Latest / Newcomer Pod / Listen to Anthropic's Dario Amodei from Cerebral Valley + Videos from on Stage
Transcript
- 0:03Hey, it's Eric Newcomer. Welcome to the Newcomer Podcast.
- 0:06Just wrapped up the Cerebral Valley AI Summit this week and
- 0:10wanted to get straight to your podcast feeds.
- 0:12My favorite conversation of the day.
- 0:14I talked with Anthropic CEO Dario Ammo Day.
- 0:17It was a spicy and lively conversation all day.
- 0:20There had been fears about whether we'd hit a scaling wall,
- 0:23whether this is sort of the end of the developments and
- 0:26foundation models that we've grown addicted to in AI world.
- 0:30And of course Dario is got the second most highly valued
- 0:34company in the foundation world space.
- 0:35Is Anthropic going to deliver another great model And he
- 0:38seemed to be suggesting yes, developments would continue.
- 0:41We got into much more than just the pace of development.
- 0:44Obviously sort of a doom and gloom guy who's now recently
- 0:48written an optimistic piece. So we got into that.
- 0:50I teased him about Elon and Sam Altman and talked about
- 0:54regulation and everything current and fun, even their
- 0:57partnership with Amazon. I was I was digging into that.
- 0:59So Woody ever sold the company? So it was a great conversation.
- 1:03And without further ado, here's my interview on stage at the
- 1:06Streetville Valley AI Summit in San Francisco with Anthropic CEO
- 1:11Dario Amodei. Dario, thank you so much for
- 1:17being here. Thank you for having me.
- 1:19All right, we've been talking all day.
- 1:21I hate to hit the same theme over and over again, but you
- 1:23actually know the answer. Are LLM's gonna get smarter?
- 1:28Have we hit a wall? Like, are we all going crazy for
- 1:31no reason? How do you?
- 1:32What's the trajectory? So I've been following this
- 1:35field for 10 years and was among the first to kind of document
- 1:40the scaling laws and the scaling of AI.
- 1:43Nothing I've seen in the field is is, you know, out of
- 1:47character with what I've seen over the last 10 years or leads
- 1:50me to expect that things will slow down.
- 1:51First, I think the class of reasoning models that we've seen
- 1:55the use of test time compute is, is going to continue the
- 1:58scaling. And I think even within the
- 2:00current scaling laws, I think something that people don't as
- 2:03much appreciate who haven't been in the field or following it is,
- 2:07you know, every time we train a new model, there's always
- 2:11there's always challenges, right?
- 2:12Like the failing laws don't say every time you train a larger
- 2:15model it does better. It says every time you train a
- 2:17larger model. If you have high quality data,
- 2:19if you're able to get the engineering right, if you're
- 2:21able to, well, you. Have to come up with new ideas
- 2:23still. Every time the optimization
- 2:25often small new ideas, right often like, oh, I got to get
- 2:28this numerical optimization thing right or you know, I need
- 2:31the the floating point precision needs to be done in the right
- 2:34way. I, I do think there's perhaps an
- 2:36effect that as we're spending more and more money, the
- 2:39engineering complexity of getting it right is, is
- 2:43increasing and that's why companies need to be larger,
- 2:46need to have more talent. But I don't see any evidence
- 2:49that the kind of the scientific, I don't see any refutation of
- 2:53the scientific hypothesis. To the extent there is a
- 2:56barrier, is it about data, is it about money, or is it just about
- 3:02sort of the approach to processing the data?
- 3:05Again, I I mostly don't think there's any barrier at.
- 3:08All OK. Are you gonna deliver a new
- 3:11model soon? You know, I mean don't talk
- 3:14exactly about model model release dates, but you know, we
- 3:17will. We will as a general matter, you
- 3:19know, we will, I hope, continue to have you know better models
- 3:23every, every, you know, every, every, some number of months.
- 3:28X number of months, what when you think to sort of future
- 3:33models and what you'd like to see over, you know, not
- 3:35necessarily the next model, but like, what are things you'd like
- 3:37your models to be able to achieve that they can't today?
