How AI Is Unlocking the Secrets of Nature and the Universe | Demis Hassabis | TED

Summary

In a fascinating conversation, Demis Hassabis, the CEO of DeepMind, discusses the transformative potential of artificial intelligence (AI). He shares his journey from being a child prodigy in chess to leading a team that developed AlphaGo, the AI that mastered the complex game of Go. Hassabis highlights how AI can unlock scientific breakthroughs, exemplified by AlphaFold, a DeepMind program that predicts protein structures with astonishing accuracy, accelerating drug discovery and disease understanding. He emphasizes the importance of responsible development as AI becomes more powerful, advocating for collaboration to ensure safe and beneficial AI systems. Hassabis envisions a future where AI contributes to a radical abundance of knowledge, potentially leading to the cure for all diseases and a deeper understanding of the universe.

Takeaways

• 🧠 Demis Hassabis founded DeepMind with the vision of using AI to answer fundamental questions about the nature of reality and consciousness.
• 🎲 Games, particularly chess, played a significant role in sparking Hassabis' interest in AI and its potential to mimic human thought processes.
• 🤖 DeepMind's early breakthroughs involved training AI to play video games, which led to the development of 'deep reinforcement learning' techniques.
• 🔍 AI's ability to find patterns and insights in vast amounts of data can complement the scientific method and potentially lead to significant scientific breakthroughs.
• 🌟 DeepMind's AlphaGo program marked a pinnacle in games-playing AI, as it learned to play Go better than any human by inventing new strategies.
• 🚀 AlphaZero expanded upon AlphaGo's capabilities, starting from zero knowledge and rapidly becoming proficient in any two-player game.
• 🧬 The protein-folding problem, a grand challenge in biology, was tackled by DeepMind's AlphaFold, which can predict a protein's 3D structure from its amino acid sequence.
• ⏱️ AlphaFold's success in predicting protein structures could significantly accelerate drug discovery and disease understanding, potentially reducing the process from years to months.
• 🤝 Open-sourcing AlphaFold's database of 200 million proteins signifies a commitment to collaborative scientific advancement and the potential for wide-reaching impact.
• 🌐 The competition between tech giants to develop advanced AI models and supercomputers raises concerns about a potential 'Moloch Trap' scenario, where competition drives riskier actions.
• ⚖️ As we approach AGI (Artificial General Intelligence), there's a call for increased collaboration and thoughtful development to ensure safe and beneficial AI architectures.

Q & A

Why did Demis Hassabis believe that building AI could be the fastest route to answer big questions in philosophy and physics?

Demis Hassabis thought that building AI could help answer big questions because he observed that in the past 20 to 30 years, not much progress had been made in understanding fundamental laws of physics. He believed AI could serve as the ultimate tool to assist in this endeavor and also aid in better understanding ourselves and the brain.

What role does AI play in scientific breakthroughs according to Demis Hassabis?

AI can process vast amounts of data, finding patterns and insights that are beyond human comprehension. It surfaces these findings to human scientists, who can then develop new hypotheses and conjectures, thus complementing the scientific method.

How did Demis Hassabis' early interest in games contribute to his journey in AI?

Demis' interest in games, particularly chess, led him to early chess computers in the mid-'80s. He was fascinated by the fact that a machine could be programmed to play chess at a high level, sparking his curiosity about how the brain creates thought processes and how they could be mimicked by computers.

What was the significance of DeepMind's work with Atari games?

DeepMind's work with Atari games marked the first time the AI system surprised its creators. The AI learned a strategy for the game Breakout that humans had not considered, which was a significant moment that demonstrated the potential of AI to innovate and learn from raw data.

How did AlphaGo's victory over the world champion at Go demonstrate a new level of AI capability?

AlphaGo's victory was significant because it not only beat the world champion at Go, a game more complex than chess, but it also developed and employed new strategies that had never been seen before, showcasing the AI's ability to innovate and understand complex patterns.

What is the concept of 'deep reinforcement learning' that was pivotal in DeepMind's AI development?

Deep reinforcement learning is a technique where AI systems learn directly from raw pixels on the screen without any prior instructions or context. They are given a goal, such as maximizing the score, and must make sense of the visual data and devise strategies to achieve the goal on their own.

How did AlphaZero differ from AlphaGo in its approach to learning games?

