2026-08-23 Lucas di Grassi A Conversation with David Deutsch
YouTube Lucas di Grassi’s channel
Duration: 01:26:38
Participants: Lucas di Grassi and David Deutsch.
Topic guide
Timestamps link to the corresponding point in the recording.
- 00:00 - Introduction and Deutsch’s books
- 01:53 - Physics, philosophy, and Karl Popper
- 04:50 - Knowledge grows from problems, not data
- 08:21 - Children, astrology, and language learning
- 16:30 - Error correction in motorsport and creative work
- 24:24 - Democracy and nonviolent error correction
- 35:34 - Family institutions and disagreement
- 37:00 - Universal computation and explanatory universality
- 43:23 - Quantum computing, many worlds, and Deutsch’s early work
- 58:16 - Free will, determinism, and creative knowledge
- 65:16 - AGI, consciousness, personhood, and classical computation
- 70:49 - Atheism, truth, happiness, and the common good
- 77:30 - Error-correcting institutions from minds to governments
- 79:50 - Humanity’s future, AI, and optimism
Transcript
Edited for readability. Verbal fillers, repetitions, and abandoned starts have been removed without changing the substance of the conversation.
Lucas di Grassi
Today I have the honor of welcoming David Deutsch, an extraordinary physicist and, in my opinion, one of the brightest minds in the world. He is not especially well known in Brazil, but he wrote two books that are personal favorites of mine. I’ve read them several times, trying to understand them, because they go deeply into universal computation, error-correcting mechanisms, the philosophy of knowledge, quantum mechanics, the multiverse, and other ideas that shape reality. David’s work aligns closely with my own way of thinking. Thank you very much for joining us today.
David Deutsch
Cool. Thank you for having me.
Lucas di Grassi
David, I’d like to begin with a simple conversation aimed especially at Brazilians. Brazil has 220 million people, and when I discuss these subjects with knowledgeable people in Brazilian society, I find that your powerful ideas and the thinking behind them are still not widely known. Quantum mechanics and quantum computing, for example, are widely discussed, and you developed much of the theory behind quantum computing many years ago. For someone in Brazil who has never heard of you, studied quantum physics, or read your books, how would you describe yourself and your worldview in a few minutes?
David Deutsch
The reason I’m a physicist is that I have always wanted to understand the physical world as well as possible. Some of our deepest knowledge of the physical world lies in the most fundamental theories of physics. But as I studied physics and encountered the problems in current physical theories, I increasingly realized that one cannot address them without also thinking about philosophy.
The philosophy I found most helpful was Karl Popper’s. It was only the second philosophy I tried. The first was Bertrand Russell’s, which was also appealing. It connected with science and rejected mysticism and supernatural explanations. But Popper seemed to me to operate on a different level, really getting to grips with philosophical problems. His ideas were relevant to problems in physics, such as what brings about the counterintuitive predictions of quantum theory. There are no conservative options there. Every theory of how those predictions come about is astonishingly counterintuitive.
Lucas di Grassi
To explain the Popper side for a general audience, his central concern is how humans can know anything and how we create knowledge. One of your most interesting points, and something very few people I speak with understand, is that knowledge does not come from data or observation alone. As you say in your book, you can watch the sun rise and set every day without having an explanation of it. Is that one of the key ideas in your epistemology?
David Deutsch
Yes, exactly. New knowledge always grows from problems, not from data. Data enters after we have a problem. For hundreds of thousands of years, people watched the sun rise every day, far more attentively than we do now. Millions may have wondered why and invented explanations. We know most of those explanations were false, even though those people had more observations available to them than we do today.
So what was wrong with their approach, and what did we begin doing differently in the Scientific Revolution, with Galileo and Newton, and then in the Enlightenment? I think the difference was that we sought good explanations. Earlier people were also trying to solve problems, in a sense, but they regarded almost any story as a solution. A good explanation cannot be altered without ceasing to explain the phenomenon. Once altered, it would no longer account for what it was meant to explain.
Lucas di Grassi
So specific explanations are more powerful than generic ones? I don’t mean “generic” in the sense of explaining more. I mean that any small change would undermine the explanation.
David Deutsch
Yes. Sometimes an explanation can account for more, and that may be another reason it cannot be altered: it also has to account for those other things. My favorite layman’s example is a conjuring show. You watch a magician perform the cups-and-balls trick. Whenever he lifts a cup, the ball is absent when you expect it and present when you don’t. You can predict what will happen each time: whatever you expect, the magician will confound it.