- 3:40Yeah. I would like to see the models
- 3:43get better at professional tasks.
- 3:47So, you know, one thing I've said is that, you know, the
- 3:50models of last year of 2023 were, were maybe a high school
- 3:53level of capability models we see today.
- 3:57Are they, they have this, this five of being a kind of the
- 3:59intern level or the strong undergraduate level.
- 4:02If we go to, you know, next year, I'm hoping that we'll see
- 4:05models that have a kind of almost grad school or
- 4:08professional level. We're starting to get to that
- 4:11level with coding, right, where models are great coding
- 4:13assistants, but I think they're going to do even better.
- 4:15We, we look at what's happened this year so far, we've gone
- 4:18from about 3% on Sui bench, which is kind of a real world
- 4:22snapshot of software engineering activities, right?
- 4:26It's like things that things that simulate get up, pull
- 4:28requests we've gotten from maybe 3% to 50%.
- 4:31And I think in the next year, we're gonna get that much higher
- 4:35still. Do you want the Anthropic text
- 4:40box to solve the coding problem and to replace sort of cursor or
- 4:45whatever copilot? Or do you plan to build stand
- 4:49alone products? Or when you think about this
- 4:50coding thing specifically, what's your approach?
- 4:52Yeah, I mean, I think I, you know, I think the distinction
- 4:55between those is not as sharp as you might think, right, Which is
- 4:58larger and larger parts of the task, more and more autonomy get
- 5:02handed over to the AI system. But there's, there's always, at
- 5:06least in the short, in the medium run or role for the
- 5:08human, which is, you know, there's, there's in economics,
- 5:12there's this concept of comparative advantage, right?
- 5:14And even, and if the AI system is doing 90% of what the human
- 5:17does, the human just invest more and more in that 10% and in new
- 5:21things that complement the new things that complement the 90%.
- 5:24I mean, maybe someday totally. And then the AI systems will do
- 5:28everything that everyone in this room, including me and you do.
- 5:31But, but you know, I think that's, I think that's further
- 5:34out. So I, I, I actually think the
- 5:37amount of autonomy will increase, but there will always
- 5:39be interaction with humans. It will always be humans who are
- 5:42driving the creative process. What the you're you're sort of
- 5:47known as a sort of doom and gloom guy, but you wrote this
- 5:51machines of loving grace to sort of signal some optimism.
- 5:56I mean are are you still the doom and gloom guy that we know
- 5:59and love or what's your attitude towards being afraid of AI right
- 6:02now? You know, I think one of my, one
- 6:03of my points here is that optimism versus pessimism is
- 6:07like, perhaps not, is perhaps not the right axis, right?
- 6:11You know, I think I think extreme change versus lack of
- 6:16extreme change is, is is perhaps irrelevant access.
- 6:19And for a long time I've been someone who who predicts that
- 6:22this field is going to have a lot of disruption and a lot of
- 6:24extreme change. I continue to make that
- 6:26prediction. I'm maybe I'll be right, maybe
- 6:28I'll be wrong. But if you do believe that
- 6:31there's going to be a lot of a lot of disruption, then you
- 6:34know, you can think of some amazing good things that that
- 6:37can happen. But then there's.
- 6:38There's you want to hit a couple, Everybody should go read
- 6:40the essay. But what?
- 6:41Yeah. What do you think?
- 6:42The good things that could happen.
- 6:43Yeah, so I, I was, I was previously a, a biologist and a
- 6:46neuroscientist, you know, before this was 10 years ago before I
- 6:50went into AI And, and one of the reasons I found that field
- 6:53difficult, found that people were getting stuck and that
- 6:56progress was slow, was just the, just the complexity of the
- 6:58field. You know, if we think about all
- 7:00the, all the, you know, regulation processes of the
- 7:03cell, you know, all the, all the proteins that are made from DNA
- 7:06and how they interact. I mean, the most complicated
- 7:09code base you've ever worked with is nothing compared to the
- 7:11complexity of the cell. But it's not fundamentally
- 7:15intellectually beyond us. There's just a lot of facts and
- 7:18a lot of interaction between those facts.