Unlike AlphaGo, which was trained on human games played on the internet, AlphaZero started with zero prior knowledge and learned entirely from random play. It was designed to be more general, capable of mastering any two-player game, not just Go.

What was the motivation behind open-sourcing AlphaFold's database of 200 million protein structures?

Demis Hassabis and his team open-sourced AlphaFold's database to accelerate scientific discovery globally. They believed that by sharing their findings, they could maximize the potential impact on biology, drug design, and disease understanding.

How does the new company Isomorphic plan to extend the work done with AlphaFold?

Isomorphic aims to extend the work with AlphaFold into the chemistry space, designing chemical compounds that can bind precisely to specific spots on proteins, potentially revolutionizing drug discovery and reducing the time required from years to months.

What is the 'Moloch Trap' and how does it relate to the competitive landscape of AI development?

The 'Moloch Trap' is a situation where companies in a competitive environment may be driven to actions that no individual within those companies would take independently. In the context of AI, it refers to the rush to release AI products and models, potentially without fully understanding the implications or ensuring safety, driven by the competitive pressure to not fall behind.

What is Demis Hassabis' vision for the future role of AI in scientific discovery?

Demis envisions AI as a tool that could potentially allow scientists to explore the entire 'tree of knowledge.' He believes AI can help solve 'root node problems' that, once solved, unlock new branches of discovery, leading to an era of radical abundance, curing diseases, and expanding human consciousness.

Outlines

00:00

🤖 AI as a Tool for Understanding the Universe

In this segment, Demis Hassabis, founder of DeepMind, discusses his motivation for pursuing artificial intelligence (AI) as a means to answer profound questions about reality and consciousness. He highlights the limitations he observed in physics and how AI could serve as an ultimate tool for scientific discovery, pattern recognition in vast data sets, and potentially unlocking new insights in both game play and scientific research.

05:02

🎲 The Evolution of AI in Gaming and Beyond

This paragraph details the progression of AI in gaming, starting with simple Atari games and moving towards complex games like Go. The discussion emphasizes the development of 'deep reinforcement learning' and how AI, specifically AlphaGo and later AlphaZero, learned to play games from scratch, achieving superhuman performance and even devising novel strategies that surprised human experts.

10:02

🧬 Solving the Protein-Folding Problem with AI

Demis explains how AI, through the program AlphaFold, has revolutionized the field of biology by predicting protein structures from their amino acid sequences. This was a significant challenge that, once solved, could greatly accelerate scientific discovery and has the potential to unlock treatments for diseases. The decision to open-source AlphaFold's database of 200 million protein structures is highlighted as a major contribution to the scientific community.

15:04

💊 Isomorphic: Extending AI into Drug Discovery

The conversation shifts towards the next steps in applying AI to scientific problems, specifically in the domain of drug discovery. Isomorphic, a new venture, aims to use AI to design chemical compounds that can bind to specific protein targets without side effects, potentially reducing the drug discovery process from years to months.

20:05

🌐 The Competitive Landscape of AI Development

The discussion addresses the competitive nature of AI development, referencing the 'Moloch Trap' and the rapid release of AI products in response to competitive pressures. The narrative touches on the need for thoughtful and responsible development of AI, especially as we approach the development of artificial general intelligence (AGI), and the importance of collaboration across different sectors of society.

25:07

🚀 AGI and the Future of Scientific Discovery

Demis concludes with his vision for the future of AI, particularly once AGI is achieved. He sees AI as a means to explore the entirety of human knowledge, which he likens to a tree of knowledge. He hopes to use AGI to conduct experiments at the Planck scale, probing the fundamental nature of reality, and believes that the scientific method is the best approach to understanding the universe.

Keywords

💡Artificial Intelligence (AI)

Artificial Intelligence refers to the simulation of human intelligence in machines that are programmed to think like humans and mimic their actions. In the video, AI is central to the theme as it is seen as a tool to answer big questions in science and philosophy, and to make scientific breakthroughs. An example from the script: 'So I thought, why not build the ultimate tool to help us, which is artificial intelligence.'

💡DeepMind

DeepMind is a UK-based artificial intelligence company founded by Demis Hassabis. It is known for creating the Go-playing AI AlphaGo and for its work in advancing AI research. In the context of the video, DeepMind is the company driving the development and application of AI to solve complex problems like protein folding. Example: 'And when you launched DeepMind, pretty quickly, you started having it tackle game play.'

