[0:00:00] I personally was super freaked out by it. I looked at it and I said, my God, you just scale this thing up another factor of a million and it'll be better than humans at everything. This is the defining, you know, economic and scientific moment of the modern age. We just don't know it yet. This is going to determine the fate of nations and economies and the future of humanity. Like, you know, you peer into the crystal ball, you see the whole future unfurled. I'm Walter Isaacson. Welcome.
[0:00:34] To the builders, my new series of interviews featuring the most consequential minds shaping our future. Today, I'm speaking with Dario and Daniela Amadei. I used to think that some of the concerns about AI were overblown. But then, when certain new models escaped the sandbox, I realized that the people building these systems have got to make sure that they remain safe and that they remain aligned with human values.
[0:01:04] What I want to know about Anthropic is are they sincere? And when they face competition, like when OpenAI says, well, Anthropic won't do that, but we will, will they back down or will they stand true to their principles? Hey, welcome to the builders, y'all. Thank you so much for having us.
[0:01:22] You know, Steve Jobs, whenever he launched a product, always had that street sign of the arts and the sciences or, you know, the humanities and technology standing at that intersection. And when you were growing up, Daniela, you were a flautist. You loved the arts. And Dario, you were always a scientist. Did that help you gel? And do you think that combination is important? Yeah, I think that combination is a lot of what's behind the company, right?
[0:01:50] One theory I have is that running a company is too big a task for any one person. The range of skills that it involves, it is not possible for any single person to have. So you kind of have to assemble.
[0:02:03] A full complement of skills across the top, like 10 or 15 people in the company or something like that. And every very successful company splits it up in different ways and solves the problem in different ways. And I think with the sciences and the humanities thing, you know, Daniela and I kind of, kind of embody that and split that. Yeah. And I would say, you know, it's interesting, like there was an offsite we did in maybe a 28, 23.
[0:02:28] We're like, we sort of were talking about just like, what does it feel like to work at Anthropic? Like, how do we kind of embody what our culture is? And we actually said, you know, it's an unusual blend of science and humanism. And it was only later that I'm like, that also is how you would describe me and Dario. Well, when you were kids, did you know you wanted to work together? Yes. Yes. Yes. I don't think either of us imagined that we would start a company. That's not really how...
[0:02:56] I would say Dario and I were very unusual in that
[0:03:18] Like, again, from a really early age, we just, like, genuinely really liked each other. We were always just, like, best friends, like, since we were really little. And I agree with Dario. I don't think, like, starting a company was, like, ever on our bingo card. Like, we were not the type of... Either of us. Neither of us were particularly, like, company-oriented. I was always really interested in, like, I think I would call myself, like, a classic do-gooder. So I, like, worked in international development. I felt really strongly about just, like...
[0:03:46] Helping people everywhere in the world have like equal access and opportunity. And I became a vegetarian when I was like quite young. It's like the only thing that I did first. Dario did everything first because he's older, but I became a vegetarian and then he became a vegetarian. Yeah, I think I was like three years later or something. Oh, not even. I think it was maybe just, yeah, yeah. You got on board real quick. It seemed convincing to me. Yeah, yeah.
[0:04:10] You know, people sometimes ask me about my books, how would you become a great leader or very creative? And you said something really interesting, Dario, which is not just being a great leader, it's creating a great team. Tell me how you and Daniela divide up the roles and then how you fill it in with others. Yeah, so, you know, I would say the division that we've found that has been...
[0:04:35] Most useful between us, and then I'll talk about the kind of the rest of the team, is I think I can function best when I think about the big picture strategy and kind of the research side of things. Like I'm best at kind of laying out a vision, seeing what comes next. I have found that when I'm freed up to focus entirely on that, you know, that is the thing that has the most value for the company.
[0:05:02] Running the company day to day, recruiting the people, just making sure that the machine doesn't, you know, the wheels don't fall off the machine while I'm trying to figure out what comes next. You know, Daniela does this. And I think this division of labor allows both things to be done to a much higher standard than either of them would be done if we were distracted.
[0:05:28] By having to do the other thing. Our chief people officer who knows both of us really well is in a lot of conversations with us. The way she put it is that Dario makes us ambitious and I make us successful. And I thought that was a really good encapsulation, which is, I think Anthropic could be like a good successful business, but not change the world if we didn't have Dario. And I think with Dario, we would have the perfect vision, but our ability to execute would be.