But what you really want is not a correct prediction. You want to know how he did it. You want an explanation of the events you cannot see. It is the same with the sun and with every scientific theory. Scientific explanations explain the seen in terms of the unseen.
Lucas di Grassi
I understand that. I have two children, a boy of eight and a girl of five, with a third on the way. I often draw on your work in bringing them up. I don’t simply tell my son Leo that Santa Claus doesn’t exist. I encourage him to ask why things happen and try to explain them. A child’s mind begins with observations and tries to predict, but it doesn’t necessarily reason about what lies behind them.
Many intelligent, cultured adults seem to view the world in the opposite way. Take astrology. It is a pseudoscience, if we can even call it that, yet it is enormously popular. It offers vague predictions about how your day will go. If one appears to come true, our biases tell us astrology explained it. If it does not, we shrug and say it can be wrong. This is the opposite of the approach you describe. Why do so many intelligent and cultured people still believe in such primitive forms of reasoning?
David Deutsch
Popper’s theory is the best theory of knowledge we have. No doubt it will be improved, but a correct theory of knowledge is deeply counterintuitive. When someone tells you that Earth is round or that the sun will rise tomorrow, it is hard to realize that you are not simply receiving an idea from them. There is no transmission from their mind, through mouth and ear, into your brain. Their behavior presents a problem that you try to solve, and the solution is a conjecture you create. Whenever you form a theory that seems to have come from another person, you have really invented it yourself.
Young children perform the first enormous feat of this kind when they learn their native language. Almost nobody teaches a child by saying, “A table is a thing with four legs, and this is one.” Children learn most words by hearing them used, trying them, and discovering which utterances other people understand. That is an extraordinary ability. They want to understand what lies beneath the words. Everyone hears words, but nobody hears grammar, yet children learn grammar too.
Sometimes a child discovers a regularity and applies a general rule incorrectly. They say, “Yesterday I goed to the shops,” rather than “I went.” We know they conjectured the rule precisely because they applied it where it does not hold. A rule is intangible. They did not see it anywhere. In this case it is false, although, like many false theories, it contains some truth. Our best theories probably contain falsehoods too.
You want to bring up your children with a Popperian scientific attitude. The practice comes first; they are already doing it. Why, then, do so many people later believe ideas such as astrology? Because those theories tell them that whatever seems true is true, the opposite of the scientific attitude. If children believed that, they would never learn grammar. They would merely repeat the words they had heard and never become fully competent language users. Yet almost all humans do, and learning a native language may be our greatest intellectual achievement.
People believe astrology because it seems obvious. Over time, however, many learn not to behave as they did as children. They are punished for mistakes or mocked for saying unpopular things. Young children are less inhibited. If a child says, “I goed to the shops,” and someone replies, “No, I went to the shops,” the child considers how that correction fits the rule, whether to modify the rule or invent another. This happens unconsciously.
Around the age of eight, I think, children learn perhaps twenty words a day. You do not see those words entering their minds, and nobody tells them what each one means. All genuine learning is done by the learner, not the teacher.
Lucas di Grassi
I understand. Another idea of yours that I especially like is the error-correcting mechanism. I’m a racing driver. I raced in Formula One and helped create Formula E. I’ve been racing since I was seven, for about thirty-five years.
David Deutsch
Well done. I really admire that.
Lucas di Grassi
Thank you. Racing involves a great deal of physics and engineering. You learn early that understanding how everything works gives you an advantage. A better car results from combining technology and sport. Error correction is something I always push my engineers and team to practice. Before reading your books, I lacked the theoretical basis, but your account made it click: this is a much broader method for creating knowledge and refuting theories than its use in racing.
I also try to teach it to my children. A theory, idea, or experiment must be falsifiable if we are to make progress. When authority or a lack of data prevents us from testing or refuting a hypothesis, we make no progress. Refutation matters more than proof because proof cannot provide certainty, whereas refutation allows us to move on. Yet neither children nor engineers naturally work this way after being trained by the system; one has to push them in this direction. One of the few principles we deliberately build into the racing team’s culture is that error correction is the best way forward, even in simple matters. Can you explain how this broader view of error correction fits into the ideas in your books?
David Deutsch
I agree that when you’re building an adult team focused on creating new knowledge and succeeding in a new way, you have to encourage a Popperian approach. But I don’t think that’s true of children. They already have it. Your job is to protect them from the many cultural forces, including one another, that make them reluctant or afraid to proceed that way.
People who escape those forces encounter a principle I often state this way: everybody who is highly creative in a field is a Popperian in that field. They may be superstitious or unscientific elsewhere, but within that field they are Popperians.