💡Protein Folding

Protein folding is a scientific challenge that involves predicting the three-dimensional structure of a protein from its amino acid sequence. It is crucial for understanding a protein's function in the body. In the video, Demis Hassabis discusses how his AI program, AlphaFold, can predict protein structures, which is a significant step towards scientific discoveries in medicine. Example: 'And it's kind of like a 50-year grand challenge in biology.'

💡AlphaGo

AlphaGo is a computer Go program developed by DeepMind that can beat world champions at the game Go, which is more complex than chess. AlphaGo is significant because it demonstrated that AI could learn and strategize at a level beyond human understanding. Example from the script: 'And then we reached the pinnacle, which was the game of Go, which is what they play in Asia instead of chess, but it's actually more complex than chess.'

💡AlphaZero

AlphaZero is an AI developed by DeepMind that can learn to play and master various games, starting from random play, without any prior knowledge. It represents a leap in AI's ability to learn and adapt. In the video, AlphaZero is mentioned as an evolution of AlphaGo, capable of mastering chess within hours. Example: 'So this is what AlphaZero was. That's why it's the zero in the name, because it started with zero prior knowledge.'

💡AGI (Artificial General Intelligence)

AGI, or Artificial General Intelligence, refers to highly autonomous systems that can outperform humans at most economically valuable work. It is a term used to describe the next frontier in AI development. In the video, the discussion around AGI focuses on the need for safe and responsible development as we approach creating machines with general intelligence. Example: 'as we get closer to AGI, we need to collaborate more.'

💡Reinforcement Learning

Reinforcement learning is a type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize a reward. It is key to how AI systems like AlphaGo and AlphaZero learn to play games. In the video, it is mentioned in the context of how AI systems learn strategies from raw pixels on the screen. Example: 'So this was in 2012, 2013, where we coined these terms "deep reinforcement learning."'

💡Large Language Models (LLMs)

Large Language Models are AI systems that process and generate language at scale, often used for tasks like text generation, translation, and understanding. They are a recent focus in AI development. In the video, the discussion around LLMs relates to their rapid advancement and the public's readiness to use these systems. Example: 'And in fact, obviously, Google research actually invented Transformers, which was the architecture that allowed all this to be possible, five, six years ago.'

💡The Moloch Trap

The Moloch Trap is a concept where competitive dynamics within an industry can lead companies to act against their own long-term interests. It is mentioned in the video to describe the situation where companies feel compelled to rush AI products to market due to competitive pressure. Example: 'And it's felt, as a sort of layperson observer, that the Moloch Trap has been shockingly in effect in the last couple of years.'

💡Isomorphic

Isomorphic, in the context of the video, refers to a new company spun out of Google that aims to extend the work done in AlphaFold into the chemistry space, potentially leading to breakthroughs in drug discovery. It represents the application of AI to new scientific domains. Example: 'And Isomorphic is extending this work we did in AlphaFold, into the chemistry space, where we can design chemical compounds, that will bind exactly to the right spot on the protein.'

💡Fundamental Nature of Reality

The fundamental nature of reality refers to the underlying principles that govern the universe at its most basic level. In the video, Demis Hassabis expresses a long-term vision for AI to help understand these principles, potentially enabling experiments at the Planck scale. Example: 'So do experiments at the Planck scale. You know, the smallest possible scale, theoretical scale, which is almost like the resolution of reality.'

Highlights

Demis Hassabis believes that building AI is the fastest route to answer big questions in philosophy and physics.

AI can help find patterns and insights in vast amounts of data, complementing the scientific method.

Hassabis's interest in games led him to AI, and DeepMind's work started with simple Atari games.

DeepMind's AI developed innovative strategies in games like Breakout, surprising its creators.

AlphaGo's victory over a world champion in Go demonstrated AI's ability to learn and create strategies.

AlphaZero, an AI developed by DeepMind, taught itself to play chess at a world-champion level in just hours.

DeepMind's AlphaFold program can predict protein structures from their amino acid sequences with high accuracy.

AlphaFold has saved an estimated billion years of PhD time by predicting 200 million protein structures in a year.

DeepMind open-sourced AlphaFold, making its results accessible to the scientific community for further research.