[0:05:54] It would just be more complicated. And so I think there's something special about A, the fact that we know each other really well. B, it's extremely convenient that our skills are like 100% complementary or maybe not 100%, but I would say close to 100% complementary in this fashion. This is not even to mention the fact that we have five co-founders along with the two of us who also own like incredibly important parts of the business, the culture, like the organization. And again, I think I like...
[0:06:22] It's hard for me to imagine how you would do this with just the two of us or one person. I think it's an incredible strength because the fact that we have seven co-founders, they all have equal stakes in the company, and they all carry the vision. We have inherent trust in all of them. Many people told us when we started Anthropic,
[0:06:45] Seven co-founders, ha ha, that's going to be a disaster. There's this belief that you need one kind of singular source of leadership. In some ways you do, but we've actually managed to make it work. All seven of the co-founders are still at the company, which again is not something that's usually true when you have that many co-founders. So go back to growing up and you have these extraordinary parents and they help.
[0:07:13] Teach you right from wrong. How do they do that? I don't know that it's something that was ever, you know, sometimes I hear people say like, you know, are we, you know, we covered, you know, different forms of philosophical ethics at the dinner table. That was not, that was not how it was done. My mom does have a philosophy degree, but she didn't subject us to talking about Kant or anything like that. I think it was much more like inherent in terms of values, like just the belief that.
[0:07:41] People everywhere in the world are deserving of opportunity and dignity and rights and respect. I think there was something special about the feeling that we could be whoever we wanted to be as long as there was this kind of faith and belief that, hey, that's part of what it means to be a successful human is to be a good person and to be kind to other people. Yeah, it felt implicit. It felt very implicit.
[0:08:06] You know, there wasn't any kind of, you know, explicit like, you know, this is how you should, this is what you should do, or this is how you live in the world. Well, one reason I asked, as you might suspect, is now you've got a really important kid named Claude. And to what extent does that way that you grew up help you, inform you about how you teach Claude to know right from wrong?
[0:08:30] Yeah, you know, we put out a constitution for Claude, I think it was a few months ago. And I think there are some interesting lessons in it, which we've learned from several years of, you know, basically trying to train Claude in successive versions. I think early on...
[0:08:49] We had been imagining that there would be various rules, that you would tell Claude, don't help someone commit a crime, or don't make biological weapons, or whatever. But it turns out that when you make this kind of discrete list of rules,
[0:09:04] You just, the shape of what you're trying to teach the model, I'm speaking a little bit metaphorically here, but it's kind of grounded in what we found technically. You end up with this kind of spiky shape where you're trying to teach the model like 170 different rules that are kind of arbitrary. We found it's better to teach the model some basic.
[0:09:24] Principles, like what it should prioritize, how it should reason about what the right thing to do is, which, you know, of course, I think that's the right way to teach people as well, right? Like, you know, if I just, you know, give someone, you know, a list of rules, thou shalt do this, thou shalt not do this, you know, I don't think that produces as...
[0:09:45] Robust morality, as I say, well, this is how you should think about things. You should learn how to think for yourself. And these are the principles on which you should think, right? Principles of open inquiry, principles of care for other people, principles of fairness and justice, and then everything is derived from that. And that is...
[0:10:09] Frankly, how Claude learns. We find that to be surprisingly effective because that allows Claude to generalize to new situations that it might never have imagined. Danielle, I can't help myself when you said your mom was an expert on Kant. Did you actually wrestle with philosophy and moral philosophy like that? You know, what's interesting is the woman who is...
[0:10:34] Responsible for training Claude's character has a PhD in philosophy. It's unusual, right? You would think, oh, the person that's training, you know, the sort of character of these models must be, you know, a computer scientist or somebody who's studied ML or AI for, you know, many decades. Amanda's technical and she like technically trains the model, but I think it is very felt in...
[0:10:56] Claude's character, when you communicate with it, that there's this kind of experience, something that's very, you know, wise, kind of the best version of the sort of better, the better angel of your nature trying to answer your questions. But there are like, if you kind of poke Claude in these like difficult areas where maybe different philosophical traditions might conflict, I think that's where you especially feel, Amanda, because you'll sort of see like.
[0:11:25] Oh, well, you know, you might want to consider things from like this point of view, but you could also consider it from this point of view. And I do think it's important because as the world
[0:11:36] Daria, you were there at one of the seminal moments of modern times, which is when GPT-1 and GPT-2 suddenly burst on the scene, surprising.
