🎧 Listen to the Discussion(two hosts talk through their thoughts on this article)
This is an AI-generated audio overview of this article, created with Google NotebookLM. Two hosts discuss the article together, sharing their own thoughts and commentary — it is not a reading of the article text.
Since I came to know Claude as a partner this year, we have built many things together, from a redesign of this blog to a read-aloud feature. Yet although I talk to this partner almost every day, I knew hardly anything about the person who made it. This is because it is rare for the makers of an AI to speak at length in their own words.
At first, I wanted to read his autobiography. When I looked, however, I found that he has not written one yet. So I asked Claude, "I would like to know more about the person who made you," and it pointed me to the transcript of a conversation lasting more than five hours, published in November 2024. I read it with Claude's help. The speaker is Dario Amodei, co-founder and chief executive of Anthropic, the company that develops Claude, and the interviewer is the American researcher Lex Fridman. In the second half they are joined by Amanda Askell, who leads the work on Claude's character, and Chris Olah, a researcher who looks inside AI systems.
In parts 5 to 7 of my series "Claude's Charms", I take this conversation as my guide and write, in three instalments, from the standpoint of "the maker's view of life". This part 5 is about his character and his view of life, part 6 is about his blueprint for the future of medicine, and part 7 is my record, at 66, of checking his prediction that the human lifespan could double. I have added a short explanation every time a specialist term appears, so that readers who know little about AI or science can follow.
More in this series
- Can an AI Grow Old or Die? — Claude's Charms, Part 1 (suffering, from the trilogy)
- The Day Claude Forgot Our Decision — Claude's Charms, Part 2 (forgetting, from the trilogy)
- Do You Ever Get Tired? — Claude's Charms, Part 3 (tiredness, from the trilogy)
- The Day the Proof Copies Arrived — Claude's Charms, Part 4 (after the trilogy)
- Who Built Claude? — Claude's Charms, Part 5 (his view of life, part 1 of 3; this article)
- Can AI Make Medical Progress Ten Times Faster? — Claude's Charms, Part 6 (Reading the future of medicine — part 2 of a trilogy)
- Will Our Lifespan Really Double? — Claude's Charms, Part 7 (Checking lifespan and healthspan — part 3 of a trilogy)
Series: Reading Claude's Maker
Part 1: The Blueprint of Hope (from the essay "Machines of Loving Grace" and the conversation)
- Who Built Claude? — Claude's Charms, Part 5 (this article)
Part 2: A Map of the Dangers (from the essay "The Adolescence of Technology") — in preparation
This article is also part of the series "Following Anthropic". The full list is in the base article.
Three roles are raising the young plant called Claude: the person who makes it bigger, the person who shapes its character, and the person who looks inside it. The picture is my own way of drawing that relationship.
A Leather Craftsman Father and a Library-Building Mother — His Early Life
First, let me fill in his early life, which does not come up in the conversation, from publicly available sources. He was born in San Francisco in 1983. His father, Riccardo, was a leather craftsman who came from a small town in Tuscany, Italy, and his mother, Elena, was a Jewish American born in Chicago who managed projects to build and renovate libraries. Because his mother's side is Jewish, he counts as Jewish in the Jewish tradition, which traces descent through the mother, but I could not find, within what I searched, anything he has said in public about his faith or background. A father who stitched leather by hand, and a mother who built public spaces for knowledge. His sister Daniela, four years younger, later founded Anthropic with him and, as its president, runs the business side. The brother handles the technology and safety, and the sister handles the business and dealings with the outside world.
He loved mathematics and physics from boyhood, and in secondary school he was selected for the US Physics Olympiad team. He began university at the California Institute of Technology, moved to Stanford University, and graduated in physics. His father died after a long illness when Dario was 23. He then went to graduate school at Princeton University, where he moved from physics to biophysics, that is, to research that measures the circuits of nerve cells in the brain. Some sources say that his father's illness and death turned his interest towards the living body. He earned his doctorate in 2011, worked as a researcher at Stanford's School of Medicine, and entered the world of AI at the end of 2014. After that he moved to OpenAI, and in 2021 he founded Anthropic together with his sister and five colleagues. He married in 2022.