Isomorphic, a new company from DeepMind, aims to extend AI's capabilities into chemistry for drug discovery.

Hassabis envisions using AI to explore the entire 'tree of knowledge' and unlock new branches of discovery.

AI's potential to accelerate scientific discovery could lead to a future of radical abundance and maximum human flourishing.

Hassabis discusses the importance of collaboration as AI approaches general intelligence to ensure safe and beneficial development.

The competitive landscape in AI development, like the 'Moloch Trap', can lead to companies making decisions they might not otherwise.

Large investments in supercomputers by tech companies are driving AI capabilities but also raise concerns about a potential race.

Hassabis is optimistic about AI's future and its role in understanding the fundamental nature of reality.

The scientific method is considered the best technique for understanding the universe, despite potential unknowable aspects.

Transcripts

Chris Anderson: Demis, so good to have you here.

Demis Hassabis: It's fantastic to be here, thanks, Chris.

Now, you told Time Magazine,

"I want to understand the big questions,

the really big ones that you normally go into philosophy or physics

if you're interested in them.

I thought building AI

would be the fastest route to answer some of those questions."

Why did you think that?

DH: (Laughs)

Well, I guess when I was a kid,

my favorite subject was physics,

and I was interested in all the big questions,

fundamental nature of reality,

what is consciousness,

you know, all the big ones.

And usually you go into physics, if you're interested in that.

But I read a lot of the great physicists,

some of my all-time scientific heroes like Feynman and so on.

And I realized, in the last, sort of 20, 30 years,

we haven't made much progress

in understanding some of these fundamental laws.

So I thought, why not build the ultimate tool to help us,

which is artificial intelligence.

And at the same time,

we could also maybe better understand ourselves

and the brain better, by doing that too.

So not only was it incredible tool,

it was also useful for some of the big questions itself.

CA: Super interesting.

So obviously AI can do so many things,

but I think for this conversation,

I'd love to focus in on this theme of what it might do

to unlock the really big questions, the giant scientific breakthroughs,

because it's been such a theme driving you and your company.

DH: So I mean, one of the big things AI can do,

and I've always thought about,

is we're getting, you know, even back 20, 30 years ago,

the beginning of the internet era and computer era,

the amount of data that was being produced

and also scientific data,

just too much for the human mind to comprehend in many cases.

And I think one of the uses of AI is to find patterns and insights

in huge amounts of data and then surface that

to the human scientists to make sense of

and make new hypotheses and conjectures.

So it seems to me very compatible with the scientific method.

CA: Right.

But game play has played a huge role in your own journey

in figuring this thing out.

Who is this young lad on the left there?

Who is that?

DH: So that was me, I think I must have been about around nine years old.

I'm captaining the England Under 11 team,

and we're playing in a Four Nations tournament,

that's why we're all in red.

I think we're playing France, Scotland and Wales, I think it was.

CA: That is so weird, because that happened to me too.

In my dreams.

(Laughter)

And it wasn't just chess,

you loved all kinds of games.

DH: I loved all kinds of games, yeah.

CA: And when you launched DeepMind,

pretty quickly, you started having it tackle game play.

Why?

DH: Well, look, I mean, games actually got me into AI in the first place

because while we were doing things like,

we used to go on training camps with the England team and so on.

And actually back then,

I guess it was in the mid '80s,

we would use the very early chess computers,

if you remember them, to train against,

as well as playing against each other.

And they were big lumps of plastic,

you know, physical boards that you used to,

some of you remember, used to actually press the squares down

and there were LED lights, came on.

And I remember actually, not just thinking about the chess,

I was actually just fascinated by the fact that this lump of plastic,

someone had programmed it to be smart

and actually play chess to a really high standard.

And I was just amazed by that.

And that got me thinking about thinking.

And how does the brain come up with these thought processes,

these ideas,

and then maybe how we could mimic that with computers.

So yeah, it's been a whole theme for my whole life, really.

CA: But you raised all this money to launch DeepMind,

and pretty soon you were using it to do, for example, this.

I mean, this is an odd use of it.

What was going on here?

DH: Well, we started off with games at the beginning of DeepMind.

This was back in 2010, so this is from about 10 years ago,

it was our first big breakthrough.

Because we started off with classic Atari games from the 1970s,

the simplest kind of computer games there are out there.

And one of the reasons we used games is they're very convenient

to test out your ideas and your algorithms.