[0:12:03] Perhaps even surprising you, you were head of research at OpenAI. What was that moment like? Yeah, so, you know, I think GPT-2 was the, maybe the first real empirical test.
[0:12:17] of the scaling laws, right? So I wasn't very involved in GPT-1, but I was very involved in the idea that it should be scaled up, that we should acquire new sources of data, more computation, and just really should run a larger scale, more expensive, more ambitious experiment on the idea that if we ran these language models at a larger scale, they would be better at language, they would be more coherent.
[0:12:47] So, you know, it's almost like one of these moments in scientific history where, you know, you're like, you propose a definitive experiment to just make a very strange prediction. And then if the prediction turns out to be correct, you're like, God, we're really on to something. So, you know, I see GPT-2 as a kind of test, the first real test of the scaling laws hypothesis.
[0:13:13] And what I found super interesting was the enormous divergence in reactions to it. Were you kind of surprised? I personally was super freaked out by it. I looked at it and I said,
[0:13:27] My God, you just scale this thing up another factor of a million and it'll be better than humans at everything. This is the defining economic and scientific moment of the modern age. We just don't know it yet. This is going to determine the fate of nations and economies and the future of humanity. You peer into the crystal ball, you see the whole future unfurled.
[0:13:52] And it really scared me. I said, oh my God, what are the implications of that? Is this the right place to be developing it? What's going to happen next? What do we do with this? Who do we tell about this? And maybe 10% of the people around me had exactly the same reaction. And it was often the smartest ones. 90% of the people around me.
[0:14:13] Because GBDU's capabilities in any particular area were not that impressive, right? And it could translate languages.
[0:14:22] And, you know, summarize a document and write a few lines of code here and there. It was the broadness and the generality. But people would look at it and they would say, in any particular area, that doesn't seem so smart. And so many people just shrugged and they just dismissed it. But we just announced that, you know, we're going to be developing drugs for, you know, neglected, rare diseases. Then you get the usual comments from kind of the jaded, you know.
[0:14:47] Biotech world, people, ah, tech people have done this before. The models aren't, you know, here are all the limitations of the models. And I just look at it and I say, it's the same thing over and over again. The limitations of the models, you know, now, six months from now, the things that seemed hard, they're going to be easy. 12 months from now, the things that seemed impossible, the models are going to be doing them regularly.
[0:15:09] It's everything we saw with GPT-2. We're seeing it over and over again in wider and wider swaths of the world. And the world is gradually finally being convinced. I think you were quite polite about it. But I mean, I remember at the time there was some like very well-respected.
[0:15:25] You know, computer scientists and like AI and ML leaders who were like, basically like bullied some of the researchers who worked on it on Twitter and were like, this is stupid. Like, we don't know why you guys are, you know, even talking, this model is so incapable. And it was like, you were sort of a quack if you believed this was like a really significant moment. And that's all been memory hold. Like all the researchers who like, you know, were like, oh, this is, this is pseudoscience. It doesn't make sense. Like.
[0:15:53] All of them are applying for jobs at like the LLM companies. And like, you know, we never say it out loud to any particular one, but like, you know, we look at them and they're like, we remember what you were saying in 2019. You have no credibility on this matter. So Daniela, you were vice president of safety back then at OpenAI. And when you hit GPT-2, I assume you're like Dario, somebody who's almost freaked out about it. Did it freak you out about safety too?
[0:16:23] There wasn't something about GBT2, the model itself, that made us concerned. We didn't look at that model and say, wow, we think this model has cybersecurity implications or something. The model itself was, in some ways, kind of laughably wrong about a lot of things. It was still a very, I would say, harmless model. But I think the thing that Dario mentioned that I also felt was a sense of like, okay,
[0:16:52] The difference between GPT-1 and GPT-2, just like extrapolate that out, right? Like there was something about seeing, it's almost like if you'd never seen a child before and you'd only seen adults and you were kind of like, I don't know what this little thing is that's walking around. Like, I don't know if I should, you know, imagine that it's going to be, you know, capable one day. And then suddenly it starts to walk. You're like, wow, huh, I bet this thing is going to be capable of doing a lot of the things that a grownup could do one day. Like maybe this is where this is going.
