Shulou(Shulou.com)11/24 Report--

Recently, Demis Hassabis, the founder of DeepMind, appeared on Lex Fridman's podcast and talked about a lot of interesting ideas.

At the beginning of the interview, Hassabis said bluntly that the Turing test is out of date because it is a benchmark proposed for decades, and that the Turing test is based on human actions and reactions, which is prone to a "farce" similar to the one Google engineer called conscious AI system some time ago: researchers talk to a language model and map their perception to the judgment of the model, which is not objective.

Since its establishment in 2015, the development of DeepMind in the field of artificial intelligence has brought surprises to the world again and again: from the game program AlphaGo to the protein prediction model AlphaFold, the technological breakthrough of deep reinforcement learning has solved the major scientific problems that have plagued human scientists for many years.

In this interview with Hassabis, he also made an interesting point that AI transcends the limitations of human intelligence. When human beings may be used to this three-dimensional world with time, AI may be able to achieve the intelligence to understand the world from twelve dimensions and get rid of the essence of tools, because there are still many deficiencies in our understanding of the world.

The following is an interview with Demis Hassabis:

1 from games to AILex Fridman: since when do you like programming?

Demis Hassabis: I started playing chess when I was about 4 years old. When I was 8 years old, I used the prize money I won in a chess game to buy my first computer, a zx spectrum, and later I bought books about programming. I fell in love with computers when I started making computer games. I thought they were magical and an extension of my mind. You can ask them to do some tasks and they will be solved when you wake up the next day.

Of course, all machines can do this to some extent, enhancing our natural abilities, such as cars that make us move faster than we run. But artificial intelligence is the ultimate performance that machines can do all learning, so my idea naturally extends to artificial intelligence.

Lex Fridman: when did you fall in love with artificial intelligence? When did you begin to learn that it can not only write programs and do mathematical operations while sleeping, but also perform more complex tasks than mathematical operations?

Demis Hassabis: it can be divided into several stages.

I am the captain of the youth chess team. I plan to become a professional chess player when I am about 10 or 11 years old. This is my first dream. At the age of 12, I reached the master level and was the second largest chess player in the world, second only to Judith Pologer. When I try to improve my chess skills, I first need to improve my thought process and think about how my brain comes up with these ideas. Why did it make mistakes? How can we improve this thinking process?

Like chess computers in the early and mid-1980s, I was used to having a branded version of Kasparov, which was not as powerful as it is today, but could be improved by practicing with it. At that time, I thought, it's amazing that someone programmed this chessboard to play chess. I bought a copy of David Levy's Chess computer Handbook published in 1984, which is a very meaningful book that allows me to fully understand how chess programs are made.

▲ Note: Kasparov, former Soviet Union, Russian professional chess player, chess grandmaster

My first artificial intelligence program was programmed by my Amiga, and I wrote a program to play Othello reverse thinking, which is a slightly simpler game than chess, but I used all the principles of the chess program, namely α-β search, etc.

The second stage is a game called "theme Park" designed when I was about 16 or 17 years old, which involves AI simulation in the game. Although it is simple by today's AI standards, it will respond to the way you play as a player, so it is also called sandboxie game.

Lex Fridman: can you tell me some of your key connections with AI? What does it take to create an AI system in a game?

Demis Hassabis: I trained myself in games when I was a kid, and then I went through a stage of designing games and writing AI for games. All the games I wrote in the 1990s took artificial intelligence as the core component. The reason for doing this in the game industry is because at that time I thought the game industry was at the forefront of technology, like John Carmack and Quake, which seemed to be played in games. We are still benefiting from it, like GPU, which was invented for computer graphics, but was later found to play an important role in AI. So at that time, I thought that there was cutting-edge artificial intelligence in the game.

In the early days, I participated in a game called "Black and White", which is the most profound example of the application of reinforcement learning in computer games. You can train a small pet in the game, and it will learn from the way you treat it. If you treat it badly, it will become mean and mean to your villagers and the small tribes you manage. But if you treat it well, it will also become kind.

