Anthropic Co-founder: Building Claude Code, Lessons From GPT-3 & LLM System Design

Tom Brown co-founded Anthropic after helping build GPT-3 at OpenAI. A self-taught engineer, he went from getting a B-minus in linear algebra to becoming one of the key people behind AI's scaling breakthroughs. And his work is paying off. Today, Anthropic's Claude is the go-to choice for developers, and his team is overseeing what he calls "humanity's largest infrastructure buildout ever." On this episode of The Lightcone, he discusses his unconventional path from YC founder to AI researcher, the discovery of scaling laws that changed everything, and his advice for young engineers entering AI today.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  When we started out, we didn't seem like we were gonna be successful at all. OpenAI had a billion dollars and like all of these, all of the star power. And we had seven co-founders in COVID, like trying to build something. And we didn't know if we were necessarily gonna make a product or what the products would look like. One thing that's interesting to look at is just that humanity is on track for like the largest infrastructure build out of all time.

• 0:25Jared Friedman?

  Tell us about the early days of Anthropic. So you had a general idea of this sort of like long-term mission that you wanted to do to, you know, not destroy humanity. But like, what did you actually work on for the first year? How did that converge on an actual product?

• 0:45Garry Tan?

  Welcome back to another episode of The Light Cone. Today, we've got a real treat, co-founder of Anthropic, Tom Brown.

• 0:52Speaker 00

  Excited to be here.

• 0:53Garry Tan?

  So, Tom, one of the things that a lot of the people watching would love to figure out is you got started in tech at the age of 21, fresh from MIT. How does someone go from that in 2009 to literally co-founding something as important as Anthropic?

• 1:12Speaker 00

  Summer 2009, Linked Language, two of my friends had started that out. I think they had seen one of our other friends, Kyle Vogt, kind of do a YC company. And so it was in the water that that's a thing that we could try to do. They started out, I was the first employee. Back then, yeah, you guys let me join for all the dinners and stuff like that, too. I could have instead gone to like a big tech company or something like that. And I think probably just as a software engineer, I might have learned more software engineering skills. But I think by being there with the other co-founders, without anyone telling us what to do, basically, we like, we had to figure out how to live, how to like, the company would die by default. I think in school, there was a lot of like a feeling of more of people would give me tasks and I would do the tasks.

• 1:56Speaker 00

  It's kind of like a dog waiting for like food to be like that to them in their bowl or something like that. And I think for that company, it was more like wolves and we have to like hunt our real like food, otherwise like where our kids are gonna starve or something like that. I think that that mindset, I think, has been like the most valuable mindset that shift that I've had for trying to do like bigger, more exciting things.

• 2:17Garry Tan?

  Yeah, big tech just teaches you to work at a big tech company, whereas it's much more fun to be a wolf.

• 2:24Speaker 00

  Yeah.

• 2:25Harj Taggar?

  How did you go from like so working at friend startup to then you started your own one?

• 2:30Speaker 00

  So linked was, um, we ran the company for a bit. I ended up going back to, to school afterwards. And then when I left school, I went to this company, Mopop.

• 2:41Harj Taggar?

  That mobile advertising thing, right?

• 2:42Speaker 00

  Yeah, yeah. I was like the first engineer there. I was like, okay, I want to be a wolf, but like, I was really bad at programming also. I was like, very, very struggling as like a, a, like software engineer. I know I want to do more, but I don't know how to do it yet. And so I think that was kind of like a experience getting to scale something. Winter 2012, one of my friends who was my smartest friend from college pitched me on, let's go and start a Y C company. We did at the time, solid stage. This was before Docker existed. And so the idea was try to make it easier to do DevOps, but Docker doesn't exist. So it's going to be a more flexible Heroku, which basically meant a more complicated like Heroku. And so we, I remember we like, we interviewed with you guys.

• 3:24Speaker 00

  I think folks didn't really understand what we were trying to build. I think we didn't really understand what we were trying to build that much.

• 3:30Garry Tan?

  When you're trying to do something new, that's actually sometimes common.

• 3:33Speaker 00

  Yeah, I think we were an outlier there because we like did our interviews and then we got called back driving back to San Francisco. And T L B had written on the board like an angry frowny face. What are you actually going to build? And so he like, wanted us to explain that. I guess we explained it enough. Or he was just like, these guys still don't know what they're doing, but maybe they'll figure it out. Halfway through, I kind of felt I still didn't actually understand what we were going to build and how we would attach a mission to it that like I wanted to, to work on for my whole life.

• 4:05Harj Taggar?

  Yep.

• 4:06Speaker 00

  And so I left, P G actually intro'd me to Michael Waxman.

• 4:10Harj Taggar?

  I didn't realize he picked the tree.