They're really fast to test.

And also, as your systems get more powerful,

you can choose harder and harder games.

And this was actually the first time ever that our machine surprised us,

the first of many times,

which, it figured out in this game called Breakout,

that you could send the ball round the back of the wall,

and actually, it would be much safer way to knock out all the tiles of the wall.

It's a classic Atari game there.

And that was our first real aha moment.

CA: So this thing was not programmed to have any strategy.

It was just told, try and figure out a way of winning.

You just move the bat at the bottom and see if you can find a way of winning.

DH: It was a real revolution at the time.

So this was in 2012, 2013

where we coined these terms "deep reinforcement learning."

And the key thing about them is that those systems were learning

directly from the pixels, the raw pixels on the screen,

but they weren't being told anything else.

So they were being told, maximize the score,

here are the pixels on the screen,

30,000 pixels.

The system has to make sense on its own from first principles

what’s going on, what it’s controlling,

how to get points.

And that's the other nice thing about using games to begin with.

They have clear objectives, to win, to get scores.

So you can kind of measure very easily that your systems are improving.

CA: But there was a direct line from that to this moment

a few years later,

where country of South Korea and many other parts of Asia

and in fact the world went crazy over -- over what?

DH: Yeah, so this was the pinnacle of -- this is in 2016 --

the pinnacle of our games-playing work,

where, so we'd done Atari,

we'd done some more complicated games.

And then we reached the pinnacle, which was the game of Go,

which is what they play in Asia instead of chess,

but it's actually more complex than chess.

And the actual brute force algorithms

that were used to kind of crack chess were not possible with Go

because it's a much more pattern-based game,

much more intuitive game.

So even though Deep Blue beat Garry Kasparov in the '90s,

it took another 20 years for our program, AlphaGo,

to beat the world champion at Go.

And we always thought,

myself and the people working on this project for many years,

if you could build a system that could beat the world champion at Go,

it would have had to have done something very interesting.

And in this case, what we did with AlphaGo,

is it basically learned for itself,

by playing millions and millions of games against itself,

ideas about Go, the right strategies.

And in fact invented its own new strategies

that the Go world had never seen before,

even though we've played Go for more than,

you know, 2,000 years,

it's the oldest board game in existence.

So, you know, it was pretty astounding.

Not only did it win the match,

it also came up with brand new strategies.

CA: And you continued this with a new strategy

of not even really teaching it anything about Go,

but just setting up systems

that just from first principles would play

so that they could teach themselves from scratch, Go or chess.

Talk about AlphaZero and the amazing thing that happened in chess then.

DH: So following this, we started with AlphaGo

by giving it all of the human games that are being played on the internet.

So it started that as a basic starting point for its knowledge.

And then we wanted to see what would happen if we started from scratch,

from literally random play.

So this is what AlphaZero was.

That's why it's the zero in the name,

because it started with zero prior knowledge

And the reason we did that is because then we would build a system

that was more general.

So AlphaGo could only play Go,

but AlphaZero could play any two-player game,

and it did it by playing initially randomly

and then slowly, incrementally improving.

Well, not very slowly, actually, within the course of 24 hours,

going from random to better than world-champion level.

CA: And so this is so amazing to me.

So I'm more familiar with chess than with Go.

And for decades,

thousands and thousands of AI experts worked on building

incredible chess computers.

Eventually, they got better than humans.

You had a moment a few years ago,

where in nine hours,

AlphaZero taught itself to play chess better than any of those systems ever did.

Talk about that.

DH: It was a pretty incredible moment, actually.

So we set it going on chess.

And as you said, there's this rich history of chess and AI

where there are these expert systems that have been programmed

with these chess ideas, chess algorithms.

And you have this amazing, you know,

I remember this day very clearly, where you sort of sit down with the system

starting off random, you know, in the morning,

you go for a cup of coffee, you come back.

I can still just about beat it by lunchtime, maybe just about.

And then you let it go for another four hours.

And by dinner,

it's the greatest chess-playing entity that's ever existed.

And, you know, it's quite amazing,

like, looking at that live on something that you know well,

you know, like chess, and you're expert in

and actually just seeing that in front of your eyes.

And then you extrapolate to what it could then do in science or something else,

which of course, games were only a means to an end.

They were never the end in themselves.