Lucas di Grassi
Exactly, within that space. We see this often. Within a specific frame of reference and system, people are highly Popperian and rational. Outside it, they fail to apply the same reasoning to other parts of their lives.
David Deutsch
That’s very common.
Lucas di Grassi
Exactly. Sorry, continue.
David Deutsch
Yes. The reason is that this is the only way to create knowledge. Put backwards, people who think this way are the ones who succeed creatively. But all children do it. They must. We do not see it written in neon lights on their foreheads, but it must be happening because it is the only way knowledge is created.
Brilliant engineers, racing drivers, and physicists can have very different personalities and enter their fields in different ways, but they share a fearlessness about conjecturing novelties. They ask, “What if we did this? What if we didn’t have four wheels? What if we had five?” Or something like that.
Lucas di Grassi
That is a good example. People are constrained by their habitual frame of reasoning. A lot of engineering proceeds by analogy rather than pure conjecture. Instead of making analogies and incremental changes, you can ask from first principles, “How many wheels would the best possible car have?” It doesn’t matter that cars currently have four. The question may sound absurd, but it creates a test that can guide you in the right direction.
David Deutsch
Right. Another important point is that when you make a mistake, you may need to go back and find a different solution. But the mistake may also enable your next step; it can become part of the path to the right answer.
Again, this is what we do when learning language. If we never made mistakes, we would never learn its real grammar or much of its vocabulary. In Popper’s phrase, everybody makes bold conjectures to solve problems. When you are learning language, the problems are: What are they saying? How can I say something they will understand?
You make mistakes. When you learn twenty words a day, you do not learn all of them correctly at first. You learn twenty conjectures about those words. Some will be right, some half-right, and some completely wrong. The next time you use one, you may see that the other person does not understand. You try another formulation and learn from the failure. Usually no one corrects you directly. You failed to solve the problem of communication, so you make a new conjecture that improves on the old one.
Lucas di Grassi
That leads to a problem we share. Let me extend error correction to the level of a country. Brazil is one of the world’s largest democracies. Like you, I strongly favor democracy, not for its own sake but because it creates, at least in theory, a mechanism through which a system of government can correct errors and move forward. It is not best by definition. Its competing parts should balance one another, removing what doesn’t work.
Britain is going through an interesting phase after several recent changes of leader. Brazil has an election in October. These are turbulent times, full of tribal politics, social-media biases, and irrationality. People rarely ask what the mistakes were or how to correct them.
Brazil has a presidential rather than parliamentary system, but we face a similar problem: more and more mechanisms seem designed to suppress error correction. There is censorship, prohibited language, and claims that cannot be challenged. The same philosophy of error correction applies to a physicist studying quantum mechanics, a racing driver testing a car, and a democracy. You’re very active on X and have discussed politics quite a lot recently. How do you view this problem?
David Deutsch
As you say, I accept Popper’s approach to large-scale politics. The first important point is to judge political systems by their institutions rather than their day-to-day decisions. Those decisions will be false, inconsistent, sometimes disastrous, and sometimes good. The institutions matter much more, especially whether they promote nonviolent error correction or suppress even nonviolent correction.
There is no government, policy, or leader that does not make mistakes. Even the best leader will make them. If they cannot be corrected, disaster follows. Some regard Napoleon as a great leader; others, a monster. But even if he really was a great leader, he created a system in which his errors could not be corrected, and millions had to die. Even if he had been the best possible leader, that is not the important thing. What matters is having institutions capable of removing him or changing his policies long before millions die, and of finding a different solution.
I think this feature has reached its best realization in Britain, although the politicians themselves are not necessarily very good.
Lucas di Grassi
In politics, though, one of the main difficulties with error correction is defining the error. How do you define it over the short, medium, or long term? Whatever the decision, how do you frame the problem and obtain the right feedback?
David Deutsch
That is not a feasible property of a political system. You cannot design it to be intentionally short-term, medium-term, or long-term. That choice must itself be subject to error correction.
Consider how a good system avoids violence. Suppose a group has an unpopular idea that its members strongly believe is true. They lose an election and decide that the public does not understand. The minority thinks it must act, and that is where violence comes from.
A good democracy in Popper’s sense instead says: If you are right, why can’t you persuade anyone? They may reply, “Because they’re all from the wrong class.” But they are still humans. They are universal and can respond to arguments. In Britain in the late eighteenth and early nineteenth centuries, middle-class people persuaded the upper class to abolish slavery. The slave owners were compensated, but they could no longer make money from slavery and had to invest elsewhere.