• 4:11Speaker 00

  Yeah.

• 4:12Garry Tan?

  So. So Grouper was a dating app, only it was novel in that you had what, three guys and three girls?

• 4:18Speaker 00

  Yep.

• 4:19Garry Tan?

  This was before A I in a lot of ways. So there was like a set of, a team of people who would manually link people up, right?

• 4:25Speaker 00

  Yep.

• 4:26Garry Tan?

  They'd meet up at a bar and shenanigans would ensue.

• 4:29Speaker 00

  Yes, reliably shenanigans. People didn't always have a great time. I think you want, you want on a. The pitch for Gruber for me for like, why I was excited for it was just, I was like an incredibly awkward kid. What I wanted to do was to basically have a thing that lets awkward people like me go out and talk to other people, for me to talk to girls and feel like I was safe doing it with like my friends around and stuff like that. And so I think who we're going to be our employees was important. I did like all of our engineering interviews, we would take someone. The only person who went on more was Greg Brockman. I think he had, he had a phase where like every single week, he would go and like post

• 5:13Speaker 00

  on Slack or HipChat at the time or whatever.

• 5:15Harj Taggar?

  He was hanging out at the Recur Center during this period, I think.

• 5:18Speaker 00

  Oh, I, I think he was at Stripe. Maybe, maybe for part of it, he was at Recur. Yeah. But he also had, I think, just like a phase where he would just at Stripe, he would just like post in their thing every like, I'm going on Gruber, who's going for like a whole year. So I, I ended up being close with Greg, which, which ended up being a connection to the OpenAI.

• 5:35Diana Hu?

  What was the journey like? Because you started as, you just graduated from MIT CS, you were 21, you became first an early employee for all these YC startups. Then you started your company just a couple of years later. And what was the path for you to eventually become the co-founder of Anthropic? It was like a long path, but it's pretty impressive. How, how did you get there?

• 5:58Garry Tan?

  I mean, it sounds like getting in touch with Greg at that moment was a big moment. And then you were one of the first, you know, couple dozen people to join OpenAI as a result.

• 6:09Speaker 00

  Yeah. So I left Gruber 2014, June 2014, and I joined OpenAI, I think a year later. I tried to like build up courage to make the switch to be a, to try to learn AI research. At the time, I was like, okay, it seems like sometime in our lifetimes, we might end up making transformative AI. If we do, that would be the biggest thing. Maybe there's some way that I could help out. But also I got like a B minus in linear algebra in college. And so it seemed like at the time you needed to be just top superstar in order to try to help out with that at all. And so I think I had like a lot of uncertainty about whether I would be able to. Help. And also, I'd had some success with startups. And so a lot of me was just like, rather than trying to retool at this,

• 6:54Speaker 00

  like, I could try to do another startup or something like that.

• 6:57Harj Taggar?

  I feel like in that period, going to work on AI research, which is not seen as like a serious, like a practically serious thing to do. And you're in a world where it's like people try and build companies and do these like really practical things. So what did your, were your friends like, oh, that's really cool. You're going to work on AI stuff? Or was it?

• 7:12Speaker 00

  Not really. I think my friends were like, that sounds, that sounds weird and bad. Kind of like, it doesn't really seem like, it doesn't seem like, like AI safety is, I think we should be aware, like, overpopulation on Mars doesn't make any sense. And my friends were also just like, I don't know if you're going to be good at that, Tom. I think that for that reason, I think I didn't try very hard for, I like, kind of flip flopped on it for like, six months trying to build up courage to do it.

• 7:35Harj Taggar?

  And what were you specifically at this point, like you're reading research papers? Like, what it, what does it look like?

• 7:41Speaker 00

  Yeah, so first I was just kind of hanging out. I built like an art car for Titanic and stuff like that.

• 7:47Garry Tan?

  That was fun, yeah.

• 7:49Speaker 00

  So I spent like a whole summer, like three months after Gruper doing that. Because honestly, I was, I was like, kind of burned out for Gruper where, I don't know, startups, like the highs are high, like the lows are low and we weren't working at the end. Our business wasn't succeeding, our revenue was going down. But I, my main job still was like recruiting engineers. And so I had to like pitch them on the stream that I'd had, but I like no longer really.

• 8:08Garry Tan?

  It sounds like a death march.

• 8:10Speaker 00

  Yeah. And so I was super burnt out and I was like, okay, Tom, like chill out, do some yoga, like do some CrossFit, build an art car.

• 8:17Garry Tan?

  What was the hindsight, like, you know, hindsight's 2020? What's the retrospective on like, Gruper obviously attracted all these really, really smart people. The graphs were up into the right and then it flatlined and maybe started declining. What happened?