Lex Fridman: the game's mapping of good and evil makes me realize that you can determine the outcome by the choices you make. Games can bring this philosophical meaning.

Demis Hassabis: I think games are a unique medium. As a player, you don't just consume entertainment passively. In fact, you are actively involved as a representative. So I think that's why games are more meaningful in some ways than other media, such as movies and books.

From the very beginning, we have thought deeply about AI, using the game as a testing ground for proving and opening the AI algorithm. This is why Deepmind initially used a large number of games as the main test platform, because games are very efficient, and it is easy to have indicators to see how the AI system is improved, the direction of thinking, and whether it is making gradual improvements.

Lex Fridman: suppose we can't build a machine that can beat humans at chess, then people will think that go is an unbreakable game because of the complexity of the combination. But in the end, the AI researchers built the machine, and humans realized that we were not as smart as we thought.

Demis Hassabis: it's an interesting journey of thinking, especially when I understand it from two perspectives (AI creators and gamers). It's more magical and a little bittersweet at the same time.

Kasparov calls chess a smart "fruit fly". I like that description, because chess is closely related to AI from the very beginning. I think every AI practitioner, including Turing and Shannon, as well as all the predecessors in this field, have tried to write a chess program. Shannon wrote the first program document about chess in 1949. Turing also wrote a famous chess program, but because the computer was too slow to run, he ran the program manually with pencils and paper to play with friends.

The emergence of DeepBlue is an important moment, it combines everything I like, including chess, computer and artificial intelligence. In 1996, it beat Garry Kasparov. After that, I was more impressed by Kasparov's mind than DeepBlue, because Kasparov is the human mind, not only can he reach the same level as computers in playing chess, but Kasparov can also do everything humans can do, such as riding a bike, speaking multiple languages, participating in politics, and so on.

Although DeepBlue has had glorious moments in chess, it actually refines the knowledge of chess masters into a program that cannot do anything else. So I think there is a lack of intelligence in the system, which is why we try to do AlphaGo.

Lex Fridman: let's talk briefly about the human side of chess. From the point of view of game design, chess is attractive because it is a game. Can you explain whether there is a creative tension between bishop (the elephant in chess) and knight (the horse in chess)? What makes games attractive and can span centuries?

Demis Hassabis: I'm thinking about that, too. In fact, many good chess players do not necessarily think about this from the perspective of game designers.

Why is chess so attractive? I think a key reason is the dynamics of different positions, you can tell whether they are closed or open, think about how different the elephant and the horse move, and then chess has evolved to balance the two. it's about three points.

Lex Fridman: so you think there is always dynamics, and the rest of the rule is to try to stabilize the game.

Demis Hassabis: maybe it's a bit like the case of chicken or egg, but the two achieve a beautiful balance, like the power of horses and knights, but their value is equal in the position of the whole universe. Over the past few hundred years, they have been balanced by humans, which I think gives creative tension to the game.

Lex Fridman: do you think the AI system can attract people to design games?

Demis Hassabis: that's an interesting question. If creativity is defined as coming up with something original and useful for a purpose, then the lowest level of creativity is like an interpolation expression, and the basic AI system has this ability. Show it millions of pictures of cats, and then give me an ordinary cat, which is called interpolation.

And like AlphaGo, it can be inferred. AlphaGo has come up with some great new ideas after playing millions of games against himself, such as taking 37 steps in the game, providing a strategy that humans have never thought of, even though we have been playing it for hundreds and thousands of years.

On top of this, there is another level, that is, whether we can jump out of the mindset and do real innovation. Can you invent chess instead of coming up with a move? Can chess, or other things like chess or go, be invented?

I think AI can do it one day, and the question now is how to assign this task to a program. We can not put high-level abstract concepts into artificial intelligence systems, they still lack something in the real understanding of high-level concepts or abstract concepts. For now, they can be combined and constructed, and AI can do interpolation and inference, but they are not really inventions.