They were just the training ground for our ideas

and to make quick progress in a matter of, you know,

less than five years actually went from Atari to Go.

CA: I mean, this is why people are in awe of AI

and also kind of terrified by it.

I mean, it's not just incremental improvement.

The fact that in a few hours you can achieve

what millions of humans over centuries have not been able to achieve.

That gives you pause for thought.

DH: It does, I mean, it's a hugely powerful technology.

It's going to be incredibly transformative.

And we have to be very thoughtful about how we use that capability.

CA: So talk about this use of it because this is again,

this is another extension of the work you've done,

where now you're turning it to something incredibly useful for the world.

What are all the letters on the left, and what’s on the right?

DH: This was always my aim with AI from a kid,

which is to use it to accelerate scientific discovery.

And actually, ever since doing my undergrad at Cambridge,

I had this problem in mind one day for AI,

it's called the protein-folding problem.

And it's kind of like a 50-year grand challenge in biology.

And it's very simple to explain.

Proteins are essential to life.

They're the building blocks of life.

Everything in your body depends on proteins.

A protein is sort of described by its amino acid sequence,

which you can think of as roughly the genetic sequence

describing the protein, so that are the letters.

CA: And each of those letters represents in itself a complex molecule?

DH: That's right, each of those letters is an amino acid.

And you can think of them as a kind of string of beads

there at the bottom, left, right?

But in nature, in your body or in an animal,

this string, a sequence,

turns into this beautiful shape on the right.

That's the protein.

Those letters describe that shape.

And that's what it looks like in nature.

And the important thing about that 3D structure is

the 3D structure of the protein goes a long way to telling you

what its function is in the body, what it does.

And so the protein-folding problem is:

Can you directly predict the 3D structure just from the amino acid sequence?

So literally if you give the machine, the AI system,

the letters on the left,

can it produce the 3D structure on the right?

And that's what AlphaFold does, our program does.

CA: It's not calculating it from the letters,

it's looking at patterns of other folded proteins that are known about

and somehow learning from those patterns

that this may be the way to do this?

DH: So when we started this project, actually straight after AlphaGo,

I thought we were ready.

Once we'd cracked Go,

I felt we were finally ready after, you know,

almost 20 years of working on this stuff

to actually tackle some scientific problems,

including protein folding.

And what we start with is painstakingly,

over the last 40-plus years,

experimental biologists have pieced together

around 150,000 protein structures

using very complicated, you know, X-ray crystallography techniques

and other complicated experimental techniques.

And the rule of thumb is

that it takes one PhD student their whole PhD,

so four or five years, to uncover one structure.

But there are 200 million proteins known to nature.

So you could just, you know, take forever to do that.

And so we managed to actually fold, using AlphaFold, in one year,

all those 200 million proteins known to science.

So that's a billion years of PhD time saved.

(Applause)

CA: So it's amazing to me just how reliably it works.

I mean, this shows, you know,

here's the model and you do the experiment.

And sure enough, the protein turns out the same way.

Times 200 million.

DH: And the more deeply you go into proteins,

you just start appreciating how exquisite they are.

I mean, look at how beautiful these proteins are.

And each of these things do a special function in nature.

And they're almost like works of art.

And it's still astounds me today that AlphaFold can predict,

the green is the ground truth, and the blue is the prediction,

how well it can predict, is to within the width of an atom on average,

is how accurate the prediction is,

which is what is needed for biologists to use it,

and for drug design and for disease understanding,

which is what AlphaFold unlocks.

CA: You made a surprising decision,

which was to give away the actual results of your 200 million proteins.

DH: We open-sourced AlphaFold and gave everything away

on a huge database with our wonderful colleagues,

the European Bioinformatics Institute.

(Applause)

CA: I mean, you're part of Google.

Was there a phone call saying, "Uh, Demis, what did you just do?"

DH: You know, I'm lucky we have very supportive,

Google's really supportive of science

and understand the benefits this can bring to the world.

And, you know, the argument here

was that we could only ever have even scratched the surface

of the potential of what we could do with this.

This, you know, maybe like a millionth

of what the scientific community is doing with it.

There's over a million and a half biologists around the world

have used AlphaFold and its predictions.

We think that's almost every biologist in the world

is making use of this now, every pharma company.

So we'll never know probably what the full impact of it all is.