• 8:30Speaker 00

  I think that when we started, the competition was like OkCupid.

• 8:35Garry Tan?

  It was all web based.

• 8:36Speaker 00

  All web based. The main problem that I think we were solving was the, it's hard to like go and put yourself out there and go like, talk to someone new and they might just be like, I don't want to talk to you, you seem weird. And so we solved that by just blind matching. Tinder came out while we were doing Gruper and Tinder solved that same problem with, both people have to show interest before you get matched. So there's also no worries about getting rejected. And I think that they just had better, that was a better solution to that same problem. So good work Tinder, good work all the Swipers. I think that that, that solved the like mission that we were trying to solve better than we solved it.

• 9:10Harj Taggar?

  And then, yeah, like, when did you get serious about AI and just how did you approach? It's

• 9:16Speaker 00

  Three months of like kind of playing and having fun. And then I ran out of money also. When I had like my personal runway, I ran out. And so I was like, okay, I think that I'm going to need six months of stealth study to have a shot at getting a job. At that point, it was DeepMind or Google Brain were the two places to do work there. Or Miri, Miri was the third one that I was like looking at. So I was like, if I want to help out with that, those are the three places to look at. I don't have any of the skills yet. I need six months of self-study to feel like I would not be a drag on them and like actually be helping instead.

• 9:49Diana Hu?

  Can you maybe explain a bit what was that self-study like? Because I'm sure there's a lot of software engineers right now in their twenties are looking to retool to become AI researchers. What was, what was that six months like? Even though, as you said, you had gotten a B minus in linear algebra, which is like core.

• 10:06Speaker 00

  Might have been a C plus. I'm not, I should check.

• 10:10Speaker 00

  I'm going to keep telling.

• 10:10Diana Hu?

  But that's pretty impressive where you got to.

• 10:12Speaker 00

  Yeah, yeah, it turned out okay. First I did a contract actually with Twitch and like earned like enough to have that six months of runway. So I did like three month contract with Twitch and then I made a plan to self-study. I don't think it's the right plan now for people to, at least in 2015. What did it look like? It was like take a Coursera course on machine learning, try to solve some Kaggle projects, read linear algebra done right. And I had a statistics textbook. I think I had YC alumni credits and so I bought like a GPU. And I would like SSH into the GPU to like work through my courses for it.

• 10:50Diana Hu?

  And this is right after, yeah, it was already after Alex's neck, right?

• 10:54Speaker 00

  It was after Alex's neck, yeah. So I was mostly doing image, image classification stuff that I was trying to learn was like the thing that all the courses would teach you to do.

• 11:02Diana Hu?

  How did you get the OpenAI job? Because you were one of the few engineers who was mostly researchers and they had a pretty stacked team of researchers.

• 11:10Speaker 00

  I messaged Greg as soon as OpenAI was announced and I was like, I'd love to help out in some way. I got a B minus in my linear algebra. But I know some engineering, I've done a bit of distributed systems work. If you guys need help, I'm like happy to mop floors if, if you guys need. I want to help out, however. And I think Greg was like, yeah, I think there's like a paucity of people who, he said paucity to it. I was like fancy word there. There's a paucity of people who know both machine learning and distributed systems. So like, yes, you should do that. I think he introduced me to Peter Abiel also to help me put together like a little course for myself too. And then I checked in on with him, I think every month or something. And then after a couple of months, he was like, oh, we actually have a project, which is, uh, we need to

• 11:53Speaker 00

  put together, we want to play a game, like play games. Can you help make Starcraft environment? And so I joined to like, help them with the Starcraft environment. So that, that ended up, I think, getting my foot in the door. I, I didn't do any machine learning work with them for the first nine months that I was there, basically.

• 12:11Harj Taggar?

  And what did OpenAI feel like at this point? Like, had it raised much funding? Did it have like an office? Is what would it, did it feel like a startup?

• 12:18Speaker 00

  So it was in the chocolate, on top of the dandelion chocolate factory.

• 12:22Diana Hu?

  This is after Greg's apartment. That's the, did they

• 12:25Speaker 00

  After Greg's apartment, yeah. So like right after Greg's apartment in the chocolate factory. When it kicked off, right? It was like a billion dollars of committed funding from Elon. It felt like it was like very solid.

• 12:34Diana Hu?

  The other interesting milestone for you was when you got to build a lot of the engineering around the training for GPT.

• 12:44Speaker 00

  Yeah, for GPT three.

• 12:45Diana Hu?

  For, and how, how, what was that? Because you got from GPT two was in TPUs, right?

• 12:50Speaker 00

  Yep.

• 12:51Diana Hu?

  And the big breakthrough in GPT three was like use more compute and using TPUs.