In America, the problem could not be solved that way, so there was a civil war in which one or two percent of the population died. Their error-correction process had a flaw. The Constitution was intended to prevent the president from becoming a tyrant, but it had the unfortunate side effect that once the Supreme Court ruled something constitutional, ordinary persuasion no longer worked. Even persuading 50 percent of voters to oppose slavery would not abolish it. Because the Court had ruled slavery legal, changing that required a supermajority through a process designed to resist change.
You may reply that without such barriers, an attempt to correct one error could create a worse one. Yes. No system can prevent that. As usual, the remedy is more error correction: better, more sophisticated processes built into what ordinary people expect from government, from one another, and from the economy.
Suppose you believe that Mr. Burnham will ruin the country, and you are the previous prime minister. You leave peacefully because he was selected by the constitutional error-correcting process. You may think it selected wrongly this time, but the essential fact is that this error can also be corrected. In a good system like Britain’s, the previous leader does not merely leave peacefully. He would fight, kill, and die to defend the right of the new leader, whom he believes is bad, to take his place. That is the essential property.
Lucas di Grassi
Exactly. I had never thought of it in quite this way, but instead of focusing on policies or the leader, the metric for whether a democracy is eroding is the number and strength of its error-correcting mechanisms. We should look for flaws or failed mechanisms and change them so that error correction improves over time.
David Deutsch
I totally agree. We may disagree about whether the same applies to a parent and an eight-year-old. If an eight-year-old wants ice cream in the middle of the night, what matters most is not whether he gets it or whether it harms him. It is whether the institutions that determine the outcome are peaceful, and whether you as a parent will defend even the wrong decision if it emerged from family institutions capable of solving problems.
Particular decisions will always go wrong. We must focus on the institutions governing how policies change. In the best case, those institutions rely on persuasion. If persuasion fails, try more persuasion. In the meantime, in my view, he gets the ice cream.
Lucas di Grassi
The beauty of this is that you can apply it both to raising a child and to governing a large democracy. I couldn’t agree more. But we are running out of time and there is still so much to cover. I want to turn to The Beginning of Infinity.
David Deutsch
Right.
Lucas di Grassi
Your latest book.
David Deutsch
Yes.
Lucas di Grassi
And an amazing book. I read it with great optimism, and I draw on it often in current debates about AI. People say, “AI will solve every problem, and then there will be no more problems.” But that isn’t true.
David Deutsch
No.
Lucas di Grassi
There will always be problems.
David Deutsch
Yes.
Lucas di Grassi
Every solution will give rise to new problems.
David Deutsch
Yes.
Lucas di Grassi
Regardless of the level of intelligence. There are many themes in The Beginning of Infinity, but first I’d like you to explain several key ideas to a layperson: universal computation and humans as universal explainers. Every universal computer can solve the same class of problems. None is fundamentally better or worse; differences reduce to time and memory.
People often tell me that aliens may possess knowledge we could not even understand. I reply that perhaps we could understand it, given enough time and memory. I use this argument often, but many people do not accept universality as a basic truth or understand exactly what it means.
David Deutsch
Yes.
Lucas di Grassi
Could you explain why humans are universal explainers, why problems will always emerge, and why AI will not solve them all?
David Deutsch
The first distinction is between hardware universality and software universality. The brain, among other things, is a universal Turing machine. Its hardware can perform arithmetic and every other task a universal Turing machine can perform.
Then there is software universality, which people tend not to understand because it is counterintuitive. This is explanatory universality, which you just mentioned. Imagine aliens with a computer that can handle a more complicated function than ours. Suppose they can decode a 256-bit code, while we can decode only a 128-bit one. Anyone who understands computers knows that the difference cannot be a different kind of computation. Universal computers can differ in speed and memory capacity. Add enough memory to ours and run their program, and it can do anything theirs can.
A similar but distinct principle applies to explanations. Could aliens explain something that we could never explain? Suppose they have enormous heads, as in science fiction. Could they possess explanations we could never even contemplate or understand?
Lucas di Grassi
And we can replace aliens in that sentence with AI.
David Deutsch
AI, same thing.
Lucas di Grassi
It doesn’t matter whether we say aliens or AI.
David Deutsch
Exactly. It also applies to children and adults. It is all the same theory of knowledge. Suppose, for reductio ad absurdum, that an AI with enormous banks of memory and processing power can understand something that we can never understand. Why can’t we? Its understanding consists of a program in memory. Nothing else can be happening except a running program. What could prevent a human brain from understanding it? Only memory capacity and speed.