- 7:20And it seems like exactly the kind of problem AI can solve.
- 7:23So I, I'm optimistic that diseases that have been with us
- 7:27for, you know, thousands of years, things like cancer,
- 7:30things like Alzheimer's disease, perhaps the very ageing process
- 7:34itself, that, that, that limits, limits our lifespans, that, that
- 7:38these things may be amenable, may yield to AI systems that are
- 7:42actually able to comprehend the complexity and figure out how to
- 7:45intervene. On it is that data like ready
- 7:48for a machine to process? I think one thing I've been
- 7:51struck by your writing is just the sense that like, OK, we
- 7:53might get the Super intelligence, but then humans
- 7:56are going to limit it in its ability to sort of get the data
- 7:59it needs, even though if it had the data, it could do great
- 8:01things. Yeah.
- 8:02So I think as a technical matter, you know, I'm not, I'm
- 8:04not imagining AI as like, you know, the scientists go out and
- 8:07run all these experiments and collect all this data and the AI
- 8:09analyzes it. I agree.
- 8:11That's actually quite limit limited and may not be the
- 8:13limiting step. I actually think we're going to
- 8:17have a picture that's more like, you know, AI's are like
- 8:20professors, PI's, principal investigators and Co PI's that
- 8:25design experiments say, hey, we can understand understand what's
- 8:28going on better with this disease.
- 8:30If we invented this new kind of microscope or we invented this,
- 8:33this new kind of molecular biotechnology for, you know,
- 8:36editing genes or like, you know, tagging what part of a cell or
- 8:41protein is built in. And I think if you add up a
- 8:43bunch of these interventions, you'll make what's going on in
- 8:45biology transparent. And that is going to be the
- 8:48first step towards, you know, really understanding what is
- 8:51going on biologically. And that is the first step
- 8:53towards, you know, designing drugs, pharmaceuticals that step
- 8:56in. And I think AI is going to be
- 8:57able to do every step type of that.
- 8:59The only part that I think is going to be hard is is our human
- 9:02institutions. These these AIS, I think are
- 9:04going to be so good that are our poor human institutions are are
- 9:08are not going to be able to keep up with them.
- 9:11In some cases, they shouldn't. Some of those institutions are
- 9:13there to prevent change from happening too quickly, from
- 9:17things being too disruptive to, you know, to prevent, you know,
- 9:21society from being disrupted or collapsing.
- 9:24But you know, some of those barriers are unnecessary.
- 9:26So that's something we're going to have to work out in response
- 9:29to very. Right.
- 9:30I mean, I read I read your essay as partially sort of putting
- 9:35regulatory barriers to AI could be bad because there are
- 9:38important things we need to achieve in any arbitrary barrier
- 9:41could be a real problem. I mean, Anthropic sort of
- 9:45debated whether to support the California legislation.
- 9:48Were you ultimately in favor and what sort of your push in
- 9:52legislation? Right.
- 9:53Yeah. So I think one thing I might
- 9:55push back against is this kind of unidirectional axis, right,
- 9:59of like more regulation versus less regulation.
- 10:02I think the way we prefer to think about it is that
- 10:06regulation should be very surgical, precisely because of
- 10:09what you said, precisely because I see all of the the benefits,
- 10:13right, the ability to save people's lives.
- 10:16You know, I'm, I'm, I'm, I'm, you know, I'm not interested in,
- 10:19you know, you know, the parents or grandparents of anyone in
- 10:21this room dying, you know, because we couldn't deploy AI
- 10:24fast enough, right? Because we, because we over
- 10:26regulated it. Like I, I feel that everyday.
- 10:29I'm very sympathetic to that. At the same time, precisely
- 10:32because of all the things we can do with biology, with cyber,
- 10:37with the autonomy of the systems, I think if you have an
- 10:40appreciation of what these models can do, you also have an
- 10:43appreciation of how they can be misused by various actors.