• 12:55Speaker 00

  Yep. So I ended up working at OpenAI for a year, left, went to Google Brain for a year, came back, and then GPT three was 2018 through 2019 was like building up to GPT three, which exactly, as you said, was like scaling things up. I think that like Dario had seen the big trend of scaling laws, basically.

• 13:13Diana Hu?

  You, you published a paper for that.

• 13:15Speaker 00

  Yeah, yeah.

• 13:15Diana Hu?

  And that's like a pretty important paper that now has withstood the test of time and we're living now the dream of it.

• 13:23Speaker 00

  Definitely like seeing that line of reliably, you get more intelligence if you spend more compute with the right recipe was the main thing that was, at least for me, it was like, this is a thing that's like happening, happening now. Cause you could look even at the time, we weren't spending very much money on the, on the training jobs at the time. And you could see that there was scaling there. And then also Danny Hernandez did a paper at the time that showed uh, how much cheaper algorithmic efficiency was making stuff over time too. And like those two things stack together. That was like, oh wow, we're going to get a lot more intelligence. Over the next few years.

• 13:56Garry Tan?

  So it was noteworthy and surprising when you saw it.

• 14:01Speaker 00

  Yeah, and I think the thing that seemed the weirdest to me is like, I'm not a physicist, but like all these physicists were doing this stuff. The like original scaling laws paper, just the like very straight line over like 12 orders of magnitude. I'm just like 12 orders of magnitude is like, just like a stupidly large amount of, I've like never seen anything go over 12 orders of magnitude. That convinced me to definitely pivot all of my work into scaling, which I hadn't been doing before.

• 14:25Garry Tan?

  Can I ask a like kind of layperson question?

• 14:27Speaker 00

  Yeah.

• 14:28Garry Tan?

  I mean, is it fair to say that the scaling law might show up in all of these other domains? Then they're like, are there like two, five, 100, 10,000 domains where the scaling law could hold that we're just not investing into?

• 14:42Speaker 00

  Yeah, so I think in physics, scaling laws hold all over the place, which I didn't know at the time. But within physics, like there's a whole field called phenomenology that basically looks at various aspects of the world and then does those types of fits. And they find these like power law distributions all over the, all over the place. This was like, I think the first one that I had ever seen in a like computer science adjacent thing, which I think was like interesting and surprising and.

• 15:12Garry Tan?

  And at the time, it was, people were mad about it. They actually were like, you're throwing money at GPUs or you're just like wasting money. This is very wasteful.

• 15:19Speaker 00

  Yeah.

• 15:20Garry Tan?

  That was sort of the vibe.

• 15:20Speaker 00

  People are still mad about that.

• 15:22Garry Tan?

  Yes, different people now, but still people mad about it.

• 15:26Speaker 00

  Yeah, yeah, I guess, yeah, the researchers were mad at that too, where it's like, it's not elegant. You're just like brute forcing it, the like gesture cap, like stack more layers, like, which I think, I think like Anthropic's slogan, I think is like, do the stupid thing that works. That was a thing where like, this was very clearly the very stupid thing that, that works.

• 15:44Diana Hu?

  Can you tell us then how you ended up collecting the last Infinity Stone?

• 15:49Speaker 00

  With Anthropic.

• 15:50Diana Hu?

  Yeah, with Anthropic, because there's very few people in the world that have basically worked at OpenAI, DeepMind and Anthropic. And you were part of the team that spun off from GPT-3?

• 16:00Speaker 00

  Yep.

• 16:00Diana Hu?

  And then started Anthropic. So how was, how was that jump?

• 16:04Speaker 00

  There were two teams there. There was the safety org and the scaling org were the two orgs that reported into Dario and Daniela. I think we had just like. Worked together extremely well. One thing I think that was great, both at OpenAI and at Anthropic was just like, we had a culture where like, everything is on Slack, 100% of things on Slack. And within that, all public channels, great communication. I think that that group also was the group that took the scaling laws the most seriously. Where it was like, okay, like, this actually is going to be transformative. There's going to be a handoff where like, humanity will hand off control to transformative AI at some point. And hopefully, like, they'll be aligned with us. And like, that will be

• 16:45Speaker 00

  a good transition that goes well. But it might not be. The stakes are incredibly high. And so I think that group was very focused on like, how do we make sure that that's taken seriously enough? And that like, we've built an institution that can handle the weight of that. That ended up being the core group that left to join Anthropic. And I think, I think it wasn't clear at all to me that like, that was the right thing for the world at the time. In hindsight, now, it seems like that was a good choice. I think what was kind of cool then, too, is when we started out, we didn't seem like we were gonna be successful at all. OpenAI had a billion dollars and like, all of these, all of the star power. And we had seven co-founders in COVID, like, trying to build something. And we didn't know if we were necessarily going to make a product, or what the products

• 17:30Speaker 00

  would look like. And so I think that what was interesting from that, too, is that all of the initial people who joined were there for the mission, too. They all could have worked somewhere else for more prestige, more, more, more money. People would have known what they were doing, etc.