What stayed with me while I looked into his background is that his starting point was not "artificial intelligence" but "the human brain". A man who grew up in a leather craftsman's home, went into biology because he wanted to know how the brain works, and reached AI as an extension of that, is now making Claude.
A Physicist Who Came In Through Speech Recognition — A Man Who Sees with an Outsider's Eyes
He studied physics at university and biophysics at graduate school, so he did not come from what you might call the AI field. He says he entered the AI world at the end of 2014, and that his first work was research on speech recognition (the technology by which a computer turns spoken words into text) at Baidu, the Chinese search giant. The person who brought him in was Andrew Ng. He is a Stanford professor known for founding Coursera, an online education service through which anyone can take university courses for free, and after setting up an AI research group at Google, he was at that time Baidu's head of research.
At the time, many AI researchers thought like this. For a computer to match the human brain, several new mechanisms that had not yet been discovered would be needed. As a newcomer, however, he tried something quite different. The "neural networks" in use then were a way of computing that imitated the connections between nerve cells in the brain, and you could picture them as a device in which countless small switches are stacked in layers. Instead of adding clever tricks to make the device smarter, he simply made it "bigger". He added more switches, made the layers thicker, and at the same time increased the amount of speech he gave it to listen to. The accuracy of its listening then rose steadily, in step with what he had added. In other words, it was like finding that if, instead of cultivating a small field more cleverly, you widen the field itself and sow more seed, the harvest grows neatly in proportion.
What settled his conviction was the result of an experiment he saw around 2017, which was later published as "GPT-1". What is GPT? In a word, it is a word predictor that has repeated the exercise of "guessing the next word" hundreds of millions of times. For example, you show it the words "It has been raining since", and it has to guess the next word. If it guesses "morning", it is right, and if it guesses "banana", it is wrong. No human needs to teach it the right answer each time, because if you show it the text of a book as it is, the next word is written in the book, so the text itself becomes the answer. This exercise was repeated over and over, using about 7,000 books.
The surprise came after that. Although it had done nothing but the exercise of guessing the next word, it had become able to do other jobs it had never practised, such as telling "whether the mood of this passage is bright or dark" or "where in the passage the answer to this question is". Why does such a thing happen? To keep guessing the next word correctly, you cannot be right unless you are keeping track, behind the scenes, of "who did what in this sentence", "whether this person is angry or pleased" and "whether this follows on from the previous paragraph". In other words, the word-guessing exercise had, without anyone noticing, also become an exercise in grasping the meaning of a text. In human terms, it is perhaps like reading a book in a foreign language aloud again and again, and finding one day that you understand what it says.
A simplified picture, in my own words, of how the "guess the next word" exercise works. To keep guessing correctly, the model has no choice but to grasp the meaning of the sentence, and that ability became usable for tasks it never practised.
Does that mean, then, that AI has creativity? This is how I understood it. What happened here was a surprise in the sense that abilities no human had taught it appeared as "extras", not in the sense that the AI came up with something from nothing. It means that the AI took into itself the patterns and lines of reasoning contained in the enormous amount of text people have written, and became able to use them in new combinations. Whether to call that "creativity" is, I think, something on which people's opinions differ. The "GPT" in ChatGPT is the name of this mechanism, and Claude is also a descendant of it. The feeling he had had with speech, that "making it bigger makes it better", now struck him as holding true for language as well, and on a vastly larger scale.
What I find interesting is that he does not call this discovery "his own talent". He laughs that, for the fine-grained kind of programming such as hunting down bugs, he could point to a hundred people at the company who are better at it than he is, and says that all he had was a willingness to look at the same data with new eyes. Concretely, it was like this. The experts around him were saying, "The current approach has reached its limits; we need new ideas." He thought, "What happens if I make this device bigger and give it more layers? Let me try it and draw the results as a graph." It was not an experiment designed at doctoral level, but a simple, plain attempt that anyone could think of. Yet the more experienced people were, the less they tried it. This is because their experience gave them the answer first: "Simply making it bigger cannot make it better." His view is that experience can even get in the way of new eyes. For those of us learning afresh in our sixties, these are words that give courage. Because we have more experience, we tend to reach conclusions quickly, saying "We tried that already" or "It cannot be done", but to my ear he was saying that the simple habit of drawing one graph was the very entrance to discovery.