Lex Fridman: to propose and optimize rule sets and set complex goals around these rule sets is something we can't do at present. But is it possible to take a specific rule set and run it to see how long it takes for the AI system to learn from scratch?

Demis Hassabis: actually, I've thought about it, and it's amazing for game designers. If there is a system that plays your game tens of millions of times, it may be able to achieve automatic balancing rules overnight. You can adjust the units or rules in the game through equations or parameters to make the game more balanced. It's a bit like giving a basic set to explore through Monte Carlo search or something like that, which would be a super powerful tool.

To balance automatically, it usually takes thousands of hours of training from hundreds of games, and balancing games like StarCraft and Blizzard is shocking, which takes testers year after year. So it's conceivable that when these things become effective enough at some point, you might want to do it overnight.

Lex Fridman: do you think we are living in Simulation?

Demis Hassabis: yes. Nick Bostrom put forward the famous simulation theory for the first time, but I don't believe in it. In a sense, we are in some kind of computer game, or our future generations will reshape the earth in some way in the 21st century.

The best way to understand physics and the universe is to understand it as an information universe from a computational point of view. In fact, information is the most basic unit of reality. Compared with matter or energy, physicists would say E=mc ², which is the foundation of the universe. But I think that information is probably the most basic way to describe the universe, and it can specify the right matter for energy or matter. So we can say that we are in some kind of simulation. But I don't agree with these ideas to discard billions of simulations.

Lex Fridman: based on your understanding of the general term machine and your understanding of computers, do you think there is anything other than computer power in the universe? You don't agree with Roger Penrose (mathematical physicist)?

Demis Hassabis: Roger Penrose is famous and has been involved in many wonderful debates. I read his classic the Emperor's New brain, and he explained that consciousness in the brain needs more quantum things. I've been thinking about what we're doing at work, and in fact, we're pushing Turing machines or classical computing to the limit. What is the limit of classical calculation? I also studied neuroscience, which is why my doctor chose this direction, whether there is a quantum presence in the brain from a neuroscientific or biological point of view.

So far, most neuroscientists and biologists would say that there is no evidence of any quantum systems or effects in the brain, which can mostly be explained by classical theory and biological knowledge. But at the same time, starting with what Turing machines can do, including AI, the process has been going on, especially over the past decade. I don't bet on how far universal Turing machines and classical computing paradigms can go, but what happens in the brain may be imitated on machines without the need for metaphysics or quantum things.

2 Al for scienceLex Fridman: let's talk about AlphaFold. Do you think human thinking comes from this neural network-like, biological computational paste, rather than working directly in the mind?

Demis Hassabis: in my opinion, the greatest miracle in the universe is that we have only a few pounds of paste in our skulls, which is also the most complex object in the brain and the universe known so far. I think this is an amazingly efficient machine, which is one of the reasons I've always wanted to build AI. By building an agent like AI and comparing it with human thinking, it may help us to know the uniqueness of the mind, and the real secrets, consciousness, dreams, creativity, emotions, and so on.

Now there are a lot of tools to do this. All neuroscience tools and FMI machines can be recorded, and AI computing power can be used to build intelligent systems. What the human mind can do is amazing. Human beings have created things like computers, and think about and study these problems, which is also proof of the human mind and helps us to understand the universe and human thoughts more clearly. It can even be said that we may be the mechanism by which the universe tries and understands its beauty.

On the other hand, the basic building blocks of biology can also be used to understand the human mind and body. It's amazing to simulate and model from the basic construction. You can build larger and more complex systems, even the whole human biology.

Another problem that is considered impossible is protein folding, and AlphaFold solves the problem of protein folding, which is one of the biggest breakthroughs in the history of structural biology. Protein is essential to all life, and every function of the body depends on protein.