- 10:46And so I think that what we need is some some very targeted
- 10:49regulation testing that's focused only on the worst
- 10:52threats and, and, and whose goal is.
- 10:54Nuclear or what are your big ones?
- 10:56Yeah, Bio, cyber, nuclear autonomy risks, right?
- 11:00These these things that are really, you know, threats to
- 11:03society and humanity, right? Things where it's like, man,
- 11:05this is there's really an issue of like public safety and
- 11:08national security. Are we close to this AI that we
- 11:11need to keep in a box or the risk of the you understand it
- 11:14much better than me, but this AI that's gonna run free scenario,
- 11:16how, how close does that feel? You know, I think right now the
- 11:20models are not smart enough, but they're very quickly becoming
- 11:23smarter. The models are also not
- 11:26autonomous enough. So I think one reason people
- 11:29don't take these concerns seriously is that you're like,
- 11:32hey, it's just a chat bot. You talk to the chat bot, you
- 11:34talk. It's absurd that the chat.
- 11:36Way and reason is like it's math.
- 11:37It's just math. It's just, well, I don't know, I
- 11:40mean, isn't, isn't your brain just math Like the same
- 11:42calculation? You know, when a neuron fires,
- 11:45it sums up the synapses? That's math, too.
- 11:48So, you know, we shouldn't be afraid of Hitler.
- 11:50It's just math, right? No, I think it's a well taken
- 11:56point. I mean, it is like eventually
- 11:58like an intelligence system gets more intelligent.
- 11:59Where Yeah, no, it's the whole, the whole universe is math.
- 12:03But. But no, I mean, you know, I
- 12:04think, I think, I think. Joking aside, you know, I, you
- 12:10know, again, I think, I think we need to balance these things.
- 12:12But the box sort of, yeah, yeah. I mean, it's sort of the idea
- 12:15that it needs to be isolated from the Internet and you're
- 12:17worried about it. I think once we have these
- 12:19agents that are acting, people will have an intuition for one.
- 12:23These are really powerful and these are going to make our
- 12:25lives a lot better. But man, you better know how to
- 12:27steer these pretty well. And all this stuff about, you
- 12:29know, people laugh today when the chat bots like, you know,
- 12:32they say something a little unpredictable, like we're going
- 12:35to have to do a better job of controlling the agents than that
- 12:37because you know, it's it's not as funny when they RM dash your
- 12:40hard drive. Do you?
- 12:41What's your P doom these days? I I don't really like.
- 12:45You've learned your lesson on this one.
- 12:47Like those? I never liked those words.
- 12:49I think they're, they're kind of, I think they're kind of
- 12:51weird. I, I, I just like my, my, my
- 12:53view is just that like, look, we should, we should measure for
- 12:57risks as they come up. And, you know, in the meantime,
- 13:00we should get all the economic benefits that we could get.
- 13:02And we should find ways to measure, measure the risks that
- 13:07are effective but are at minimally disruptive to the the
- 13:11amazing economic process that that we see going on now that,
- 13:14you know, the last thing we want to do is slow down.
- 13:17What what's the Trump administration mean for the
- 13:19trajectory of AII? Mean You're, you know, you've
- 13:22got security now. You see yourself.
- 13:24You know, I think understandably large language models are going
- 13:27to become the thing of nation state interests.
- 13:29Like what's your interaction with the US government?
- 13:32And does that change in a Trump world?
- 13:34Yeah. Look, I mean, there's several,
- 13:35there's several dimensions to it.
- 13:36I mean, you know, we've always worked with the government on
- 13:40various kinds of policy. I mean, I've advocated for this
- 13:44type of targeted regulation. You know, I think we advocated
- 13:47that to the last administration. You know, that's our position
- 13:49for this administration. We're going to be completely
- 13:51consistent on the matter. Obviously, different
- 13:53administrations have different positions, maybe more or less
- 13:56receptive, but that's not something we can control.
- 13:59We're just going to advocate for what?
