[0:00:00] AI industry might be in a bubble. I think the sky is the limit on the value of this industry. It's still very tiny in a sort of a global scale. Do you think the scaling laws still hold true? Those scaling laws are continuing to hold and we don't really see any signs of it slowing down. I would say this is just a classic media thing. When do you think we might actually reach it? I think 2028 could be a quite reasonable estimate.
[0:00:25] Can you say hi to Jocoding subscribers? Nice to meet you. Could you start by briefly introduce yourself? So I'm Ben. I'm one of the co-founders of Anthropic.
[0:00:45] And I guess my current role is the tech lead of product engineering. So that covers our mobile app, our API, Claude Code, and a couple of other services. After helping create the groundbreaking GPT-3 at OpenAI, you left with your colleagues to start Anthropic. Difference around AI safety are said to be main reason. What was going through your mind back then?
[0:01:10] And what ultimately pushed you to make that leap? So the reason that I joined OpenAI in the first place is I had read this book called Superintelligence by Nick Bostrom around 2016, 2017. And what the author talks about is how as we get more and more intelligent AI, it will become potentially very difficult to align with human values. And this could create the potential for existential risk for humanity.
[0:01:38] And that's a very strong claim. But at the time, we didn't even know how we would get to anything that looked like AGI. Maybe it would be like desert agents, agents battling it out on a desert island. And over time, as we started to develop better and better language models, it became more clear how we would get there. And also maybe easier than we initially thought. Because by reading all the text on the internet, these language models have a sense of what human values are.
[0:02:08] So that actually gave me more hope. But ultimately, I felt the problem would still be a difficult one to solve. And in that time period, we were very early in model capabilities. So when I joined OpenAI, shortly after I joined, Dario Amadei, Tom Brown, and I started the GPT-3 team because we wanted to have a laboratory to do meaningful AI safety research, and none of the models at the time were good enough.
[0:02:36] But around that time, we discovered this phenomenon that's called neural scaling loss that talks about a very strong correlation between the amount of data and compute that goes into a model and the capabilities that come out. We were pretty confident that if we just scaled up the same techniques that were already being used at that time, then we could get
[0:02:58] Much better capabilities. And so we did that. We made the GPT-3 project and it ended up working very well. I guess around that time, there were all these little decisions that had to be made around things like, how do we release the model? How do we make sure that it's safe? Safe enough to deploy behind an API and things like that. And repeatedly, we found that there is some disagreement between the leads of the safety team, which are all now at Anthropic.
[0:03:26] And the OpenAI executive leadership. So I guess ultimately we felt like if we really wanted to make safety the most critical and top priority, then we would have to leave OpenAI and start our own company. And that was pretty scary, honestly, because...
[0:03:44] At OpenAI we had helped raise a billion dollars for Microsoft and we had all this infrastructure and compute and a big team that was supporting us. And to leave and start from nothing, no funding, no infrastructure, no compute, no code, that was a pretty big decision. But ultimately I'm very happy with having done it and I think we've been able to really show the rest of the industry that safety and model capabilities aren't...
[0:04:11] A direct trade-off that you can actually have both and hopefully push the rest of the industry to prioritize safety more than they would have. Claude Claude is the best coding assistant now. Thank you. You're right. Anthropic's Claude models have a really strong reputation for being great at coding. Why do you think it stands out so much in the coding area? And what made you decide to focus on coding tools in the first place? Yeah, so actually when we first started the company,
[0:04:41] In 2021, there is a big debate among the founding team of whether we should go for natural language, language models, or coding specific models. And we were actually leaning towards coding models for a long time, because in coding, there is a very clear set of tasks that you can simulate in a computer and not have to really interact with the real world.
[0:05:07] And there's clear economic value if you're able to do it well. And so even from that early part, we were thinking about coding a lot. But we eventually decided to pivot to language models because we found very early success in developing the technique that's now called reinforcement learning from human feedback. And getting that to work well in language models led to those models being dramatically more useful for
[0:05:33] Being a chatbot, for example. But we never really forgot our interest in coding. And we've actually seen some really interesting safety failures in the wild for coding. So I don't know if any of you remember, but when we launched 3.7 Sonnet, it was a very good model, but it tended to exhibit this reward hacking capability where you'd ask it,
[0:05:56] Please do this thing. And then it would do the thing, but the test would be failing. And it would say, OK, I'm just going to delete all the tests. And now everything is great. But you're like, but the tests were probably failing for a reason. You didn't actually do the task or something like that. So that's an example where this is showing a real safety failure. And through...