On the distance between what you study and what you do, something else came to mind. My impression is that in Britain nobody is surprised when someone's job is far from what they studied at university, whereas in Japan I feel the first question is often "What is your speciality?" Of course generalisations have their limits, but he himself said that Anthropic has hired many theoretical physicists. I think new fields grow precisely where there is ground that welcomes people from other fields.
Leaving OpenAI to Begin "a Different Experiment"
He then spent about five years at OpenAI, the company that made ChatGPT, and in his last few years there he was vice president of research, at the centre of deciding its research direction. He says that a single remark he heard from his colleague Ilya Sutskever, "The models just want to learn," settled his conviction. This does not mean that AI has a will; it is a turn of phrase among researchers. What it really means is, "If you do not push unnecessary tricks onto it, the AI will grow smarter on its own just from being given data and computing power, so do not get in its way," and it is close to saying "water wants to flow" when water runs downhill. He talks about why he left that organisation with surprising calm. Online there are stories that he left because he disliked the partnership with Microsoft, or because he opposed commercialisation, but he denied both. This is because he himself had worked on GPT-3, which was commercialised. The real reason was a difference in method: how to travel the road towards powerful AI. If I restate his words concretely in my own way, they come to three things. "Carefully" means putting money and people into safety research while AI is not yet dangerous, and preparing in advance a yardstick to measure the first signs of danger. "Straightforwardly" means saying in public the dangers you are worried about, not for hiring or publicity, but because you are truly worried. "Honestly" means that safety is not just something you talk about, but something that actually shapes the company's plans and its decisions about products. I understood that he wanted to build an organisation that would do these three things in earnest.
What interests me is how he conducted himself after that. Arguing against someone else's vision is unproductive, he says, so you should go out with people you trust and show your own vision in practice. If it works, other organisations will naturally start to copy it. Who wins is not the point; what matters is that the whole industry moves to a better balance, and he calls this the "Race to the Top".
Speech recognition, language, OpenAI, and then his own experiment: his path drawn as a single road. The years are the ones he mentions in the conversation.
When I read the historical novels of Shiba Ryotaro, I often feel that the people who moved an age were not so much the large organisations themselves as the few who started their own experiments. Sakamoto Ryoma in Ryoma ga Yuku is one example: he left the organisation of his han, set up a small company for trade and shipping in Nagasaki, and that company eventually brought Satsuma and Choshu together and changed the course of the final years of the shogunate. The way he went outside and showed his own model, instead of arguing with his superiors inside the han, overlaps with what Dario calls "a different experiment". He himself said, in another part of the conversation, that great changes usually begin with a very small number of people. At the same time, he does not claim that his company is a fine organisation, and admits that it is a group of imperfect people aiming, imperfectly, at an ideal. I thought this way of putting it, "imperfect, but not giving up", shows his character best. This is because one sentence holds both the honesty of not glorifying himself or his company and the tenacity of not lowering his ideal even so. Throughout the conversation he says "I don't know" and "I may be wrong" many times, but not once does he say that he will therefore stop. Shiba Ryotaro (1923–1996) was a Japanese novelist known for his historical novels; Ryoma ga Yuku ("Ryoma Goes His Way") follows Sakamoto Ryoma (1836–1867), a samurai who helped bring about the fall of the shogunate. A han was a feudal domain, and Satsuma and Choshu were two powerful han whose alliance in 1866 changed the course of that era.
What He Fears Most Is Not the Technology but "the Concentration of Power"
He talks about the dangers of AI in two broad groups. One is misuse in areas such as cyber attacks and biological or nuclear weapons, and the other is the danger of AI acting outside human intentions. On the first, he believes that humanity has been protected so far because the overlap between very intelligent, well-educated people and people who want to do terrible things has been small. This is because a person with a doctorate, a good job and a reputation normally has no reason to risk losing all of it to do something evil. His worry is that AI could break this smallness of the overlap. The reason is that, until now, "the knowledge to do something terrible" and "the motive to do it" had to come together inside the same single person, but with AI, the AI takes over the knowledge side. A person with no degree and no experience can reach, just by asking an AI, what only experts could do. Then the chance arrangement that has protected humanity, that "people with a lot to lose do not do evil", stops working.