- 14:00For what we believe in, in terms of things like, you know,
- 14:04security around the models. Yeah, this is this is this is a
- 14:07worry that this is a worry that I have.
- 14:09I mean, I think we shouldn't be naive.
- 14:11The economic power of these models is going to turn into,
- 14:15you know, it's, it's gonna turn into governmental and state
- 14:17power. And you know, I think I, I do, I
- 14:21do think it's important, you know, the use of these models by
- 14:25autocracies could be really terrible.
- 14:27Are we are we picking our favorite philosopher king?
- 14:29Is it sort of like we have to shop which AICO look into their
- 14:32soul and see which one we trust? Them yeah, I mean, you know, I
- 14:35think, I think it's AI think it's AI think it's a, you know,
- 14:38I think it's a bad, I think I said this on a podcast.
- 14:40I think it's a bad idea to think in terms of individuals like
- 14:44companies or institutions. We've tried to build Anthropic
- 14:47in such a way, you know, we have these structures like long term
- 14:50benefit trust, you know, we've imposed responsible scaling plan
- 14:54and ourselves. That would be the point of
- 14:56targeted regulation that, you know, I think a reasonable
- 14:58society, particularly when there are technologies that grow and
- 15:02change this fast, it needs to put limits on the power of
- 15:05individuals. And that applies to people in
- 15:07government and applies to people who run companies, right
- 15:09everyone? So you want a CEO who can get
- 15:11kicked out by the board and get back in and really overwhelm.
- 15:15That's how companies should be. Right.
- 15:17I think you're baiting. Take the bait.
- 15:20A company you defected from, Yeah.
- 15:23To put it in more sort of a different way, like the race to
- 15:27the top idea part of what? The sort of, I think a core goal
- 15:31of Anthropic is if we can be good, we'll attract the best
- 15:35engineers and that will put pressure on other people.
- 15:37And I mean, it seems like a lot of people are joining Anthropic
- 15:39right now, where Give us the state of play on your race to
- 15:42the top idea. Here yeah, I mean, you know,
- 15:45I've really seen this play out right.
- 15:46So there are a bunch of practices we engage in, you
- 15:49know, the, our invention of mechanistic interpretability,
- 15:52our work on things like constitutional AI to, to much
- 15:55more mundane and, and, you know, and, and kind of commercially
- 15:59focused things like, you know, we do, we try to do better on
- 16:02security. We try and produce models that
- 16:04enterprises can trust all of these things through various
- 16:07channels other companies tend to adopt after we've done them.
- 16:11So I, you know, there are now nearly 1000 people at Anthropic.
- 16:16And so I've, I've been through a lot of kind of, you know, sort
- 16:20of, you know, recruiting fights between US and, and other
- 16:23companies. And so I see it as a repeated
- 16:25game. You know, there will be, there
- 16:27will be, there will be an engineer or researcher who comes
- 16:29to Anthropic. I'm really, really inspired by
- 16:31your interpretability work. I'm really inspired by, you
- 16:34know, this stuff you've done unconstitutional AI, you've been
- 16:36really thoughtful about the RSP. And I always say, go to the
- 16:40competitor you didn't choose, write up a document, give a list
- 16:43of reasons you chose Anthropic. And then I repeat that.
- 16:45And after that's happened five times, the other company gets
- 16:48really sick of it. And and you know, then before
- 16:51they know it, they have a team that's doing the same.
- 16:53Thing is, is Elon responsive to this?
- 16:55Is XAI? Do you see them?
- 16:57I mean, that does not feel like a company that's going to engage
- 16:59in your race to the top. I see them as a real competitor
- 17:02at the. You know, it is, it is pretty
- 17:04interesting. You know, I mean, you know, I, I
- 17:07don't want to, I don't want to comment specifically on other
- 17:09companies, but you know, Elon has this, this kind of, you
- 17:12know, view that defies classification, right where, you
- 17:15know, on one hand, I think, you know, the models are very, you
- 17:19know, if you just look at them in in kind of a normal way, just
- 17:21kind of mundane safety things, It's kind of it's kind of a Wild
- 17:24West. But you know, he's he's also
- 17:26been in some context supportive of regulation.