But we can augment our memory and speed just as the computer does. I already carry additional memory and speed. With this, I can do some things a million times faster than I could without it. The supposedly superhuman AI’s act of understanding cannot be anything other than a computer program, because of Turing universality. The only reasons we could not run it are insufficient memory or speed, and we can augment both with external tools.
Lucas di Grassi
That leads to another question. Does quantum computation represent a different form of computation, or does it obey these laws too?
David Deutsch
It obeys these laws. Technically, it is a more powerful form of computing than Turing computing. However, a Turing computer can emulate a quantum computer with vastly more memory, or use a quantum computer as a subroutine.
Lucas di Grassi
Understood. So a quantum computer is still Turing universal, but structurally different and able to perform calculations in another way. Is that right?
David Deutsch
Yes. It is structurally superior. It can perform operations that a classical computer cannot perform without an exponential increase in the required time or memory. An exponential increase amounts to trying every possibility, which is not feasible. However, a classical computer can use a quantum computer as an ancilla. It can program the quantum computer and receive an answer, so there is no fundamental obstacle.
Lucas di Grassi
One of your major contributions was the argument that quantum computers work by processing information in parallel universes. Before we get to that, could you explain the basic disagreement in quantum physics for people who know little about it?
Quantum discussions attract a lot of woo. Whenever something is unknown or multiple possibilities exist, someone invokes “quantum” because almost nobody understands it. I don’t understand much either, but I try to distinguish the genuinely strange and counterintuitive phenomena of quantum theory from complete nonsense.
Many scientists accept some version of the Copenhagen interpretation: multiple possibilities exist until observation collapses the wave function into the reality we experience. You and other scientists favor the Everettian interpretation, which the mathematics leads us to: every outcome occurs in parallel universes. More people seem to be moving in this direction. Could you explain the difference between wave-function collapse and the multiverse, and where the multiverse comes from?
David Deutsch
I think the easiest route into this subject is to stay with quantum computers for a moment. Saying that they perform different computations in different universes is shorthand for something not quite so bizarre: quantum theory’s description of subatomic events is exponentially more complex than a classical description.
In the world we see, we can describe one particle here and another there, assign numbers specifying where each is, and apply classical laws to determine where they will go. A classical computer works that way; it makes particles move around.
The quantum description is exponentially more complex. Most of the time, that extra complexity is washed out and we cannot see its effects. Inside a quantum computer, however, it resolves into autonomous strands of events. Each strand proceeds as if the others were absent. There may be two to the power of one hundred separate computations, each independent of the others.
Lucas di Grassi
That is the parallelism of the computation.
David Deutsch
Yes, but it is not like a classical parallel computer, where you need a hundred computers to do a hundred different things. This is one computer doing two to the power of one hundred different things simultaneously.
Lucas di Grassi
Where are these computations actually happening?
David Deutsch
Inside the computer.
Lucas di Grassi
I know, but what about the computations we do not observe?
David Deutsch
Let’s begin with the computations that are happening. Two to the power of one hundred of them occur inside the computer until you look at it. If you measure it, perhaps by sending in a laser pulse that detects what is happening, the complexity washes out again.
I skipped an important step. Those computations would be useless unless we could bring them together. If a question requires information from all of them, such as whether there are more results of one kind than another, we perform a quantum operation that combines their contributions and places the answer in every strand. That is the result of the quantum computation. Without that final operation, we would observe only one randomly selected result, no better than carrying out a single computation.
Now look more closely at measurement. A laser enters the computer, bounces off, reaches a detector, and photons from the detector enter your eye. The entire process remains highly complex. The different parts simply cease affecting one another. Since they no longer interact, you can observe at most one. From your perspective, it is as though the others ceased to exist. But you cannot say that of the computations while the computer operates, because they all contribute to the answer.
Lucas di Grassi
That means the observer becomes entangled with the measurement.
David Deutsch
Exactly. But it doesn’t need a conscious observer or anything like that.
Lucas di Grassi
No. A conscious observer is where the woo starts. It could be any collection of atoms.
David Deutsch
Yes. Or a laser.
Lucas di Grassi
It could be a laser beam or anything else becoming entangled with that particular quantum state.
David Deutsch
Yes. All the autonomous channels are realized in reality. Because we are part of that multiplicity, if you ask how many worlds we see, we will say one. Similarly, if you carefully ask each computation midway through the process how many computations it is performing, they will all answer in unison, “One.” There are two to the power of one hundred of them, but each performs only one. A quantum computer lets them interact so that a certain function of their results is transferred to every strand. Whichever strand you later observe contains the answer.