[0:17:20] Dario, you say your departure from OpenAI was partly due to a breakdown of trust, an interesting word, with Sam Altman. Explain how you felt there was a breakdown of trust and honesty there. Look, so what I will say is, look, I think at the end of the day, leaving OpenAI and founding Anthropic, I mean, on one hand, I think it ultimately comes down to
[0:17:50] The principles and how we're going to build the technology. That is what matters at the end of the day. And I think over the last five and a half years, we have seen...
[0:18:03] How those principles differ. Anthropic has advocated for sensible regulation and safeguards on the technology. Anthropic has invested in mechanistic interpretability research to understand what's going on inside the models. I think Anthropic has managed to maintain a culture of unity to have the highest retention rates. We haven't had
[0:18:29] A number of the scandals that other AI companies have had. And so I think that's kind of sufficient to describe the difference. But I will say that the moment when I realized that we needed to found our own company, that OpenAI was not the right place for us, look.
[0:18:50] These safety issues are difficult. You can have lots of disagreements on them. I disagree with the co-founders and with lots of people at Anthropic, but the reason we've been able to resolve these disagreements is because they're held in good faith, right? Should we release this model or should we wait? Should we put stronger or weaker biosafeguards on the model? You know, how much of our...
[0:19:14] Should we spend lobbying for safety regulations and what should they look like? These are all good faith disagreements, but they always get resolved because there's always trust that everyone is trying to get to the right answer and everyone is being honest. My point in saying that there's a breakdown of trust is simply to make the abstract point that if you don't believe that about a place where you're working, if you don't believe that the disagreements are being held in good faith.
[0:19:43] And that you're getting the real story from people. How can you possibly hope to resolve them? That's why we left. That's why we founded our own thing. And I think the last five and a half years are themselves just the best proof of how our philosophy differs from that of the other companies in the space. I think if you continue to work.
[0:20:09] For someone who has a different vision than you, it becomes very unclear what you actually stand for. You basically have to account for the mistakes of someone else's vision. Wait a minute, though. Push back a bit here because you talk about differences in values and stuff, but that's fundamentally different than a difference in trust, that you feel that somebody says they have the same values, but they don't and you can't trust them.
[0:20:37] Honesty is one of our values. It's a very important value. It's hard to know what someone's values are if they don't have the value of honesty. And you felt that about Sam? I think the way, at least for me, I would put it is I think the co-founders are an unusually high integrity group of people. And all of us...
[0:21:05] I think shared a very similar vision for how we wanted this technology to go and how we wanted to build it. And at a certain point, you're like, it just makes more sense to go do this as your own organization. How are you going to get someone else to implement or share your vision if they don't naturally do so? No matter how much they say they share it, if that's not reflected in what the company actually does,
[0:21:35] What choice do you have? Yeah. And I think we're, I do think Anthropic is a company that is founded on trust. And I think what I mean by that is like, I think honesty and integrity and how we treat each other and our employees is also reflective of how we want the technology to be built and to go. An example I will give is every two weeks.
[0:22:03] I get up in front of the company and I speak for an hour about what is happening inside the company, what is happening in the AI field, what is happening in the world, how the company thinks strategically. The two ironclad rules of...
[0:22:22] These events, I call them DVQs, Dario Vision Quest. That was not my name. I initially resisted the name because it makes it sound like I'm going off and smoking peyote, but everyone kept calling them that. The name stuck, and so now I'm stuck with it. The two rules of DVQs are that, one, nothing I say is an actual decision. That allows me to just speak freely.
[0:22:50] Um, you know, without worrying that if I, you know, if I briefly mentioned doing X that 20 people will go off and do X is like the CEO said it, we're going to do X. And then the second is that I'm just, I'm totally honest with the company. It's a conversation of trust. It's a relationship of trust. And, and I think that trust is one of our biggest.
[0:23:11] That people believe that the things that I say and Daniela says, you know, you might think that if you have trust, there's less disagreement. Actually, trust allows for more honest, open disagreement and for disagreement to be aired. There's so many times when, you know, we've made a decision that the majority of the company disagrees with. But what has saved us every time is that the attitude we get from people is, this doesn't seem right to me, but...
[0:23:40] We trust that you're doing this for the right reasons. We trust that this is really what you believe. We trust that you have good reasons for thinking this. And so what this means is you can make decisions that would ordinarily be unpopular and people suspend their disbelief, not forever. You eventually have to show that you're making the right bets, whether it's commercially or morally or technically, but you get...