CA: But you're continuing this work in a new company

that's spinning out of Google called Isomorph.

DH: Isomorphic.

CA: Isomorphic.

Give us just a sense of the vision there.

What's the vision?

DH: AlphaFold is a sort of fundamental biology tool.

Like, what are these 3D structures,

and then what might they do in nature?

And then if you, you know,

the reason I thought about this and was so excited about this,

is that this is the beginnings of understanding disease

and also maybe helpful for designing drugs.

So if you know the shape of the protein,

and then you can kind of figure out

which part of the surface of the protein

you're going to target with your drug compound.

And Isomorphic is extending this work we did in AlphaFold

into the chemistry space,

where we can design chemical compounds

that will bind exactly to the right spot on the protein

and also, importantly, to nothing else in the body.

So it doesn't have any side effects and it's not toxic and so on.

And we're building many other AI models,

sort of sister models to AlphaFold

to help predict,

make predictions in chemistry space.

CA: So we can expect to see

some pretty dramatic health medicine breakthroughs

in the coming few years.

DH: I think we'll be able to get down drug discovery

from years to maybe months.

CA: OK. Demis, I'd like to change direction a bit.

Our mutual friend, Liv Boeree, gave a talk last year at TEDAI

that she called the “Moloch Trap.”

The Moloch Trap is a situation

where organizations,

companies in a competitive situation can be driven to do things

that no individual running those companies would by themselves do.

I was really struck by this talk,

and it's felt, as a sort of layperson observer,

that the Moloch Trap has been shockingly in effect in the last couple of years.

So here you are with DeepMind,

sort of pursuing these amazing medical breakthroughs

and scientific breakthroughs,

and then suddenly, kind of out of left field,

OpenAI with Microsoft releases ChatGPT.

And the world goes crazy and suddenly goes, “Holy crap, AI is ...”

you know, everyone can use it.

And there’s a sort of, it felt like the Moloch Trap in action.

I think Microsoft CEO Satya Nadella actually said,

"Google is the 800-pound gorilla in the search space.

We wanted to make Google dance."

How ...?

And it did, Google did dance.

There was a dramatic response.

Your role was changed,

you took over the whole Google AI effort.

Products were rushed out.

You know, Gemini, some part amazing, part embarrassing.

I’m not going to ask you about Gemini because you’ve addressed it elsewhere.

But it feels like this was the Moloch Trap happening,

that you and others were pushed to do stuff

that you wouldn't have done without this sort of catalyzing competitive thing.

Meta did something similar as well.

They rushed out an open-source version of AI,

which is arguably a reckless act in itself.

This seems terrifying to me.

Is it terrifying?

DH: Look, it's a complicated topic, of course.

And, first of all, I mean, there are many things to say about it.

First of all, we were working on many large language models.

And in fact, obviously, Google research actually invented Transformers,

as you know,

which was the architecture that allowed all this to be possible,

five, six years ago.

And so we had many large models internally.

The thing was, I think what the ChatGPT moment did that changed was,

and fair play to them to do that, was they demonstrated,

I think somewhat surprisingly to themselves as well,

that the public were ready to,

you know, the general public were ready to embrace these systems

and actually find value in these systems.

Impressive though they are, I guess, when we're working on these systems,

mostly you're focusing on the flaws and the things they don't do

and hallucinations and things you're all familiar with now.

We're thinking, you know,

would anyone really find that useful given that it does this and that?

And we would want them to improve those things first,

before putting them out.

But interestingly, it turned out that even with those flaws,

many tens of millions of people still find them very useful.

And so that was an interesting update on maybe the convergence of products

and the science that actually,

all of these amazing things we've been doing in the lab, so to speak,

are actually ready for prime time for general use,

beyond the rarefied world of science.

And I think that's pretty exciting in many ways.

CA: So at the moment, we've got this exciting array of products

which we're all enjoying.

And, you know, all this generative AI stuff is amazing.

But let's roll the clock forward a bit.

Microsoft and OpenAI are reported to be building

or investing like 100 billion dollars

into an absolute monster database supercomputer

that can offer compute at orders of magnitude

more than anything we have today.

It takes like five gigawatts of energy to drive this, it's estimated.

That's the energy of New York City to drive a data center.

So we're pumping all this energy into this giant, vast brain.

Google, I presume is going to match this type of investment, right?