Proteins are specified by their genetic sequences (also known as amino acid sequences) and can be regarded as their basic building blocks. They will fold into a three-dimensional structure in the body and in nature, which determines its function in the body. In addition, if you are interested in drugs or diseases, you want to use a drug compound to block the action of proteins, as long as you understand the three-dimensional structure of the binding sites on the protein surface.

▲ Note: in July 2021, DeepMind published AlphaFold predictions for the first time through a database established in cooperation with the European Molecular Biology Laboratory (EMBL). The initial database contains 98% of all human proteins.

Lex Fridman: the essence of the protein folding problem is, can you get an one-dimensional letter string from an amino acid sequence? Can the three-dimensional structure be predicted immediately through calculation? This is a major challenge in the field of biology for more than 50 years. Christian Anfinsen, a Nobel laureate in 1972, first stated that it was possible to move from amino acid sequences to three-dimensional structures.

Demis Hassabis: Christian Anfinsen's sentence opens up 50 frontiers of computational biology that are trapped and not well done.

Before the advent of AlphaFold, this was done through experiments, and it was very difficult to crystallize proteins. Some proteins could not crystallize like membrane proteins, and they had to use expensive electron microscopes or X-ray crystal analyzers to obtain three-dimensional structures and visualize their structures. With AlphaFold, two people can predict three-dimensional structures in a matter of seconds.

Lex Fridman: there is a dataset that trains on this dataset and how to map amino acids. It is hard to believe that this small chemical computer can be calculated in some distributed way, and it is very fast.

Demis Hassabis: maybe we should talk about the origin of life. In fact, the protein itself is a magical little creature and animal machine. Cyrus Levinthal, the scientist who proposed the Lewenthal paradox, roughly calculated that the average protein may be 2000 amino acid bases long and can be folded in 10 to 300 different ways. In nature, physics solves this problem in some way, and proteins fold up in your body in milliseconds, or a second.

Lex Fridman: this sequence has a unique way to form itself, and it finds a way to remain stable in the face of great possibilities. Dysfunctions may occur in some cases, but most of the time it is a unique mapping, and this mapping is not obvious.

Demis Hassabis: if there is usually a unique mapping of health, what exactly is the problem when you get sick? For example, there was a guess about Alzheimer's disease that it was misfolded by folding beta-amyloid in the wrong way, so that it became entangled in neurons.

Therefore, to understand health, function, and disease, it is extremely important to understand how they are structured and what these things are doing. The next step is that when proteins interact with something, they change shape. So in biology, they are not necessarily static.

Lex Fridman: maybe you can come up with some solutions to AlphaFold. Unlike games, this is a real physical system. What is very difficult to solve? What is relevant to the solution?

Demis Hassabis:AlphaFold is the most complex and probably the most meaningful system we have built so far.

We initially built AlphaGo and AlphaZero related to games, but the ultimate goal is not just to crack games, but to use them to guide general learning systems and meet real-world challenges. We want to focus more on scientific challenges such as protein folding, and AlphaFold is our first important point of proof.

In terms of data, the number of innovations requires more than 30 different composition algorithms that are put together to crack protein folding. Some major innovations revolve around physics and evolutionary biology, establishing hard coding to constrain things such as bond angles in proteins, but without affecting the learning system, so the system can still learn physics from cases.

Assuming that there are only about 150000 proteins, even after 40 years of experiments, only about 50, 000 protein structures will be found. The training set is much less than the amount of data commonly used, but uses a variety of techniques such as self-extraction. Therefore, when using AlphaFold to make some very confident predictions, it is very important for AlphaFold to put it back into the training set to make the training set larger.

In fact, in order to solve this problem, a lot of innovation is needed. AlphaFold produces a histogram, a matrix of paired distances between all molecules in a protein, which must be a separate optimization process to create a three-dimensional structure. To make AlphaFold truly end-to-end, you can skip the intermediate step directly from the base sequence of amino acids to the three-dimensional structure.