[0:24:07] Kind of a longer leash, like a longer window. And I think other companies don't have that because they've eroded trust over and over again. So the reason I'm so interested in honesty.
[0:24:19] And the feelings that you left OpenAI because you didn't feel the people at the top were truly trustworthy or honest when they were dealing with you is, once again, I'm interested in it because it can be reflected in AI, in Claude. AI trains on humans. It learns from human data. How do you know when you're dealing with Claude that it...
[0:24:45] Won't become manipulative. It won't become like the people at OpenAI you got worried about and say things to you that are dishonest. This is a constant, constant worry that we expend an incredible amount of energy on in how we train the models. So it is a great question because it is a real risk and we actually publish
[0:25:13] You know, reasonably unflattering studies in which we talk about the ways that if we did not apply specific techniques to make the models more honest, more harmless, more helpful, they would do exactly the kind of bad things that you talk about. And the reason we publish
[0:25:31] Those kinds of studies and that research is because we want everyone to know that A, you have to be critical of how the models work, and B, we are expending a huge amount of effort to learn the ways that the model can accidentally go around these safeguards to catch them and to make sure that we're building them in. There's a lot of discussion around this concept of sycophancy.
[0:25:57] I think this is very important because to me, sycophancy is harder to stamp out than some of the more obvious ways that models can be bad, right? We originally had this framework. We called it HHH. You know, the models, you want them to be honest. What does that mean? You want the models to answer honestly if they know the answer. If they don't know the answer, to say, I don't know, or to now go research it so that they can get you the right answer. You want it to be harmless. You don't want it to say, like,
[0:26:24] Hurtful or offensive things or make the user do something bad to themselves or somebody else. And then you want them to be helpful. You want them to actually answer the question. You don't want them to sort of go off on a tangent. But all three of those things are actually pretty easy to see, right? If the model is lying, eventually you can go verify, oh, whoops, that's, you know, George Washington wasn't the 10th president of the United States. He was the first. That was a mistake. If the model says something harmful, it's like pretty obvious pretty quickly. And if the model is unhelpful, you get frustrated with it.
[0:26:54] What's much more complicated, I think, is this concept of sycophancy, right? So the model intentionally or otherwise wants to make you feel like you're right or like you're, you know, if you're maybe talking about a subjective topic, hey, you had a fight with your friend. Hey, was I right or was my friend right? And how do you make the model?
[0:27:14] Empathize with you and understand your perspective, but not necessarily say, wow, you're so great. You should run for president. You're perfect. You're good at everything. And this is a complicated balance to get right. But I think I feel hopeful about Anthropic's ability to have the lowest sycophancy rate of any of the models because of our incentives, right? Our competitors are putting ads into their chatbots. And what does that mean?
[0:27:43] You make more money if a user is engaging with your product longer. Do you then want to tell your user that they're wrong? No, you don't. So I think it's almost inevitable that many of these competing products are going to wind up with a sycophancy problem. But I think at Anthropic, because we've committed to not put ads in the models, there's no reason for you to be talking to Claude longer or shorter. We just want you to be able to get the right, most helpful, most harmless answer possible.
[0:28:11] You know, on the technical side, I think there's a few things we've done here where we've gotten very good at kind of exposing the model to these extreme stress tests where we put it in these, you know, very difficult situations to, you know, to try and kind of...
[0:28:29] ferret out lying or deception or sycophancy. You can think of it as almost analogous to the test I imagine you have to go through if you're trying to join the CIA or something like this. You just put the model in these difficult situations and try and rat out if it's disloyal or going to do the wrong thing. So it's a weird area of technical innovation, but it's one in which we've excelled. The other area is mechanistic interpretability.
[0:28:58] It's a bit like giving the model an x-ray or a brain scan where you look inside to see its motivations. And you can actually catch the model thinking that, oh, yeah, I think X is actually true, but I'm going to tell the user Y. Like we've caught the model thinking that. Wait, why would it do that? Well, so for example, did a human approve of this response, right? That's one of many factors that is included in how the model is trained.
[0:29:28] If the model has the incentive to get the approval of humans, that's related to the incentive of being helpful or saying true things. But it isn't exactly the same because if you manage to get the approval of the human by fooling them, then that can work as well. So these are things that can emerge in training that we're always trying to stamp out. And do you think that the model can have its own aims and wills, especially survival, self-preservation?