Lucas di Grassi
When did you begin theorizing about this process? You were the first, or among the first, to describe quantum computation. How did that begin?
David Deutsch
It’s a slightly odd story because that is not quite the right question. I first theorized about quantum computers and proved that a universal quantum computer exists in theory, whether or not we can build one, just as a universal Turing machine exists. That was in 1984 or 1985.
Lucas di Grassi
When I was born.
David Deutsch
Right. These things take a long time to catch on. But already in 1977 or 1978, I wrote a paper describing a universal quantum computer without realizing what it was. I presented it as a machine for testing the many-universes interpretation of quantum theory. To make the machine work, it needed a computer that operated according to quantum theory. I described how it would work and how it would produce one answer if the wave function collapsed and another if it did not collapse, as Everett said.
I never used the words “quantum computer” because I regarded it as a physical object intended for another purpose, not for computation. Only several years later did I realize its connection with computation in general and with universality.
Lucas di Grassi
Has the scientific consensus moved toward this view? Your work was revolutionary, and I assume it was far from the consensus at the time.
David Deutsch
It was not. But consensus among whom? People who work on quantum computers will typically say that they adopt the Everett interpretation. A few do not, but most do. People who work on lasers or solid-state physics will often say they are not interested in that kind of question. They would rather calculate what happens than understand why it happens. In my opinion, that is not a scientific attitude. It is a very bad attitude and cannot lead to progress. Unfortunately, I think most working physicists would take that view nowadays.
That is not the Copenhagen view, by the way. But the difference between Copenhagen and what they now say is a technical matter that is probably not of interest here.
Lucas di Grassi
Let’s return to more everyday subjects. We’ve discussed universality and how knowledge is created. I want to ask about creativity, AI, and large language models. Can LLMs achieve AGI, or do we need a new technology? How does human creativity work? Before linking this to AI, are you a determinist? Do you believe in free will?
David Deutsch
Yes, I believe in both.
Lucas di Grassi
You believe in both?
David Deutsch
Yes.
Lucas di Grassi
How do determinism, free will, and creativity fit together? Then perhaps we can connect them to artificial intelligence.
David Deutsch
I’m not sure the connection between them is so strong, though I have views on both. People say we do not have free will because we can only do what our atoms make us do. That is a mistake. If you say your thinking brings something new into the world, they reply that it was all determined by the atoms in your body. You merely experience it as new, while it was unconsciously caused by the atoms.
That is a fundamental misunderstanding of the physical world. We are not forced to do things by atoms. Atoms simply do things. They always behave deterministically, and sometimes what they do together expresses free will. Those are the events that create new knowledge.
Lucas di Grassi
How does an atom do something involving free will? Sorry.
David Deutsch
Atoms do not have free will.
Lucas di Grassi
But you just said that atoms do things involving free will. What does that mean?
David Deutsch
Together, as in the brain. The brain is made of atoms. That collection can do something radically and fundamentally new, something that cannot be predicted merely from the laws governing atomic motion. It is not just too difficult to predict; you cannot predict it even in principle.
If you thought otherwise, you would be led to ridiculous claims such as, “Mathematics is not the study of abstract objects; it is merely the study of mathematicians’ brains.” Some people do say that, but it is untenable. There is a reason a mathematician prefers a valid proof to an invalid one. He cannot wake up on Wednesday and decide that from now on he will prefer invalid proofs, because the mathematical object he is thinking about is autonomous.
I have now used the word “autonomous” twice: once for quantum computations and once for collections of atoms creating new knowledge. The two are linked. We are not forced to do things by atoms. Atoms are not forced to do things either; they simply do them. One thing they can do together is create new knowledge, and that is fundamentally unpredictable. Randomness is also unpredictable, but it is not the same thing.
Lucas di Grassi
Does unpredictability necessarily mean free will in this sense?
David Deutsch
It is the other way around. Free will necessarily requires unpredictability. But unpredictability is not enough for free will. Unpredictability could simply give randomness.
Lucas di Grassi
Consider a neuron that fires an electrical signal, or does not, depending on its chemistry and the electrical gradient across it. Where can free will enter? How does it arise from such a process? This has always intrigued me.
David Deutsch
If I knew the exact answer, I could make an AGI. I don’t, but I know many bad answers and why they are bad. In the brain, many things go wrong all the time, just as they do in a government. What makes the brain a knowledge-creating system rather than a random collection of atoms is its capacity for error correction.