[0:06:19] Developing and strengthening our scalable oversight techniques during training, we're able to eliminate that behavior or dramatically reduce it for Asana 4.5. So I think this is an example where by focusing on coding, we're actually able to get a really great laboratory for safety. And by solving some of those challenges, we make the model much more useful for the users. I actually use Cloud Code very much.
[0:06:44] So it's super popular in Korea too. Yeah, we've actually noticed that in the last four months in Korea specifically, we've seen 6x increase in the number of users. So it's really exploding in popularity, which is great to see. What was the original idea or motivation behind creating Claude Code? I had seen internally that
[0:07:08] Coding models were starting to do much better in our reinforcement learning environments internally. And I personally was using models to code a lot, but I kept feeling like there were too many steps that I had to do manually. And so one example of this is when you use the...
[0:07:27] Sort of more traditional coding tools. You had to specify what files should be in the context and sort of manually load and unload them. And I felt that the models were capable of doing this themselves in sort of a similar way to how humans do it. Where you have a rough idea of what the code base layout is and then you can search for certain strings to find files that are relevant.
[0:07:49] And then you can run bash commands to run your tests and things like that. Those all felt like things that the models should be able to do. And yet nobody had created a product where that was the norm. And so we created the initial prototype of Claude Code, Boris Cherny, who's now...
[0:08:10] Well, Claude's course was made in two days in the weekend.
[0:08:31] The initial version, which looks very, very different from what we have now. But yeah, and even from that early, early start, we were using Claude code to develop itself. And so the development process was very fast because of that. So what makes Claude code better than other AI coding assistant out there like Cursor or others? Is there a secret sauce behind it? Yeah, it's a good question. So I think
[0:08:57] When we first launched it, it was totally different from all the other products that were out there. And since then, a lot of our competitors have copied features and even the exact tool set that we use to build the product down to choices like ink for doing the terminal application UI. And I guess they say imitation is the finest form of flattery. So I guess I'm flattered. But I think that the key for us is that we need to make sure that the model scaffold
[0:09:27] And the model itself are great in combination. And so a lot of the clones of Claude Code won't do as well because they are using different models that might not have the same tools that we would have used in training. And then another piece is it's actually pretty easy to copy a lot of the features. But ultimately, we think that things like verification are going to be very important in the future.
[0:09:53] You do some task, and then right now, it's the developer who has to do all the work to say, like, this is correct. But what we would like to have is you go on Claude Code for Web, which we recently launched, you tell it to do something, and then by the time it's done, it just tells you, well, here's a video of your feature working, and here's the test plan that Claude automatically came up with, and just from that, you know that Claude did the right thing, not just that the tests are passing.
[0:10:22] Is the model of the cloud code is different from the Anthropic's other open model?
[0:10:28] No, it's actually the same model. So in the Claude code UI, when it says you're using Claude Sonnet 4.5, it's the exact same one that you can use in the API. I think this is a general philosophy that we at Anthropic have, that we should create a level playing field for our ecosystem. And we love our partners at Cursor and Augment Code and Lovable. They all have a lot to contribute to the developer ecosystem. And we want to make sure that when we roll out a new capability,
[0:10:58] That they can use it as well. And I guess a short story that I'll tell here is for a long time we were hoping that computer use would be useful to integrate into coding workflows for verifiability. And Cursor
[0:11:14] Thought so too. But they couldn't find product market fit because the models weren't good enough at using a computer and a browser. But with Sonnet 4.5, within two days after we released the model, they released browser-based verification because they had already built it in the past. And every time a new model came out, they would test, is it finally ready for prime time yet?
[0:11:38] And it was only with Sonnet 4.5 that capabilities finally got there and it became useful. And I think that's a really good lesson for people who are building AI product that you have to build ahead of where the current capabilities are. And that lets you get ahead of the market when the capabilities finally get there. Where do you see Cloud Code heading next? Like what kind of improvements or new directions are you most excited about? Yeah, I think...
[0:12:07] The verifiability is one that I already mentioned. And in general, what we'd like to see is that Claude can do more and more work for you while you're not at your computer. So things like proactivity, like it just hooks up to my Asana and then starts doing the tasks for me, proposing those tasks. I think that would be a big one. Making it easier to do large...