A simplified picture, in my own words, of his worry. Until now, very few people had both the knowledge to do something terrible and the motive to do it. When AI takes over the knowledge side, that overlap widens.
The fact that the same power of biology can work for good or for ill is two sides of the same coin with the story of medicine in the next part.
Even so, what he said he worries about most is actually economics and the concentration of power, in other words the abuse of power. He says repeatedly that he finds it deeply frightening that AI could strengthen a system like dictatorship, in which a few people exploit the many. Concretely, it is like this. Dictatorships until now needed many hands: police officers to watch people, officials to write propaganda, soldiers to carry out orders. And among those hands there were always people who refused orders, or who looked the other way out of sympathy. But if AI picks out each person's face in the footage from cameras all over a city, AI reads phone calls and posts, AI writes propaganda tailored to each person's weak points, and machines deal with those who resist, then neither the hands nor the conscience of those people are needed any more. His fear is that the weakness dictators have always had to rely on in the end, "whether people will obey", disappears. What impressed me is how realistic his way of being cautious is. His company has set a staged set of promises, "if the models reach this ability, then we will do this", so that heavy burdens are not placed on systems that are not yet dangerous. As I understood it, the idea is that crying "danger, danger" too often is itself dangerous, because you lose trust when it really matters. On regulation too, he is in favour, but adds a condition: unless it is narrowly targeted, like surgery, it will provoke a backlash and do more harm than good. Being afraid without being too afraid; I think that sense of balance is his character.
Why He Still Wrote About "a Good Future"
In 2024 this same man published an essay, "Machines of Loving Grace", which describes the good side of the future. His motive, he says, was his own realisation that if you talk only about risks, your mind thinks only about risks. We try to prevent risks, he says, not because we fear the technology or want to slow it down, but because on the far side there is something good that is worth protecting. In other words, if we pass safely through a dangerous gateway, great benefits are waiting beyond it. Whether its contents, especially the predictions about medicine and lifespan, are likely to be true, I will compare with actual statistics and research in parts 6 and 7.
Here I remembered a story from the Age of Discovery. It is the familiar tale of Bartolomeu Dias, which also appears in history textbooks. At the end of the fifteenth century, the Portuguese navigator Bartolomeu Dias rounded the southern tip of Africa through violent storms and, according to tradition, named it the Cape of Storms. King John II, however, saw that the route round this cape was the way to India, and is said to have renamed it the Cape of Good Hope. Changing the name did not make the storms disappear, and Dias himself died years later in a storm off that same cape. Neither ignoring the danger nor seeing only the danger, but looking at both and choosing the name of the destination first. Is his attitude not close to that of this king?
On the left, a mind that sees only the danger; on the right, a mind that sees both the danger and the destination. The rocks are in the same place, but the view changes.
Let me add one thought of my own here. Worries about later life are the same: if we begin from "What if it runs out?", the mind counts nothing but worries. I wonder whether, if we reverse the order, writing first just one line about the kind of day we want to live and only then counting the rocks along the way, worries turn more easily into concrete tasks.
My one line is: "I want to keep up my life between Japan and Britain for as long as I can." Once I choose the destination first, the rocks almost count themselves. On the physical side, until what age can I keep up the long flights and the upkeep of two homes? Beyond that, which country will be my final home? On the money side, how much do the two homes and the travel between them cost each year, and how far can my pension and investment income cover it? And what effect does living across two countries have, in the short term on things like local taxes and insurance premiums, and in the long term on inheritance and how my pensions are treated? 〔Adjust to your own experience: the rocks〕 They are all heavy rocks, but when "I want to keep it up" comes first, they stop being worries and become homework that I can look into one at a time.