From machine learning, it can also be found that the more end-to-end, the better the system, and the system is better at learning constraints than human designers. In this case, a three-dimensional structure is better than an intermediate step, because it has to move on to the next step manually. The best way is to let the gradient and learning flow through the system, from the end point to the desired final output, and then to the input.

Lex Fridman: the idea of AlphaFold, which may be an early step in a long journey in biology, do you think the same approach predicts the structure and function of more complex biological systems, multi-protein interactions? as a starting point, can it simulate larger and larger systems and eventually simulate things like the human brain and the human body? Do you think this is a long-term vision?

Demis Hassabis: of course, once we have a strong enough biological system, treating diseases and understanding biology is my top priority on To Do List, which is one of the reasons I personally promote AlphaFold. AlphaFold is just the beginning.

AlphaFold solves the huge problem of protein structure, but biology is dynamic, and everything we study is liquid protein binding. It is my dream to react with molecules, build pathways, and eventually form a virtual cell. I've been talking to a lot of biology friends, including Paul Nurse, a biologist at the Creek Institute. For biology and disease discovery, building a virtual cell is incredible because you can do a lot of experiments on the virtual cell and then go to the lab to verify it in the final stage.

In terms of the discovery of a new drug, it takes about 10 years from setting a target to having a candidate drug, which may be reduced by an order of magnitude if most of the work can be done in a virtual cell. In order to realize virtual cells, it is necessary to establish an understanding of the interaction between different parts of biology. Every few years, we talk to Paul about this. After AlphaFold last year, I said it was finally time for us to do it, and Paul was very excited. We have some cooperation with his laboratory. On the basis of AlphaFold, I believe that the Biology Society has made some amazing progress, and it can also be seen that some communities have been doing it since the open source of AlphaFold.

I think artificial intelligence systems may one day solve problems like general relativity, not just by dealing with content on the Internet or public health care. It will be very interesting to see what it can come up with. It's a bit like our previous debate about creativity inventing go, not just coming up with a good go move. If you want to win a prize like a Nobel Prize, what it needs to do is to invent go, not to be designated by human scientists or creators.

Lex Fridman: many people do see science as standing on the shoulders of giants, but the question is how much have you really achieved on the shoulders of giants? Maybe it just absorbs different types of results from the past and finally provides groundbreaking ideas from a new perspective.

Demis Hassabis: this is a big mystery. I believe that in the past decade or even the next few decades, many major new breakthroughs will occur at the intersection of different disciplines, and some new connections will be found in these seemingly unrelated areas. People can even think that deep thinking is a cross-discipline between neuroscience thought and AI engineering thought.

Lex Fridman: you have a paper on magnetic control of tokamak plasma through deep reinforcement learning, so you are looking to use deep reinforcement learning to solve nuclear fusion and control high-temperature plasma. Can you explain why AI finally solved this?

Demis Hassabis: over the past year or two, our work has been very interesting and fruitful, and we have launched a lot of my dream projects, which are related to the science field that I have collected over the years. If we can participate in the promotion, it may have a transformative impact, and the scientific challenge itself is a very interesting question.

At present, nuclear fusion faces many challenges, mainly in physics, materials, science and engineering, and how to build these large-scale nuclear fusion reactors and accommodate plasma.

We work with the Lobsang Federal Institute of Technology (EPFL) and the Swiss Institute of Technology, who have a test reactor that is willing for us to use. This is an amazing test reactor on which they try all kinds of pretty crazy experiments. What we are looking at is, what is the bottleneck when entering a new field such as nuclear fusion? What is the underlying problem that hinders the operation of nuclear fusion in terms of the first principle?

In this case, the plasma control is perfect. The plasma is 1 million ℃, hotter than the sun, and obviously no material can hold it. So there has to be a very strong superconducting magnetic field, but the problem is that the plasma is quite unstable, just like holding many stars in a reactor, predicting in advance what the plasma will do. you can move the magnetic field in millions of seconds to control what it does next.