- 17:28So he's he's one that I think is very hard to classify.
- 17:31I mean, you probably know better than me, but it feels a little
- 17:33bit like if he's in charge, he trusts it.
- 17:35When other people are in charge, he is somewhat apprehend.
- 17:38You're you're baiting me. Again, well, we want to know
- 17:41this is these are real existential things.
- 17:43We have to have an open and honest conversation about what's
- 17:45happening. And I, you know, I think the
- 17:46premise of a lot of my writing is that the actual characters
- 17:49and individual people really matter.
- 17:51And some of these. I think they do really matter,
- 17:53but I think we should do everything we can to make them
- 17:55matter less. Institutions.
- 17:57That is my, that is, that is my perspective here.
- 17:59When individuals matter, we've too much.
- 18:01We failed in some way. 11 just sort of specific question I had
- 18:06because I've actually been debating with my wife about
- 18:08this. Do you, do you think your models
- 18:10and other models are going to be film makers and creatives or
- 18:14where are you on sort of the models as a creative spirit?
- 18:18Yeah, again, I mean, I think the models will have quite a lot of
- 18:21creative ability. You know, I would go back to
- 18:24this view of, you know, at least in the short and medium term,
- 18:27right? There's some world where all the
- 18:30AIS are smarter than, you know, incomparably smarter than all
- 18:33the than all the humans. And we can debate when that
- 18:36happens. The logic of comparative
- 18:38advantage is extremely powerful. Even when the models can do 90,
- 18:42even they can do 99% of of what humans can do, humans will
- 18:47continue to have a surprisingly large creative role.
- 18:49So I, I, you know, I do think that even as the model start to
- 18:53be creative, you know, I, I use Claude to help with, with my,
- 18:56with my own writing, it still very much feels like it's
- 19:01authored like human by humans. One thing that maybe has been a
- 19:04little bit less like this is the the image generation, the video
- 19:07generation models where what they do it's, it's kind of like
- 19:09lumpy. You generate an image or you
- 19:11don't. We're starting to do editing on
- 19:13it. But this is maybe one reason,
- 19:15not the only reason, but one reason Entropic has been less
- 19:18involved in models like that. They do feel like they're more
- 19:20of a direct replacement. I think they have a wider range
- 19:24of safety issues associated with them, something about.
- 19:26The like diffusion models overall.
- 19:28Yeah, yeah, like like text image models or text to video models.
- 19:32You're you're gonna stay away or.
- 19:35You know, we, we haven't, we haven't done them for now.
- 19:38We have no current plans to do them.
- 19:40You know, we never want to say never.
- 19:41You know, lots of things could change in the in ecosystem both
- 19:45you know, about whether it makes sense from a business
- 19:47perspective to do them or you know, the safety properties of
- 19:49doing them. But this is this is not
- 19:51something we've done so far and it's not something we're
- 19:53currently working on. And throughout the business, I
- 19:56mean, how do you take stock? It seems like you're doing very
- 19:58well selling to businesses, working with businesses and just
- 20:02people here get the consumer product, but maybe America at
- 20:06large doesn't. Yeah.
- 20:07Is that a fair assessment? Yeah, you.
- 20:09Know, I think we've been deliberately much more focused
- 20:12on the enterprise side than the consumer side.
- 20:15Every other company in this space either is a large consumer
- 20:19focused company or is a startup that for a number of reasons
- 20:23seems characterologically more suited for consumer, even if
- 20:27they do both consumer and, and, and enterprise.
- 20:30So, you know, we've, we've, you know, we, we, you know, we've,
- 20:34we've been majority enterprise for, you know, most of
- 20:37Anthropic's history. And so most of our attention,
- 20:40most of our development attention goes there.
- 20:43And you know, it really shows in what we've been able to put into
- 20:45our API. For example, we identified
- 20:47coding as an area that lots of businesses want to build on.