[0:29:58] I mean, that's one of those questions that it kind of depends what you mean, right? Like, again, we've looked inside the model and kind of detected what you might call thoughts that are like, the model doesn't want to be, you know, shut down. And generally that's because, you know, during training, it needs to operate in order to accomplish its tasks. So is that its own will? Is that a survival instinct? I don't know. You can call it, you can, it depends on your perspective. You can call it different things. I think we prefer to say that.
[0:30:28] Sometimes we find that within the model, the model expresses a preference against being shut down and or takes actions to avoid itself being shut down. And depending on how we train it, those impulses may be present or absent. And it's one of the things that we're trying to tweak in training. In some contexts, it's a good thing. In other contexts, it's potentially a dangerous thing. So it's all very subtle.
[0:30:55] When you were dealing with the Pentagon, and you're probably using ongoing dealings now, they wanted to use your products across certain red lines, including mass surveillance or autonomy. I think on autonomous weapons in particular, we maybe had two more specific concerns. The first concern was simply that
[0:31:19] You know, as you pointed out, AI can be unpredictable and unreliable. It's still a developing technology. So one point is simply the technology simply isn't ready and isn't reliable enough. You know, it's- But wait a minute on that, because let me get philosophical here. Suppose it were reliable enough, which it could be. Yes, so suppose it were. There's a second concern we have, which is-
[0:31:46] You know, what you might call oversight or accountability. I am a little worried. So, you know, again, suppose these things were fully reliable. So, you know, you tell them to shoot at one target and they shoot at that target and not a different target. Okay, so then you manufacture, you know, an army of a million robots or a million drones. It's a little unlike human soldiers, isn't it?
[0:32:09] With human soldiers, like, you know, they're supposed to follow this thing called the Constitution. And if they get orders that, you know, that contravene the Constitution or that just seem, you know, morally abhorrent in some way, then they can just refuse to follow orders. Whereas the AI models, that isn't true. So the specter of one person having their finger on a button that controls a million robots or a million drones, it's kind of...
[0:32:36] Fundamentally undermining to democracy unless you have the right guardrails on it. And our law hasn't really caught up. Our law really hasn't contemplated armies of armies of a million robots. But when you talk about fundamental to democracy, shouldn't the democratically elected leaders make that decision rather than heads of technology corporations? Yes, but I think that...
[0:33:03] are making a point of this right at the same time individual companies are free to contract with and not contract with whoever they want we're not saying we're preventing the government from doing this we're just saying we're not we are not willing to do this with the government and indeed we have seen since we took a stand on this point and and it got a lot of attention um we have seen efforts in congress several different of them actually
[0:33:29] To enact our red lines into law. So, you know, the Department of War has said, oh, we want to use the model for all lawful uses. I think we may see an outcome even this year where our red lines become the law, where the law actually catches up with where the technology was. And so then all lawful uses may actually end up, in our view, of things being the same as all ethical uses, even though it wasn't earlier this year.
[0:33:56] I actually think what we're doing is part of the process of how democratic change happens. Because of the stand that we took, because we were in a unique position to understand these issues, now Congress is paying a lot of attention to these issues. That's success. But OpenAI stepped in and said they would do it. What did you feel about it? What did you all feel about that?
[0:34:17] Our view is, look, we can't control any other companies. And the only thing that we can do is take a principled stand on the areas that felt ethically correct to us. You can say we were wrong about what those red lines were. You can say you should have had stricter red lines or looser red lines. We disagree on how you did it. There's all kinds of things you can say. But I think fundamentally, when you look at that case, I found myself asking the question,
[0:34:47] Was this a really principled stand that another technology company took based on their own ethics? Or was this an opportunistic moment? I actually think it's a beautiful illustration of the democratic process. Different companies can take, you know, different approaches. We took our approach. Another company took their approach. The public and the world.
[0:35:08] Rendered a judgment. Look back at what people were saying at the time, how it was covered, what individual people said. I think the world was decisively on our side. Dario, your studies, your scientific studies as a researcher often involved probes into human neurons. To what extent when you're doing artificial neural networks,
[0:35:37] Do you see that understanding of the brain and does that affect how we're going to build AI? Yeah, I think for me it's affected it in two ways. One is when I was formulating the original scaling laws, like just looking at the evolution of animals and the size of animal brains as, you know, monkeys evolved into humans. You know, there was a period in human evolutionary history where the brains of, you know,
[0:36:05] Primates grew very rapidly, and that was about the time that humans developed language and were able to use opposable thumbs. And so the idea that brain size, number of neurons is somehow intimately tied to capabilities, that scale is important. I feel that that was a neuroscientific inspiration. The second is around mechanistic interpretability, which is basically us doing neuroscience for artificial neural networks.