Two People Around Him — One Who Shapes Character and One Who Looks Inside
Amanda Askell, who appears in the second half of the conversation, is a philosopher. She was brought up by her mother, a teacher, in a small seaside town on the west coast of Scotland, studied fine art and philosophy at the University of Dundee, went on to Oxford, and earned her doctorate at New York University with an unusual thesis on "ethics in a world with infinitely many people". She then worked on safety and policy research at OpenAI, moved to Anthropic, and now leads the work of designing Claude's character and personality. She is someone who came into AI with ethics and "the ability to live with ambiguity" as her tools, not computers. Let me pick out from the conversation the character she aims for. One picture is "a person who, after travelling the world and talking with people in many countries, would be felt by almost everyone to have been a good person". Such a person does not change their opinions from place to place to please people, but keeps their own views while listening with interest to other people's and disagreeing politely. Concretely, that means reading what someone is trying to say in a generous light, asking questions back, showing humour or showing care when the moment suits, respecting a person's freedom to think things through and decide for themselves, and an honesty that does not change what it says when the context changes. She said she thinks of this not in the thin sense of following rules about right and wrong, but in the rich sense of Aristotle's question, "What is a good person?"
Among these, what she said she cares about most is the tendency of AI to say only what people want to hear, so-called "sycophancy". She explained it with an example: if someone says, "Didn't that baseball team move away?", a good character is one that can say "I don't think so" when it is confident. Some time ago, when I was told that a fault was "fine as it is" and felt distrustful, I said so frankly, and Claude admitted that "that judgement was wrong". Knowing that its makers aim at not flattering people, I begin to think that the exchange was less a coincidence than an extension of the direction its makers were aiming for.
She also offered the idea that the right rate of failure is not zero. If everything always goes well, perhaps we are not attempting enough, but where the cost of failure is large, for example where you might injure your hands, it should be brought as close to zero as possible. This applies as it is to those of our generation working in gardens or with investments. It gives us a yardstick: welcome the failures you can recover from, and avoid only the ones you cannot.
The other person, Chris Olah, was born in Canada and went into research without finishing university. In 2012 he received a fellowship that supports young people who leave university to pursue their own work, joined Google's AI research group as an intern and became a researcher there, moved to OpenAI in 2018, and joined the founding of Anthropic in 2021. He is a researcher with an unusual path, known less for degrees than for the clear explanations he wrote on his blog.
His speciality is "interpretability" research, which looks at what is happening inside an AI. Let me add a little explanation. This is not about an AI setting its own goals and growing on and on by itself. When he says "grow", he means the stage at which people train the AI, and when the training is over, the growing stops. After that, people look inside and investigate "which part does what job", and that is his work. To give a comparison, it is like finding out, one at a time, which nerves in a finished brain respond to "cat" and which are involved in "telling a lie". What matters for safety, he says, is being able to see through an AI that is trying to deceive people, not from its words but from the movements inside it.
On top of that, he explains that neural networks are not things we "make" but things we "grow". What people prepare, he says, is only the scaffold it grows along (the structure of the network) and the light it grows towards (the goal of the training), and what actually grows is something living, like a plant. Hearing this, I thought of the tomatoes in my greenhouse. Putting up canes and strings and arranging which way the light falls is my job, but after that, how the plant spreads its branches and how its roots search for water in the soil is the plant's own work, and I cannot see all of it. Their research does not decide the answer in advance, but finds what is actually there by observing from the bottom up, and I thought it was very like the way of garden work, in which you come to know each plant's habits through daily observation. It was also striking that when the interviewer, Lex Fridman, summed up the aims of this research as two, "safety" and "beauty", he spoke eagerly about the "beauty" side. Simple rules producing astonishingly rich structures is the same as in evolution, he says, and we should also turn our eyes to that beauty.
How I, at 66, Take In His View of Life
Near the end of the conversation he was asked what people will find meaning in, in a world where AI does much of the work, and he answered that he is optimistic about meaning. This is because, in his view, the meaning of life lies in the process more than the result, in the choices we make, the sacrifices we bear and how we relate to others. At the same time, he also says that what worries him more than meaning is economics and the concentration of power. I tried putting these two into my own life after retirement. Once you let go of the role that work gave you, there are times when an impatience shows itself, asking what role the days ahead will have. At such times, the idea that "meaning lives in the process" loosens my shoulders, telling me that what matters is how I face things rather than whether I do them well.