• 17:45Harj Taggar?

  Well, stayed at OpenAI.

• 17:47Speaker 00

  Exactly. Yeah, that, that, exactly. That's been an interesting thing, then, that I think has been like, the key to like, letting our culture, or like, let our org scale. We're like, 2,000 people now. But we still have a thing where it doesn't seem like politics have creeped in. And I think a lot of that is like, the first 100 people, all were just there for the mission. So like, if something starts to go wrong, they'll like, raise their hand and be like, it seems like this person might not be acting for the, for the mission.

• 18:09Garry Tan?

  YC's next batch is now taking applications. Got a startup in you? Apply at ycombinator.com/apply. It's never too early, and filling out the app will level up your idea. Okay, back to the video.

• 18:23Jared Friedman?

  Hey, tell us about the early days of Anthropic. So the, the seven of you broke off from OpenAI. You had a general idea of this sort of like, Long-term mission that you wanted to do to, you know, not destroy humanity. But like, how did, what did you actually work on for the first year? How did that converge on an actual product?

• 18:42Speaker 00

  So first year, the main thing that I tried to do was just build the training infrastructure that we needed to train a model and then get the compute that we needed to train the model. Those were like my two main projects. All the other things that you need to do when you're like, starting up a company too. So like, set up a Brex account and like, I don't know, like all of that, all of that stuff. We started out with seven co-founders. Within like a few months, I think like 25 folks from OpenAI overall had joined. So we had like a pretty substantial team that like already knew how to work together too. And so that helped us get up and running faster.

• 19:18Jared Friedman?

  And at what point did you launch the first product? And when did things begin to actually start working?

• 19:22Speaker 00

  So the first product that we launched was after ChatGPT. We had like a, maybe nine months before ChatGPT, we had a Slack bot version of like Claude one.

• 19:33Garry Tan?

  Oh yeah, we had that in the YC Slack, actually.

• 19:36Speaker 00

  Yeah, yeah, yeah. Yeah, I remember like Tom Blomfield adding all of you guys to it also.

• 19:42Garry Tan?

  That's really cool.

• 19:43Speaker 00

  And then I think that at the time though, we didn't know whether or not we wanted to launch it as a product. We didn't know if doing so would be good for the world at the time. I think we hadn't really thought through our theory of impact that much for like how we actually will make stuff work well. Plus, I think actually in hindsight, like if we had tried to launch it, we like wouldn't have had the serving infrastructure to have done it. And I think because we weren't sure whether or not we wanted to, we like hesitated for too long on building that infrastructure, which I think is learning for, for me.

• 20:13Garry Tan?

  I mean, at this time, ChatGPT had not launched yet.

• 20:15Speaker 00

  ChatGPT hadn't launched. And so I guess we didn't know that it would be a big deal too.

• 20:18Diana Hu?

  This is around the pandemic, 2022?

• 20:21Speaker 00

  This is summer 2022, yeah. And then ChatGPT launched fall 2022. And then we, we launched our API after that and then Claude AI after that also. I think it didn't seem like it was working basically until Claude 3.5 encoding. I think like really, really like through that whole time that. Until about a year ago, it seemed like it wasn't clear that we were going to end up being like a successful company.

• 20:48Diana Hu?

  We just saw that in the startups because we kind of get a bit of a vibe check in terms of what is the preferred model for startups. So all of 2023, OpenAI, OpenAI was the response.

• 20:59Speaker 00

  Yeah.

• 20:59Diana Hu?

  Then things started to turn in 2024 is when we saw Claude 3.5 and especially Sonnet.

• 21:07Speaker 00

  Yeah.

• 21:07Diana Hu?

  Was starting to get market share per se in the YC batches, going from single digit to at some point like 20 and to 30%. And especially for coding.

• 21:16Speaker 00

  Yeah.

• 21:16Diana Hu?

  Became the default choice, which was very interesting. Can you tell us about how that emergent behavior and the spikiness on that particular skill?

• 21:25Jared Friedman?

  Must be 80% now or 90.

• 21:27Diana Hu?

  Yeah, for coding even more, especially now Claude code. What was that? Was that on purpose or just kind of happen?

• 21:33Speaker 00

  I think that we invested more in trying to make the model really good at code because we wanted the model to be good at code was one thing.

• 21:41Diana Hu?

  Then you did it.

• 21:43Speaker 00

  And then I think seeing, seeing the reaction of everyone to it was like, okay, yeah, like let's go much harder on that also.