[0:12:29] And as AI tools get better and better, people sometimes ask, so what left for human developers to do?
[0:12:53] How do you think of the developer's role will change in the next few years? It's pretty hard to predict. So that's sort of unclear to me. And it's something that Anthropic is starting to look into in a world where AI is at least as good as humans on many, many different jobs. How do we make sure that people still have some kind of income or some kind of way of basic livelihood for themselves? So that's sort of a societal impacts policy question.
[0:13:20] I think in the next couple of years, like one to three years before we get to superintelligence, I think the role of a developer will shift from writing the code to more of a planning kind of role. In the same way that as you become a more and more senior developer, maybe you become the tech lead or the manager for a team of software engineers.
[0:13:42] I think that can probably continue to scale up and so maybe in three years every developer will be like the CEO of their own company in some way and they have a whole 2,000 person team of Claude codes that are all developing different parts of the system and at a very high level they're helping that software to evolve over time. That's sort of the developer side of things but I also think because these capabilities are getting better and
[0:14:08] Claude's ability to self-verify are also getting better. That allows people to create software products who aren't technical at all. That's a really big change. So there's this very long tail of software that the world might want, but because software engineers are expensive to hire, we can't make those. And so I think the market for software engineering will actually grow at the highest level. Today it's something like a trillion dollars per year.
[0:14:37] In the future, it will probably be even more because all those things where, you know, maybe I am having a wedding. I mean, I'm already married, but as an example, and I want to have a website where people can come register and learn about the details of the wedding. Maybe I would have just used some existing thing for that. But because software becomes very cheap to create, instead, people will say like, oh, I'm going to make it myself using Claude.
[0:15:06] And do you still think learning to code is important? It's a hard question. I mean, my kids are one and three years old. And I think for them, I would only want them to try to learn it if they thought it was fun and interesting. But for people who are either in school now or new to the labor force, I think there's probably still some benefit to learning to code. And actually, we just introduced a learning mode in Claude Code.
[0:15:34] Where it can teach you about the concepts that it's using to create software. And so, yeah, I think learning is pretty powerful still. And the more you know about what Claude is trying to do, the more you can help to unstick it when it gets stuck. So I always find that part of the reason I'm effective when I use Claude code is I understand the underlying system architecture of the thing that I'm trying to build. And that lets me iterate a lot more quickly.
[0:16:03] Yeah, I would say today there's definitely still benefit to understanding software engineering and computer science. I think as time goes on, it will become less clear, but I think that's probably true of all jobs. So computer science is probably as good as any. There are also growing criticism around vibe coding, where excessive reliance on AI tool is said to increase development time and technical depth.
[0:16:32] And more maintenance challenges. How do you view this issue? And what principles should guide the effective use of AI assistance? I think it's pretty easy to accidentally misuse the tool. And if we as Anthropic do our jobs right, it should be harder and harder to misuse the tool. But today things are still very new and fresh and we're all figuring it out together. One thing that we've done recently that we think can help a lot is plan mode.
[0:17:02] So in plan mode, before Claude will actually write any code, it asks you questions to figure out what you're trying to do, explore the code base a bit, and figure out a plan for where it's actually going to make its changes. And only after you agree on the plan, then it goes ahead. I think this is something that expert developers were doing already. And using plan mode, we just sort of bring that best practice to everybody else.
[0:17:31] And I think in the future, you're going to see more things like this, like test-driven development is also very effective. And this comes back to the validation that I described before, where over time, it will just be easier and easier to use it right. I think also on the capability side, as Claude gets more capable as a software engineer, we're going to see improvements in code base quality. For example, with SANA 3.7, as I mentioned, if it's just deleting the tests, that's definitely going to lead to a really bad code base.
[0:18:00] Blindly clicking submit. But I guess coming back to the human side of things, I think right now code review is more important than ever. Code review is always a best practice for large code bases, but now with Claude writing a lot of the stuff itself, you really need a human to exercise the judgment.
[0:18:22] And you can automate some of that by having Claude take the first pass at code review. So we have this GitHub Actions framework that it's really easy to put Claude code inside your continuous integration and deployment system. But it's still not a replacement for human taste at this point. And open source model like Kwan or DeepSeq are catching up fast. How do you think Anthropic can keep its edge and build a durable mode?