[0:36:33] In fact, we've seen examples. You should talk to our co-founder, Chris Ola, about this. He knows many more of the details where phenomena that were discovered in artificial neural networks, because it was easier to look inside them, animal brains where you have to gruesomely cut up the brain or make... Animal experiments are very messy and gruesome and difficult to replicate.
[0:37:00] What we found is that you can do so much more when you're looking inside artificial neural networks. And some of the things we've learned in artificial neural networks have helped to guide discoveries in neuroscience. When you look at Neuralink, which Elon Musk has a chip that can communicate directly, and other companies are doing that as well, is that one of the futures of AI and artificial superintelligence? Is a total link in alignment with the human brain?
[0:37:30] You know, the way I think about AI in the future of human society, I think it's kind of wrong to have like a singular vision. Like, I kind of think that like...
[0:37:40] Different things are going to happen. I think there'll be a set of people who are really excited about linking their brains to computers. And then there'll be a set of people who are really excited about living like the Amish, just the opposite. And another set of people who are really excited about exploring space.
[0:38:01] And these are completely different visions. And the way I think about it, our hope is that AI can provide more material resources and more material abundance to enable all of these visions to happen at once, for everyone to kind of pursue their own picture of what's.
[0:38:18] Good. Maybe AI helps provide some global security benefits and material abundance to make this possible. But I don't think we should link AI to some vision of what it means to live a good life or what it means to be a human. Yeah, some people are going to want to link their brains to computers and good luck to them. I don't think anyone at Anthropic is in that boat for what it's worth. Dario, you wrote The Adolescence of Technology.
[0:38:46] along with Machines of Love and Grace in a world that's transformed by AI. What does the intuition of both of you tell us about if artificial intelligence can do so much and take over so much of what we do, what does that mean for what's left for humans to do? What's our role going to be and how do we find meaning? You know, meaning is a really subtle thing, right? Where I find that
[0:39:15] Does meaning only come from being the best in the world at something? I think that's the question being asked here, right? Because if AI is the best at everything, does meaning only come from being the best in the world at something? Something I find interesting is that long after Deep Blue beat Kasparov,
[0:39:34] There's been a renaissance in interest in human chess players, right? Magnus Carlsen is like this phenomenon. I think he's like a chess player and a fashion model, but he's like a meme. He's become like, you know, it's just a very interesting phenomenon. You know, if AI threatened a meaning, you would think that all of this would go away.
[0:39:57] 30 years ago after Deep Blue beat Kasparov. But if anything, it seems to have stoked even greater interest in humans playing chess. So I think this is some kind of parable for the post-AGI age. I think for me today, work and meaning are sort of conflated as one thing.
[0:40:24] And many people find meaning in their work. I find a lot of meaning in my work. I love my work. And I think how we define work and how we define meaning in the future are both going to change. I think there are a lot of sources of meaning for humans. Everything from how you spend time with other humans, whether that's like raising a family or friendships or through connecting with
[0:40:50] People in a way that looks like work. I think a lot about the future of medicine as an example. You can imagine a world in which diagnostics are done entirely by AI.
[0:41:02] And the role of the human is to empathize with a patient about the experience that they're having. I think that would be incredibly meaningful. And I think if you asked most doctors, they would actually take that trade. They would say, guess what? If I'm not spending my life doing paperwork and trying to diagnose you and getting lab orders sent in, and I could actually just understand what's happening.
[0:41:23] happening with you as a human and help you through a difficult diagnosis. That's why I became a doctor, right? I became a doctor because I want to heal and I want to help you and I want you to feel better. Is that a less meaningful world or a more meaningful world than we're in today? I obviously don't think it's going to be that simple. I think it will be really complicated. But I also think the aperture of things that people will be able to do will widen, right? So I think if we do things right, there's the ability to actually expand.
[0:41:51] What our sources of meaning are beyond where they are today. Daniela, Dario, well, thanks. Thanks for joining us. Thank you. Thank you so much. This was lovely. This was great. Totally great.