• 21:50Jared Friedman?

  And this is before 3.5 Sonnet. You'd already invested enough in coding to realize that that was really promising and you said, I decided to double down.

• 21:57Speaker 00

  I think this really was like individuals within the org being like, we want to do coding before 3.5 Sonnet. And then when we saw 3.5 Sonnet's really good product market fit, that was good signal to like, go, go for that.

• 22:08Jared Friedman?

  And do you guys know, like, the day that you guys launched 3.5 Sonnet, did you know that you had something really special? And this was going to be the turning point for the company? Or were you as surprised as OpenAI when they launched ChatGPT and it just like unexpectedly took off?

• 22:21Speaker 00

  Yeah, I wish that, I wish that we had like more foresight on that. But no, I think, I think it was surprising for us to like how, how big of a deal it was. And then I think 3.7 Sonnet also like surprised us by how much it unlocked like agentic coding. I think for, for each of these things, yeah, we move quite fast in rolling them out. And so we really often don't know what the results are going to be there.

• 22:43Diana Hu?

  I think it's what made a lot of these coding agent startups work. I mean, there's a crazy story of Replit winning, going to 100 million in just 10 months, right?

• 22:54Diana Hu?

  There's Cursor, of course, a story and all built on all these with Sonnet.

• 22:59Speaker 00

  I think that all of those things have been surprising to me. And then also, just like in my working with Claude too, like I think I continue to be surprised by like the type of stuff that it can do. And I do think with each one, there's like more stuff that kind of unlocks. But one of my friends was telling me that she had some code that she, some code source tool that she wanted to modify, but she didn't have the source code for it. She had the compiled binary and she was like, Claude, can you, can you decompile this? No way. Yeah, can you, can you disassemble the assembly? And Claude, Claude chewed on it for 10 minutes and like made a C version of it. And so then she had the thing that she can modify. It's insane. And she's like, yeah, like if I spent three days on it, I probably could have gotten the hex tables and like written a little code. But like, it did the whole thing, made up variable names for them, et cetera.

• 23:42Speaker 00

  So I do think that like, we keep getting surprised by stuff that model has memorized all the hex tables, it can think through, try to work through it. I think we're going to continue to be surprised by that sort of stuff, too.

• 23:51Jared Friedman?

  If you pull like the Y C founders, they prefer using anthropic models for coding by like a huge margin. It's much larger than what you would predict if you just looked at the benchmark results.

• 24:01Speaker 00

  Yep.

• 24:02Jared Friedman?

  So there, there seems to be some X factor that makes people really like these models for coding. Do you know what it is? And is it intentional in some way, or it just came out of the black box somehow?

• 24:13Speaker 00

  I think that the benchmarks, benchmarks are like easy to game. Where I think that all the other big labs, I think, have teams where they like, their whole job with the team is to like make the benchmarks scores good. And we don't have such a team. And so I think that, I think that that is probably the biggest factor. You don't teach to the test. We don't teach to the test. Cause I do feel like if you start doing that, then like it has weird, bad incentives. Maybe we could like put that team under marketing or something like that and then ignore all the benchmarks. But I think that that's one reason why there's some train test mismatch there.

• 24:45Diana Hu?

  So the evaluations are more qualitative internally? We have your internal benchmarks.

• 24:49Speaker 00

  We have internal benchmarks, yeah, but we don't publish them.

• 24:52Jared Friedman?

  And is it the internal benchmarks that the teams are really focused on improving?

• 24:56Speaker 00

  That's right, yeah. So we have internal benchmarks that the team focuses on improving. And then we also have a bunch of tasks. Like, I think that accelerating our own engineers is like a top, top priority for us, too. And so we do a ton of like dog food in there to make sure that it's helping with our folks, too.

• 25:12Garry Tan?

  Going back to Golden Gate Claude, there's a lot of sort of interpretability seems like it's a big part of it. And then most people would say that Claude's personality just feels better.

• 25:23Speaker 00

  Yeah.

• 25:24Garry Tan?

  And then how do you sort of at once be very quantitative, but then also build evals around personality?

• 25:30Speaker 00

  The evals for personality are kind of complicated, too. For like, how do you tell if Claude has a good heart or something like that? It's like hard to know. But I do think that that's like Amanda Askel's team's mandate. I think she describes it as like being like a good world traveler where Claude goes and talks with all sorts of people from different backgrounds and each of the people should come to that being like, I feel good about this conversation that I've had. Interprly, I think, is like a long-term bet, right? Where it's like, right now, the models aren't that scary, but at some point, they're going to be more scary. And so I think the hope there is to have some ability to know what's actually going on under the hood when it becomes more intense.

