof last year, our system got, you know, silver medal in that.”
Hassabis says biologist Paul Nurse, a mentor of his, founded the Crick Institute and won the Nobel Prize in 2001.
“He runs the the, you know, he founded the Crick Institute and and won the Nobel Prize in in 2001.”
Hassabis says he has wanted to build a virtual simulation of a cell for about 25 years.
“I've had this idea, you know, of wanting to do that for maybe more like 25 years.”
Hassabis says he would start a virtual cell project with yeast, a well-understood single-celled organism that Paul Nurse studied.
“And I'd probably start with a yeast cell, and partly that's what Paul Nurse studied, because a yeast cell is like a full organism, that's a single cell, right?”
Hassabis says Google DeepMind's neural-network weather systems outperform traditional fluid-dynamics weather models that run on supercomputers.
“We've created the the best weather prediction systems in the world, and they're better than traditional fluid dynamics sort of systems”
Hassabis estimates roughly a 50% chance of AGI within five years, by about 2030.
“My estimate is sort of 50 chance by in the next five years, so, you know, by 2030, let's say.”
Hassabis proposes testing an AI by giving it all knowledge up to 1900 and seeing whether it could derive special and general relativity like Einstein.
“have a cutoff of knowledge cutoff of 1900, and then give the system everything that was, you know, that was written up to 1900.”
Hassabis says Google DeepMind has done work with Commonwealth Fusion on AI for plasma containment in fusion reactors.
“We've done lots of work on that with Commonwealth Fusion and also, uh, one could imagine reactor design”
Hassabis argues that cheap clean energy would solve water access through desalination, which he says currently is too expensive for most countries.
“the water access problem goes away because you can just use desalination. We have the technology, it's just too expensive.”
Hassabis says a new major Gemini version takes roughly six months for a full training run and productization.
“maybe it takes, you know, roughly six months or something to to do a new kind of full run and the full productization of a new version”
Hassabis says the smaller Gemini models, Flash and Flashlight, are often distilled from the largest Pro model.
“we also have different sizes pro and flash and flashlight that are often distilled from the biggest ones, you know, the flash model from the pro model.”
Hassabis says Isomorphic Labs was spun out of the AlphaFold work to do drug discovery.
“we spun out companies like Isomorphic off the back of AlphaFold to do drug discovery, and it's going really well.”
Hassabis says that when DeepMind started in 2010 he did not pay himself for a couple of years because they could not raise money.
“I remember when we were starting out back in 2010, you know, I didn't even pay myself for a couple of years”
Hassabis predicts AI's impact will be around ten times that of the industrial revolution and arrive roughly ten times faster.
“I think what we're gonna see is something like probably 10 times the impact the industrial revolution had, and but 10 times faster as well, right?”
Hassabis did a neuroscience PhD studying how the brain works, focusing on imagination and memory in the hippocampus.
“to look at how the brain works, especially imagination and memory. I focused on the hippocampus”
Hassabis says Roger Penrose worked with neuroscientists to seek quantum mechanical effects in the brain, which Hassabis says have not been convincingly found.
“He collaborated with a lot of good neuroscientists to see if he could find mechanisms for quantum mechanics behavior in the brain.”
Hassabis says Google built a version of its Gemma model, called DolphinGemma, trained on dolphin and whale sounds.
“A version of our system was trained on dolphin and whale sounds.”