If you think of it as a reinforcement learning prediction problem, it seems perfect, with a controller that can move the magnetic field and cut, but previously used a traditional controller. I want a controllable rule that they can't react to plasma right now, it has to be hard-coded.

Lex Fridman:AI finally solved nuclear fusion.

Demis Hassabis: last year we published a paper on solving this problem in the journal Nature, fixing the plasma in a specific shape. In fact, it's almost like carving the plasma into different shapes, controlling it and keeping it there for a record time. This is an unsolved problem of nuclear fusion.

It is important to include it in the structure and maintain it, and there are some different shapes that are more conducive to energy generation, called drops, and so on. We are talking to many fusion startups to see what is the next problem that can be solved in the fusion field.

Lex Fridman: another fascinating thing in the title of the paper is to push the frontier of density functions by solving the problem of fractional electrons. Can you explain the work? Can AI model and simulate any quantum mechanical system in the future?

Demis Hassabis: people try to write an approximation of the density function and a description of the electron cloud to see how two elements interact when they are put together. What we are trying to do is to learn a simulation, to learn a chemical function that can describe more chemical types.

So far, AI can run expensive simulations, but it can only simulate very small and very simple molecules, and we cannot simulate large materials. Therefore, to establish functional approximations to show its equations and describe what electrons are doing, all material science and properties are controlled by how electrons interact.

Lex Fridman: summarize the simulation through the function to approach the actual simulation results. The difficulty of this task lies in running the complex simulation and learning the mapping task from the initial conditions and simulation parameters. What will the learning function be?

Demis Hassabis: it's tricky, but the good news is that we've done it, and we can run a lot of simulations on the computing cluster, that is, molecular dynamics simulations, which generate a lot of data. In this case, the data is generated. That's why we use game simulators to generate data, because we can create more data at will. If there are free computers in the cloud, we can run these calculations.

3 AI and Human Lex Fridman: how do you understand the origin of life?

Demis Hassabis: I think the ultimate use of AI is to accelerate science to the extreme. It's kind of like the tree of knowledge. If you imagine that this is all the knowledge to be acquired in the universe, but so far, we have almost touched the surface of it. AI will speed up this process and explore the knowledge tree as much as possible.

Lex Fridman: my intuition tells me that the tree of human knowledge is very small, considering our cognitive limitations. Even if we have tools, we still can't understand many things. This may be the reason why non-human systems can go further.

Demis Hassabis: yes, it's very possible.

But first, these are two different things. Just like what we understand today, what the human mind can understand, what we need to understand as a whole, there are three concentric trees that you can think of as three bigger trees, or explore more branches of this tree. With AI, we will explore more.

The problem now is, if you think about what we can understand in general, there may be things that can't be understood, such as things outside the simulation, or things outside the universe.

Lex Fridman: because the human brain has become accustomed to the state of this three-dimensional world with time.

Demis Hassabis: but our tools can go beyond that. They can be 11 dimensions, 12 dimensions.

The example I often give is that when I played chess with Gary Kasparov, we talked about chess and things like that. If you are good at chess, you can't Gary his moves, but he can explain it to you. You can think of it as ex post facto reasoning. There is a further explanation that you may not be able to invent it, but you can understand and appreciate it as much as you appreciate Vivaldi or Mozart.

Lex Fridman: I want to ask some crazier questions. For example, do you think there are alien civilizations outside the earth?

Demis Hassabis: my personal opinion is that we are alone at the moment. We already have all kinds of telescopes and other detection techniques, trying to find signals from other civilizations in space, and if there are many alien civilizations doing this at the same time, we should hear noises from outer space. But the truth is, we didn't get any signals.

Many people will argue that there are alien civilizations in the world, but we have not really searched properly, or we are looking for the wrong band, or we may have used the wrong equipment. We do not realize that the form of alien existence is very different, and so on. But I don't agree with these views. in fact, we have done a lot of exploration. if there were so many alien civilizations, we should have discovered it a long time ago.