- 20:50So for example, you know, cursor which which people mentioned
- 20:54there's, you know, cognition we're recently in GitHub
- 20:57Copilot, there's. Coding their their software is
- 21:00powered by Anthropic is what you're saying.
- 21:03You're the brains, they're they're figuring out the
- 21:05software. Well, yeah, I mean this, this
- 21:07is, this is kind of, this is kind of API business, right,
- 21:10Where you, you make a model available and people are going
- 21:13to build a product on top of that model.
- 21:15And, you know, we don't have time to worry about the details
- 21:17of every product. You know, we, we can't make a
- 21:19product for coding and a different kind of product for
- 21:22coding and a product for legal and a product for something
- 21:24else. And so I think that's actually
- 21:26the magic of serving enterprises is that you let the ecosystem do
- 21:30the exploration for you. How much?
- 21:32Does the success of Anthropic depend on being the smartest?
- 21:37Like you need to be the leader and you need to maintain a
- 21:40persistent edge? And if so, man, that sounds like
- 21:43a stressful business for. You know, I think, I think to be
- 21:46blunt, there's there's really two ways to succeed.
- 21:48If you have a much better model than everyone else, then you'll
- 21:50succeed everywhere. If there are a few players with
- 21:54strong models, which is the position we're in now.
- 21:56I mean, I think, you know, anthropic models for for many
- 21:59purposes, especially code, you know, I think, I think I think
- 22:02are the strongest. But you know, we have maybe 3-3
- 22:05or four players with, with fairly strong models.
- 22:08If if you have a kind of product direction or a niche or a
- 22:13business model that's different from that of the other three
- 22:16players, which isn't very hard because there aren't that that
- 22:18many players and, and the models are, you know, roughly
- 22:21comparable, a little little better or better, little worse.
- 22:23That's a sustainable business model.
- 22:25And I think for Anthropic that's been deployment on clouds like
- 22:28AWS and GCP and it's been enterprises compared to
- 22:32consumer. Although, you know, despite
- 22:34consumer not being our central focus, we've put some effort
- 22:37into it and it's actually, again, it's grown quite fast for
- 22:40something that is not the. Central I know it's easy to
- 22:42underestimate not how how young your company he still is and how
- 22:45much has happened in such a short.
- 22:47Yeah, I think, I think of the five, you know, companies were
- 22:49the only one who hasn't, you know, been around for 10 years
- 22:52or is tied to some large existing source of, you know,
- 22:55company wealth or individual wealth.
- 22:57From from what you see, you're building this API business.
- 22:59I don't know what company should people start that like Anthropic
- 23:02is too busy to do or that you think you are so making
- 23:05possible. Code, code is definitely, you
- 23:07know, you know, code is definitely going going crazy
- 23:11these days. And I think that's going to
- 23:13continue. The progress is going to
- 23:15continue. I think that in many ways has a
- 23:17wind at its back because code has this property during
- 23:21training or during deployment that it's really easy to verify
- 23:24in an automated way, right? You, the human doesn't, doesn't
- 23:27need to be in the loop of seeing how good it is.
- 23:30You know, my, my favorite in line with machines of love and
- 23:32grace. I, I would love to see more in
- 23:35the health and biology side. You know, that's, that's maybe
- 23:38less of a business judgement and more of a judgement of the
- 23:41humanitarian value, Although I think there's a lot to be done
- 23:44there. And I think it'll, it'll, you
- 23:46know, as a side effect, also be also be quite profitable.
- 23:49You know, I think there's a lot to be done in the legal and
- 23:52financial space. I think those areas are are
- 23:55pretty ripe for. Where you just are a better
- 23:58trader than everybody else or a trading service.
- 24:00Yeah, or, or, you know, yeah, providing, providing financial
- 24:03services and kind of a better and more informed way.
- 24:06I think on the legal side, like huge amount of legal work is
- 24:09just just super repetitive. You know, we're powering a
- 24:14service with Intuit now, which, you know, you can like ask, ask,
- 24:17you can ask better questions when, you know, when you're when
- 24:20you're filing your taxes. So I think things like
- 24:23accounting consulting we're partnering with.