• 26:09Harj Taggar?

  Then more recently, Claude Code's been a real success. Can you talk us through like, how did that project get started internally? And again, was it like, did you know this time it was going to work or was it a surprise?

• 26:20Speaker 00

  Claude Code was an internal tool also. So like try to help out our engineers within Anthropic that Boris had hacked together.

• 26:31Harj Taggar?

  There's an internal Anthropic engineer wanting to build it for themselves.

• 26:33Speaker 00

  For other internal engineers, yeah, for him and other internal engineers. And then I think, yeah, I think we definitely didn't know that it would be successful out there. And I think to some degree, like, we really had fully just bet on the API before that with the intention being like, There's like so many startups out there with so many good ideas. Who are we to like, figure out what the right product is to build on top of this stuff? Everyone out there is going to build better stuff than us. And so put all of our effort into just making the best possible API. And I think that this surprised me as like, okay, like we actually were able to make something that like as a product was like better than the other products out on the market for this adentic use.

• 27:14Speaker 00

  I have like some theory that like, part of that came from like a mind shift of seeing Claude as like the user for this thing too. For like LinkedIn, we were like trying to build things for teachers. We're like our users for Groupon. It was like single people in New York mostly, I guess. For this, I think really the like users are the developers, but also I think the users is Claude. It's like, give Claude the right tools that Claude can actually do that effectively. Help Claude get the right context to work effectively. This team was like the most focused on Claude as like a user, which I think is like a bit weird.

• 27:50Jared Friedman?

  It makes sense that you guys would understand Claude the best.

• 27:52Speaker 00

  Yeah, I do think that that's a place where like startup founders though, like can do that too. And I think that that's probably a rich vein for people to like make tools that are better for models as users.

• 28:04Garry Tan?

  That's the perfect anthropomorphization of like the LLM itself. Like the agent is one of the stakeholders, is one of the users that you would go after and try to like empower.

• 28:13Speaker 00

  Yeah, yeah, totally.

• 28:14Diana Hu?

  Which actually makes a lot of sense why you guys actually got MCP to work, to do tool calling. Because a bunch of other labs had tried to do something and the standard that stuck, that really took off was yours.

• 28:29Speaker 00

  Yeah, I think that that seems like a similar one too, where it's like a model focused.

• 28:33Harj Taggar?

  Going back to Claude Code, so like success is really exciting. It's also scary for like Cursor and other companies that have built on top of the API. Like, what's your advice to founders building products? Like, how should they think about building on the API, but also worrying about like Anthropic or one of the labs building something better than they can build?

• 28:51Speaker 00

  I think I was kind of surprised that Claude Code, like we did build a thing that was like, like the best in the market there too. It's not super clear to me what the. Big advantage was for us for Claude code besides more empathy for Claude.

• 29:05Harj Taggar?

  I think that's actually really interesting insight. Like it seems like the thing that, yeah, you were building for a specific user that you knew really well that other people wouldn't have thought to build for versus like you had some like intrinsic technology advantage.

• 29:17Speaker 00

  Yeah, like I think a startup could could have done that same thing too, right?

• 29:19Harj Taggar?

  Yeah.

• 29:20Speaker 00

  I think we're the most like developer focused lab. I think we're the most like API focused lab too. So I think we want to make sure that we have the best platform for people to build stuff on because this thing is growing so incredibly quickly. Like we're not going to be the fastest at figuring out all the ways that we need to empower Claude to do the work that connects Claude to the entire human business. That's like human, human world is all designed for humans, but like we need to get the models to be able to be productive members of the economy.

• 29:49Harj Taggar?

  Are there like ideas or areas you would love to see developers building in or like areas you don't you you think are like underappreciated right now?

• 29:58Speaker 00

  Yeah, Claude code is like, how do you get Claude to be a useful pair programmer kind of? Or like junior engineer, you've got like a suite level two or three or something like that that you can work with or like very spiky because also it can do the like weird disassembly stuff that like a super high level suite would struggle with less good at knowing what type of work to do. Needs kind of a lot of handholding, needs a lot of context from it. That's like one very particular subset of work that can be done. If you look at like all the stuff that happens in businesses, besides that, it's like a very tiny fraction of like all the work that's done in

• 30:38Speaker 00

  businesses that like a smart person who knows how to code and like use lots of tools, but doesn't have that much context yet would want to do. So I think I think finding ways to coach Claude or coach whatever model to like do useful tasks for businesses seems like there's just like a huge, huge space there.

• 31:01Jared Friedman?

  So Tom, a big part of your job is like. Owning all the compute infrastructure that makes Anthropic work. Can you talk about like, what is the compute infrastructure behind this giant thing now?