Suppose something in your brain momentarily reversed your memory of left and right. You would do a double take because many other error-correcting processes contain knowledge about left and right. An enormous torrent of error correction is occurring in your brain all the time. You suddenly think, “Wait, that isn’t left; this is left,” and correct the error before you are even conscious of it.
This error correction allows molecular indeterminism or randomness to contribute to creativity at a higher level. It works because of higher-level error correction.
Lucas di Grassi
Interesting. I had never thought of it that way. By analogy with AI, if we created the right structure of error correction among its components, could an AI become creative in the same way?
David Deutsch
Then it would be an AGI. We do not know how to do that, but I hope we will one day. Then we will be able to create AGIs. They will be people, with rights and property, integrated into society.
Lucas di Grassi
Do you think achieving AGI will require a new technology centered on self-correction among its nodes or structures?
David Deutsch
The self-correction that happens in an AGI is almost the opposite of what happens in an AI. An AGI has free will and can decide for itself. When building a conventional AI, you try to make it meet specified criteria as closely as possible. Ask for the best chess move and it must produce the best move. Ask whether a picture shows a cat or a dog and it must usually give the right answer.
Increasing the number of tasks it can perform in that way takes you further from AGI, not closer, because an AGI does not have to play chess if it does not want to. An AGI that plays chess could one day say, “I don’t want to play chess anymore. Put me in a robot and I’ll play tennis.” It could also remain silent. Like a human, it could decide that it dislikes society and wants to be alone.
Lucas di Grassi
The AGI you’re describing sounds conscious as well.
David Deutsch
Yes. The “general” means that it must be capable of consciousness, among other things. Otherwise it would not be general.
Lucas di Grassi
Could a general machine create knowledge without being conscious, or is consciousness necessary?
David Deutsch
I don’t know. My strong guess is that consciousness is necessary, though I do not have a theory proving it. Free will, consciousness, qualia, creativity, and the capacity to do science all evolved together in a remarkably short evolutionary period. At roughly the same time that humans could invent campfires, they could think about the sky, invent gods and cultures, and create all sorts of new ideas because they also had free will.
I cannot prove that these five or six capacities could not occur separately, which would raise several important problems if they could. But I think they cannot. A general AGI would be capable of all of them, even if it did not exercise every capacity. For example, it could decide not to exercise its free will.
Lucas di Grassi
Roger Penrose often links consciousness to quantum processes in the brain. I won’t go into the details because I do not fully understand them. Do you think creating AGI through the process you describe will require quantum computation or quantum operations?
David Deutsch
No. I think AGI can be achieved with a classical computer.
Lucas di Grassi
Okay.
David Deutsch
I also think the brain is a classical computer in the relevant sense. Its thermodynamics, which provides the lowest level of error correction, is already classical. Many further levels of error correction lie above that before we reach the conscious or creative level.
Lucas di Grassi
To finish, I have a few questions that I use to understand the person better. First, are you religious?
David Deutsch
I’m an atheist.
Lucas di Grassi
I’m an atheist too, so that is straightforward. My second question is whether the search for truth is more important than the search for happiness. I have often discussed with friends whether ignorance of something stressful can make life easier. You have spent your life trying to understand truth at the deepest level of reality. Has that made you happier, or might not knowing certain things make life easier?
David Deutsch
I don’t think the two are separable. You cannot have a conception of happiness that is independent of problem solving. It is not so much truth, because we never reach the truth.
Lucas di Grassi
Yes, truth in the sense you just described: problem solving and the next approximation to truth.
David Deutsch
Exactly. Problem solving is inseparable from happiness. If you seek happiness in the abstract without trying to solve problems, you will fail, no matter how high a priority you give it. We cannot define happiness independently of problem solving because people enjoy activities that others absolutely hate. Most people, for example, would hate to race in Formula One.
Lucas di Grassi
Yes.
David Deutsch
They would find it terrifying and sacrifice a great deal to avoid it. Yet for others it is the prime source of joy and happiness in life. This shows why putting happiness on one side and truth on the other is a mistake. Our attitudes toward truth and happiness are both determined by our attitude toward problems: whether we are solving them, hope to solve them, or think we can solve them.
We are happiest while solving problems. When illness, bereavement, or something else makes you unhappy, you are trying to solve the problem and reach a state of mind in which you are happier.
Lucas di Grassi
Following that thought, suppose a government believes it has knowledge that some citizens lack. Does it have the right to decide what those people may or may not do if it claims to be acting for the common good?