DH: Well, I mean, we don't talk about our specific numbers,

but you know, I think we're investing more than that over time.

So, and that's one of the reasons

we teamed up with Google back in 2014,

is kind of we knew that in order to get to AGI,

we would need a lot of compute.

And that's what's transpired.

And Google, you know, had and still has the most computers.

CA: So Earth is building these giant computers

that are going to basically, these giant brains,

that are going to power so much of the future economy.

And it's all by companies that are in competition with each other.

How will we avoid the situation where someone is getting a lead,

someone else has got 100 billion dollars invested in their thing.

Isn't someone going to go, "Wait a sec.

If we used reinforcement learning here

to maybe have the AI tweak its own code

and rewrite itself and make it so [powerful],

we might be able to catch up in nine hours over the weekend

with what they're doing.

Roll the dice, dammit, we have no choice.

Otherwise we're going to lose a fortune for our shareholders."

How are we going to avoid that?

DH: Yeah, well, we must avoid that, of course, clearly.

And my view is that as we get closer to AGI,

we need to collaborate more.

And the good news is that most of the scientists involved in these labs

know each other very well.

And we talk to each other a lot at conferences and other things.

And this technology is still relatively nascent.

So probably it's OK what's happening at the moment.

But as we get closer to AGI, I think as a society,

we need to start thinking about the types of architectures that get built.

So I'm very optimistic, of course,

that's why I spent my whole life working on AI and working towards AGI.

But I suspect there are many ways to build the architecture safely, robustly,

reliably and in an understandable way.

And I think there are almost certainly going to be ways of building architectures

that are unsafe or risky in some form.

So I see a sort of,

a kind of bottleneck that we have to get humanity through,

which is building safe architectures as the first types of AGI systems.

And then after that, we can have a sort of,

a flourishing of many different types of systems

that are perhaps sharded off those safe architectures

that ideally have some mathematical guarantees

or at least some practical guarantees around what they do.

CA: Do governments have an essential role here

to define what a level playing field looks like

and what is absolutely taboo?

DH: Yeah, I think it's not just about --

actually I think government and civil society

and academia and all parts of society have a critical role to play here

to shape, along with industry labs,

what that should look like as we get closer to AGI

and the cooperation needed and the collaboration needed,

to prevent that kind of runaway race dynamic happening.

CA: OK, well, it sounds like you remain optimistic.

What's this image here?

DH: That's one of my favorite images, actually.

I call it, like, the tree of all knowledge.

So, you know, we've been talking a lot about science,

and a lot of science can be boiled down to

if you imagine all the knowledge that exists in the world

as a tree of knowledge,

and then maybe what we know today as a civilization is some, you know,

small subset of that.

And I see AI as this tool that allows us,

as scientists, to explore, potentially, the entire tree one day.

And we have this idea of root node problems

that, like AlphaFold, the protein-folding problem,

where if you could crack them,

it unlocks an entire new branch of discovery or new research.

And that's what we try and focus on at DeepMind

and Google DeepMind to crack those.

And if we get this right, then I think we could be, you know,

in this incredible new era of radical abundance,

curing all diseases,

spreading consciousness to the stars.

You know, maximum human flourishing.

CA: We're out of time,

but what's the last example of like, in your dreams,

this dream question that you think there is a shot

that in your lifetime AI might take us to?

DH: I mean, once AGI is built,

what I'd like to use it for is to try and use it to understand

the fundamental nature of reality.

So do experiments at the Planck scale.

You know, the smallest possible scale, theoretical scale,

which is almost like the resolution of reality.

CA: You know, I was brought up religious.

And in the Bible, there’s a story about the tree of knowledge

that doesn't work out very well.

(Laughter)

Is there any scenario

where we discover knowledge that the universe says,

"Humans, you may not know that."

DH: Potentially.

I mean, there might be some unknowable things.

But I think scientific method is the greatest sort of invention

humans have ever come up with.

You know, the enlightenment and scientific discovery.

That's what's built this incredible modern civilization around us

and all the tools that we use.

So I think it's the best technique we have

for understanding the enormity of the universe around us.

CA: Well, Demis, you've already changed the world.

I think probably everyone here will be cheering you on

in your efforts to ensure that we continue to accelerate

in the right direction.

DH: Thank you.

CA: Demis Hassabis.

(Applause)