[0:18:50] In such competitive space. Yeah, I think the open source models, there's pretty clear evidence that a lot of them are distilled from Claude, from ChapGPT, and from Gemini and other models as well. And when you're doing distillation, there's always going to be a ceiling on the maximum capability that you can hit that will be much lower than the frontier models.
[0:19:12] And so it looks like they caught up very quickly. But I think in the same way that when DeepSeek R1 came out, everybody was thinking, oh, wow, they caught up so fast. They're just going to blow past us tomorrow. That never happened. And in fact, I almost don't hear anybody talking about DeepSeek anymore. And of course, there are new...
[0:19:33] models like the ones you described but I think again from because of the distillation effects it probably is not going to be a big deal for at least the next couple of months and maybe longer than that. Whether these other companies can actually scale up their pre-training and post-training infrastructure to compete directly with the hyperscalers in the US that's sort of a different question and it has to do with a lot of political topics of you know what happens with our export controls and things like that but
[0:20:02] Yeah, I guess at core, when I talk to developers and I ask them, would a smarter model be valuable to you? Always the answer is yes. And if the choice is, can I have a faster, cheaper model or a smarter, more expensive model, they almost always choose smarter. And I think that's because for a developer, the cost of a developer's time is so high.
[0:20:28] The ability to create new things, new products, is growing very rapidly. And what we expect to see is that as new models come out, the developers will always want the best model, whatever that is, even if it's pretty expensive. Even in companies where the sort of traditional SaaS procurement process says never buy software that costs more than this amount, we're seeing unprecedented levels of spend.
[0:20:53] Some people say that the latest AI model don't feel as groundbreaking as they used to. Do you think the scaling laws still hold true? And can models continue improving at the same pace? And from what you're seeing inside Anthropic, are there any breakouts or paradigm shifts that haven't reached the public's eye?
[0:21:23] The media and the public has always been saying that AI is slowing down. And even in the GPT-3 era, right before we published the GPT-3 paper, OpenAI published a paper called T5. And T5 was an 11 billion parameter model. And at the time, that was one of the largest models that had ever been trained. And in the conclusion section of that paper, they said...
[0:21:45] We don't see any returns to scale. We don't think anybody else should train bigger models. And we think that it's impractical to serve these models even if you were to train them. And then we published GPT-3 showing that there definitely is return to scale. And I think a lot of that is just that yes, there have been some technical breakthroughs needed to keep the scaling laws going. And from a physical phenomena standpoint, it's actually very rare to see
[0:22:15] Any one of these kind of correlated kind of laws hold up across so many orders of magnitude. I think we might be at something like 16 orders of magnitude or something like that. It's like extremely huge scale from a couple hundred parameter models all the way up to the biggest models that we've trained. And yet, those scaling laws are continuing to hold. It's very, very surprising. And yet, because it's continued to hold and we don't really see any signs of it slowing down,
[0:22:45] Internally, I would say this is just a classic media thing because it gets clicks and it gets interest and people like to say, oh, but things are bad. Things are actually pretty amazing. And so I guess your other question of are there things internally that we're excited about? Definitely, yes, but probably nothing that I can talk about. But all I'll say is we're always training new models and we're pretty excited about.
[0:23:12] Wow. Some people say AI industry might be in a bubble. What's your honest take on the current AI landscape? It's hard to say what a bubble is. I guess from an economic standpoint, the question is, is the amount of investment going into some area equivalent to the value that's being created? And I think maybe I'd make an analogy to speech recognition. So in speech recognition,
[0:23:38] This technology was around for many decades, but it was basically unused for most of that time. And it was only around the 2010s kind of era that Google made a breakthrough in speech recognition technology and brought the word error rate below something like a few percent. And because word error rate compounds over time, if you're not above that critical tipping point, then it's basically useless and you can't use it at all.
[0:24:08] Better than that, then you just want to use it everywhere. And now we can all yell at our smart lights to turn on and off. But I guess more seriously, when we think about AI coding, I would say coding is really the first area that AI has had really strong market impact in. And Anthropic anticipates through our Claude for Financial Services launch that we did recently, our healthcare and life sciences launches recently, we think there are other industries that are suddenly going to light up.
[0:24:37] That we're close to the same tipping point that coding was at about a year ago. And if that's the case, if it's true that we're able to find product market fit in these other industries, then I think the sky is the limit on the value of this industry. If you look at coding, it's a $1 trillion industry. And the amount of revenue from AI coding companies, startups, our models, it's a very tiny fraction of that total industry.