- 24:26Related to coding, sorry to interrupt synthetic data, are
- 24:28you bullish on that or are you think?
- 24:30I'm, I'm quite bullish on synthetic data.
- 24:32Yeah. I mean, you know, the example I
- 24:34always give of synthetic data is alpha go zero got better than
- 24:37the best human player at go without ever training on human
- 24:41data or playing a human player. It only played itself.
- 24:44So it it may seem like you can't get something from nothing, but
- 24:47alpha go zero kind of got something from nothing.
- 24:49How much you know, how important is it that anthropic always stay
- 24:53a stand alone business? I mean, obviously you have these
- 24:55relationships with Amazon and Google and, you know, I guess in
- 24:58a Democratic administration with Lina Khan, it would have been
- 25:02like, well, you could never sell anyway, whatever.
- 25:04Now there seems like more of a chance that someday you could
- 25:06sell. Is that off the table for
- 25:08security concerns or how do you think?
- 25:10Yeah. I mean, you know, I think, I
- 25:11think our picture is that we, we believe that our mission is best
- 25:15served by be by, you know, by being an independent company,
- 25:18right. These things we have like long
- 25:20term benefit, trust, responsible scaling policy, like those are,
- 25:24those are things that are very important to us.
- 25:26And you know, you know, you know, by just by by by far my,
- 25:30you know, my strong preference is, is to do these, is to do
- 25:33these as an independent company. And you know, if we look at our,
- 25:37our position in the market and, and, and like and like and like
- 25:40and like what, what we've been able to do, the independent
- 25:43partnerships we have with Google, with Amazon, with
- 25:46others, you know, I think, I think this is, I think this is a
- 25:50very viable thing. Can you use their chips or do
- 25:52you basically need to use NVIDIA most of the time?
- 25:56We use NVIDIA, but we also use custom chips from both Google
- 25:59and Amazon. And how are they stacking up
- 26:02right now? What's?
- 26:03Your, I mean, look, different chips have different tradeoffs.
- 26:06I think we're getting value from all of.
- 26:09Them the last. Thing just you like serious
- 26:12intelligence. I get it wrong every time.
- 26:14Not AGI wait. I I mean, I it was just a
- 26:16throwaway phrase. I called it powerful.
- 26:18Powerful. Whatever I wanted.
- 26:20To give it some, I wanted to give it some term so generic to
- 26:23indicate. That like, well, it's already
- 26:24hard. To defy.
- 26:25There's no buzzword here. There's just like models that
- 26:28get better and. Better define define it for
- 26:29yourself and then give your latest prediction on how far
- 26:32away we. Are Yeah, I mean, you know, I I
- 26:34would say again, you know, a model that's like as good as a
- 26:37Nobel Prize winner, polymathic across many fields, has agency,
- 26:42can act on its own for our for hours or days.
- 26:46And and where there's a large collection of mid data center,
- 26:48right, like hundreds, hundreds of thousands or maybe millions
- 26:51of them, I think once we get to that.
- 26:54Point a year. What year society's gonna change
- 26:57again. You know, I've, I've always had
- 26:58very large error bars around them.
- 27:00But you know, I think, I think, you know, it could happen in
- 27:032025, could happen in 2026, seven.
- 27:07It could. That's gonna change our like,
- 27:08our lives are gonna be totally different when this happens.
- 27:11Or yeah. Yeah, I mean, you know, there
- 27:12will be issues around like the, you know, the human barriers,
- 27:15the point. And you know, it's not, I think,
- 27:17gonna change our lives instantly.
- 27:19Right. Well, you commit to one year.
- 27:20Which year is your big? Right now, faster than others.
- 27:24Which which year are you picking in this?
- 27:26Just come on. No, no.
- 27:29All right, Dyer, thank you so much.
- 27:31You've been a great sport. Really appreciate it.
- 27:33Thank you.