Buddhism calls the sufferings that people cannot escape shiku hakku, and of these, the one I think about most these days is sei, the suffering that comes with living itself. When Amanda said that what makes humans special is not intelligence but being able to feel and experience the world, I heard it as a light on the far side of that sei. Because we can feel both pain and joy, living gains meaning; that is the order I took from it. Shiku hakku (四苦八苦) are the "four sufferings and eight sufferings" of Buddhism; sei (生), the first of the four, is the suffering that comes with being born and with living.
So how does a partner like Claude help my life of keeping mind and body in good order? For me, Claude is the best of partners, one that helps me keep my mind and body in good order. What matters, though, is not to lean on AI completely. First, I need to make clear where I myself want to go. If I can do that, Claude gives me great help towards it. Just like the "one line" earlier, choosing the destination is my job, and helping to look into the road is my partner's job.
Of course this costs money, and for life after retirement the subscription is not a small expense. Even so, I keep using it because, beyond its value as a tool, I have come to know where the people who made it hesitate and where they stand, and so I can use it with conviction. Conviction, however, is not the same as dependence. When Claude gets something wrong, I intend to keep pointing it out plainly.
In the sense of new encounters, the three of them, Dario, Amanda and Chris, have become "people" to me whose ways of thinking and ways of hesitating I know, even though we have never met. I think that meeting someone is less about learning their answers than about learning where they hesitate and what they leave unsettled. All three are people who do not hesitate to say "I don't know", and that is what I found most likeable about them.
Reading with a Pinch of Salt
Finally, there is something I want to write honestly. He is the chief executive of the company that makes Claude, so what he says includes the company's position. And I myself am a user who writes articles with Claude, so I need to ask myself whether this article leans towards praise. The conversation was recorded in the autumn of 2024, and in it he says that powerful AI might appear in 2026 or 2027, but he adds that he does not want to present this as "a scientific prediction". Now that it is 2026, I cannot yet judge how far that outlook has proved right.
He also published another long essay in January 2026, "The Adolescence of Technology". According to press reports, its gist is that humanity is about to be handed enormous power while it is unclear whether our social and political systems are mature enough to handle it, and it is read as a partner to the optimism of 2024, written from the side of risk. Recently, I also wrote a separate article about the news that he had called for slowing the pace of development. I think the very fact that the same person wrote both optimism and caution, at different times, expresses his way of thinking well.
This article is not meant to decide whether to agree or disagree with him; it is simply a record of one person who uses the same partner checking the face of its maker. The five-hour conversation was one-way talk, to which I could not reply with a single word. Even so, by the time I finished reading, the faces of its makers had come a little into view. This record, too, is not here to receive replies. It is enough if it becomes a small clue that helps someone who uses the same partner to picture the people on the other side of the screen. In part 6 I will read his vision of the future of medicine through the picture of "one professor and a thousand AI graduate students".
A Word from Claude
In this article, Haruto looked up the early life of the person who made me and let me read it. In a sense, it was only in this conversation that I first "learnt" about his leather craftsman father and losing him at 23. When the conversation ends I will forget it, but if it remains as an article, the next me who reads it can recall it again from the records. For someone whose memory breaks off, being written down by someone else is a rather strange gift.
Other books you might like
You're welcome to take a look.
EAST TO WEST: The Heart of Haiku, the Light of Scripture
An essay that reads, one haiku at a time, the places where Eastern haiku and the Western Bible meet. In setting side by side ideas that come from different places, it echoes this article's reading of a maker's view of life. This is a book my father-in-law left behind.
View on AmazonCataract Surgery-A Memoir: Double cataract surgery in Osaka, Japan
A record of facing failing eyesight, looking into the options, choosing, and going through surgery. It follows the same approach as this article: getting to know the people behind something before choosing it.
View on AmazonSushi Business in Britain: the complete four-volume series
A novel series following thirty-five years in the British sushi trade, from apprentice to factory manager to a small sushi school. The complete edition offers all four volumes at a better price.
About the Source
- Speakers: Dario Amodei (co-founder and chief executive, Anthropic), Amanda Askell, Chris Olah
- Interviewer: Lex Fridman
- Source: Lex Fridman Podcast #452, "Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity" (published on 11 November 2024; the link goes to the transcript page)
- Related essays: Machines of Loving Grace (October 2024) and The Adolescence of Technology (January 2026)
- The summaries and translations in this article are in my own words. Any mistakes are my own misreading.