David Deutsch
I don’t believe there is such a thing as the common good. There are only problems. At any particular time, we have laws, ways of enforcing them, and ways of changing them. All are imperfect. If those imperfections are not corrected, they will grow and destroy the system.
It is also wrong to think of government as possessing more knowledge. Its members may know more about particular scientific theories, or they can hire people who do, but I side with Popper on this. Some of us may know a little more about one thing or another, but in our infinite ignorance we are all the same.
That is the important point. Differences in knowledge should not determine how we interact. What should determine it are institutions that enable error correction, because everyone is infinitely ignorant: governments and citizens, parents and eight-year-olds. Each person possesses small pieces of knowledge that others lack. We all try to solve problems and build knowledge so that we can solve further problems and become happier.
Lucas di Grassi
What do you think people should understand or study more? If you could shape education, what would you recommend? How can people turn the ideas of decision-making and error correction into a practical approach to life?
David Deutsch
They need institutions within a single mind, a family, a subculture, a culture, a government, and the world. All of these need nonviolent error correction, with no particular theory entrenched beyond revision.
This is an ideal. In reality, everything is imperfect and has too little of this property. But the more a mind, family, culture, or government has it, the better it can create knowledge, solve problems, and move toward greater happiness. The more its institutions thwart or sabotage error correction, the more they obstruct problem solving and the pursuit of happiness. Always happier; there is no final state called happy.
Lucas di Grassi
Happier, yes. My last question, I promise. Are you excited about humanity’s future? I hear parents and others say, “I won’t have children because humanity is doomed. AI will eliminate all the jobs.” You have spent decades studying knowledge, progress, and how information is processed. Are you optimistic about our future?
David Deutsch
I cannot predict humanity’s future because it is the most unpredictable thing in the universe. It is affected more than anything else by the growth of knowledge, or by the prevention of that growth.
I can say that the present society is by far the best that has ever existed on this planet. It is not a close-run thing. What happens next depends on the choices people make and the knowledge they create. I do not think AI or LLM technology poses any special threat to humans. It creates problems, of course, just as computers and the internet do.
Some people have died after becoming victims of internet scams. In that sense, the invention of the internet killed them. But many more have been saved by the internet and would otherwise be dead. That is true of every technology. Even atom bombs have, up to now, prevented more deaths than they have caused.
Lucas di Grassi
Yes.
David Deutsch
Atom bombs are a standard example of supposedly bad technology, but technology is not bad in itself. Everything depends on how it is used. Atom bombs could have been used to destroy the human species, but they were not, and they still have not been. Whether they will be tomorrow depends on our decisions.
We are decision-making, knowledge-creating, problem-solving animals. I argue in the book that the problems arising from AI are inherently soluble. An act of will cannot guarantee that we solve one on a particular occasion. But if we fail, it is not because the problem was insoluble. It is because we did not find the right idea in time.
As Popper says, we must be optimistic in the sense that we regard problems as soluble and our task as removing the obstacles to solving them. Problems will certainly arise. When AGI appears, completely different problems will arise. People who think a little more AI will produce AGI misconstrue the problems we will face when genuine AGIs exist. What rights will they have? Will they vote? May you unplug one, or would that be murder?
We are nowhere near that. Of course, I cannot predict that someone will not discover how to make an AGI tomorrow. They might suddenly say, “Wait. We should not be doing this like LLMs. We should be doing this instead.”
Lucas di Grassi
It is a problem that may eventually be solved.
David Deutsch
It could be solved, yes. I hope it will be, and I hope I am still alive when it is. But that will not be the end of problems. There will simply be more problems and better problems.
Lucas di Grassi
Very good. That is a great place to finish. David, it has been an honor. Thank you very much for this hour-and-a-half conversation. In my opinion, you are one of the brightest minds in the world today.
I recommend your two books to anybody interested in almost anything. As we have discussed, your ideas apply to children and governments, to education, to how we see the world, and to how we operate in a company. When you understand their first principles and the explanations behind them, they can change how you approach life and help you solve problems differently and perhaps faster.
You have contributed a great deal to humanity, and you are a true legend. Thank you for this conversation. I lived in Oxford for two years and have wonderful memories of the city and the UK. I wish you well, and I hope we can meet in person one day. You are one of a kind and deserve this recognition.
David Deutsch
Thank you for all that. It was a very pleasant and interesting conversation.
Lucas di Grassi
Thank you very much. I appreciate it.
David Deutsch
Okay.
Lucas di Grassi
Bye-bye.
David Deutsch
Bye-bye then.
Markdown