Interestingly, if the earth is a lonely civilization, it's comforting from a Great Filters point of view, which means we've already passed the filter.

Going back to the question about the origin of life you just asked, life originated from something incredible, and no one knows how it happened. I wouldn't be surprised to see some kind of life form of a single cell, such as bacteria, outside the earth. But because of its ability to capture and use mitochondria for me, the difficulty of multicellular life is unprecedented.

▲ diagram note: the large filter theory mentioned by Demis Hassabis

Lex Fridman: do you think you need to be conscious to be truly intelligent?

Demis Hassabis: I personally think that consciousness and wisdom are doubly separated, so we can realize consciousness without wisdom, and vice versa.

For example, many animals are self-aware, socialize and dream. They can be defined as having a certain degree of self-awareness, but they are not intelligent. But at the same time, artificial intelligence that is very smart at a certain task, they can play chess or perform other tasks very well, but they do not have any self-awareness.

Lex Fridman: some time ago, an engineer at Google thought that a certain language model was perceptual. Have you ever encountered a perceptual language model? If there is "perception" in a system, how do you understand this situation?

Demis Hassabis: I don't think any AI system in the world is conscious or perceptual. This is how I really feel when I interact with AI every day. The so-called perception is more of our brain's own projection, because it is a language model, closely related to wisdom, so it is easy to personify the system. That's why I think the Turing test is flawed because it's based on human reaction and judgment.

We should talk to top philosophers about consciousness, such as Daniel Dennett and David Charmers, and others who think deeply about consciousness. At present, there is no accepted definition of consciousness, and if you ask me, I think the definition of consciousness is the feeling that information brings when it is processed.

Lex Fridman: let me ask you a dark personal question. You said to create the most powerful super artificial intelligence system in the world. As the old saying goes, absolute power leads to corruption, and you are likely to be one of them, because you are the one most likely to control the system. Would you think about that?

Demis Hassabis: I think about what defenses against this kind of corruption all the time.

The tools or technologies in the best interests of mankind allow us to enter a radical world, and we are faced with many daunting challenges. AI can help us solve problems, eventually lead mankind to the ultimate prosperity, and even find aliens. The creator of AI, the culture that AI depends on, the values that AI has, and the builders of AI system will all influence its development. Even if the AI system will learn on its own, most of its knowledge will carry a certain residue of existing culture and creator values.

Different cultures make us more divided than ever before. Perhaps when we enter an era of extreme affluence, when resources are not so scarce, we do not need fierce competition, but can turn to better cooperation.

Lex Fridman: when resources are severely constrained, some atrocities occur.

Demis Hassabis: resource scarcity is one of the causes of competition and destruction. All mankind wants to live in a kind and safe world, so we must solve the problem of scarcity.

But this is not enough to achieve peace, because there are other things that can lead to corruption. AI should not be run by just one person or organization. I think AI should belong to the world, belong to human beings, and everyone should have a say in AI.

Lex Fridman: do you have any suggestions for high school and college students? If young people have the desire to engage in AI, or want to influence the world with their own power, how should they get a career that they are really proud of? How to find the ideal life?

Demis Hassabis: I always like to say two words to young people. The first sentence is, where is your real passion? Young people should explore the world as much as possible. When we are young, we have enough time to take the risks of exploration. To find the connection between things in your own unique way, I think this is a good way to find the passion.

The second sentence is, know yourself. It takes a lot of time to understand what is the best way to work, when is the best time to work, what is the best way to learn, and how to deal with stress. Young people should test themselves in different environments, try to improve their weaknesses, find out their unique skills and strengths, and then hone them. These are your future values in this world.

If you can combine these two things, find your passion and develop your own unique and powerful skills, then you will gain incredible energy and make a huge difference to the world.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.