• 31:11Speaker 00

  One thing that's interesting to look at is just that humanity is on track for like the largest infrastructure build out of all time now.

• 31:18Jared Friedman?

  This is going to be larger than the Apollo project, larger than the Manhattan project?

• 31:21Speaker 00

  It'll be bigger than both of them next year if it keeps on the current trajectory, which is like roughly 3X per year increase in spending on AGI compute, which is just bonkers. Yeah, like 3X per year is wild. I think it's going to keep up on the 3X per year trajectory. It's already locked in for that for, for next year. And then it's a little bit open for, for 2027.

• 31:46Garry Tan?

  I mean, anecdotally, internal to YC, we can't get enough, you know, credits across all of the top frontier models, including Claude. So you gotta help us out a little bit.

• 31:57Speaker 00

  Yeah.

• 31:58Garry Tan?

  We're just, I mean, everyone's bottleneck, literally every, you know, it's like, give me more intelligence. I can't have enough.

• 32:03Speaker 00

  Yeah. And I know you guys have been looking at more hardware startups also for like more accelerators. I think that we will see more accelerators coming online to 2027. That's a good, a good space also, like data center tech, I think is a big one.

• 32:15Jared Friedman?

  Where are the bottlenecks for you guys now? Is it like getting enough electricity, getting enough GPUs, getting construction permits?

• 32:23Garry Tan?

  Power, people are using jet engines to get power. That's nuts.

• 32:27Speaker 00

  Overall, for the build out, I think power is going to be the biggest bottleneck, especially power in the US. Like we want to build in the US. That's one of our biggest policy goals is to like, get the US to like, build more data centers, permit more data centers, make it easier to build.

• 32:42Garry Tan?

  Is the answer renewables or is it nuclear?

• 32:46Speaker 00

  I definitely, I feel like, yes, yes, all of those things. I wish, I wish the nuclear was easier to build.

• 32:52Jared Friedman?

  And Anthropic is the only major lab that uses not just one kind of GPU, but the GPUs from three different. Manufacturers, can you talk about that and how that strategy has played out?

• 33:02Speaker 00

  Yeah, yeah. So we use GPUs, TPUs and Tranium. Downside of doing that is that we split our performance engineering teams across all of those platforms, which is a ton of extra work. The positive thing is it gives us the flexibility to both, one, like soak up that extra capacity because there just is more of those altogether than just one. And then two is we can use the like right chips for the right jobs, where some chips will be better for inference, some chips will be better for training, and we can match the right chips to the right jobs. So yeah, I think that that's kind of the trade off there.

• 33:36Diana Hu?

  I guess one cool thing is just connecting the dots through your career and how all of this compounded. Because you were the one engineer building that change of the architecture from TPUs to GPUs back at OpenAI that got GPT-3 to actually scale. And now you're in charge of that at a much, much bigger scale years, years later. I don't know if that kind of connected dots for you.

• 33:59Speaker 00

  The big move from TPUs to GPUs at OpenAI, I think, was partly driven just that PyTorch was a better software stack on top of them than TensorFlow on top of TPUs. And I think that that then unlocked fast iteration where like, if you have like a good reliable software stack, then you can experiment quickly, just like build a whole system that works. I think that that's the thing that we really strive for now at Anthropic too, is a challenge of having many more platforms is that it's harder to write all the good software. I think building the muscle of knowing how to build that software well so that all of the people who build on top of that low level can have a great experience with it is the most important thing there.

• 34:38Diana Hu?

  Do you have advice for kind of like a younger Tom version of yourself, who now you've seen and went through this crazy journey? If someone was you back in the twenties, living today, and they wanted to ride and join the AI revolution, what would you say to them?

• 34:54Harj Taggar?

  And very specifically, something we see from a lot of here, from a lot of college students at the moment is they They don't know what like if they should stay in college, like are there going to be jobs for them?

• 35:03Speaker 00

  Yeah.

• 35:04Harj Taggar?

  What like, how is the world going to change and what should they do?

• 35:07Speaker 00

  Taking more risks, I think, is is wise. And then also trying to work on stuff where your friends would be really excited and impressed if you did it, or a more idealized version of yourself would be really like proud of yourself if you succeeded at it. I think it's like probably the thing that I would I would try to tell a younger version of myself.

• 35:27Garry Tan?

  More intrinsic, less extrinsic, like don't chase these other credentials and getting the degree or whatever, you know, working at Fang, like those are just

• 35:35Speaker 00

  Yeah.

• 35:35Garry Tan?

  irrelevant.

• 35:36Speaker 00

  Yeah.

• 35:36Garry Tan?

  As of today.

• 35:37Speaker 00

  Yeah, exactly.

• 35:38Garry Tan?

  That's all we have time for today. We'll see you guys next time.