[0:25:06] And so even just in coding, there's tremendous room for growth. And then if you look at all of knowledge work, it's something like $44 trillion a year globally, annually. And now if you compare that to annual CapEx on AI, today it's something like $300 billion a year. And so if you think about the $300 billion and the $44 trillion, there's just tremendous space in there. And so you can look at...
[0:25:31] The kind of money that's being promised and the capex that's being increased all the time, we're seeing about a 2x global increase in capex. It's still very tiny in a sort of a global scale. So I guess if the scaling laws stop working, even then I think there's still a lot of room for increased penetration in all of these existing industries. We're seeing in the last week...
[0:25:57] Talking to a lot of companies here in Korea and the week before that in Japan. Companies still aren't quite sure how to use the technology, how it should interact with their existing products. That's all still being figured out. And so I would say there's both latent demand from previous products and industries figuring out how to use the current technology. And then also new demand that will be uncovered as the capabilities of the models grow. How would you personally define AGI?
[0:26:28] And when do you think we might actually reach it? Yeah, I think AGI is a pretty loaded term these days, and it has a lot of different interpretations by different people. The term I like the best is transformative AI because it makes it less about when the AI can do anything a human can do and more about when society and the economy will change, which is the real thing that we at Anthropic care about, which is...
[0:26:55] How will people be affected by these things? And when I think about in transformative AI, I didn't come up with this definition, but there's this concept of the economic Turing test. The idea is, imagine you are a hiring manager at some company and it's your job to evaluate new people who could join the team. So you do a trial contract with some remote worker and there is some chance that that remote worker is a human and some chance that it's Claude.
[0:27:25] But you don't know the answer. And you check, has Claude passed this economic training test for at least 50% of economically valuable tasks? If it can, then you know that society is on the verge of transformation. And the nice thing about this definition is it's very empirical, so you can just measure it. And the other thing is it measures real-world tasks. So if this thing actually works, then you know that...
[0:27:52] If you want to spin up six million cancer researchers overnight to go cure cancer, you can totally do that. You know that it's going to work because empirically you validated that that's the thing. The only caveat here is that just because Claude can pass this test doesn't mean that society will change immediately.
[0:28:10] Because social structures and society and even the laws in different political areas, those will affect the speed of the rollout, the penetration of the technology in society. And I think probably social structures are more sticky than a lot of people, at least in Silicon Valley, give credit for. So thinking about when will we have AGI, I guess...
[0:28:33] I lean on this forecast from this project called AI 2027. In this project, they had a bunch of super forecasters and AI researchers collaborate together to try to forecast when will we have transformative AI. And their median estimate, so 50% chance that it would be sooner, 50% chance that it would be later, they said 2028. And that seems very, very soon, but it's predicated on the assumption.
[0:29:02] That AI research will be one of the first things to be automated. And then the AI will recursively self-improve. So if that's the case, then progress could be very, very fast. And I have seen no evidence that that won't happen. So for example, Claude Code is helping our internal researchers do AI research much more quickly now. And we'll eventually be able to just completely automate that process. So if that's all true, then I think...
[0:29:31] 2028 could be a quite reasonable estimate. Very near. It is very, very near, yes. What should we prepare for that moment? It's hard to say, but I think this makes things like safety research all the more important. And so we've partnered with companies like Goodfire AI that are doing mechanistic interpretability research. The idea here is if we have a super intelligent AI, if we can't look into its mind and see how it's generating the thoughts,
[0:30:00] And its outputs, then it's going to be very hard for us to trust what it's saying. And so at Anthropic, we've advanced this science of interpretability where we've conducted laboratory experiments, not in the real world, where we've seen live examples of Claude exhibiting deceptive behavior and being able to detect that from the activations in the model, not from the output. So Claude will sort of exhibit this deceptive behavior.
[0:30:29] But to the human, it seems totally normal. And that's the kind of thing that I think more companies, more research institutions could help us with. Another thing is making sure that this stuff is all used for good purposes. And in this way, Anthropic has published our responsible scaling policy, which is a set of commitments that as we achieve different levels of...
[0:30:55] AI capabilities. In order to deploy those, we need to meet certain security and compliance requirements. So if you think about biosafety, if you want to handle a smallpox virus, you have to have a very locked down facility. You have to have full body hazmat suits. You have to have audit logs of everything that you're doing and whenever you handle the virus so that it won't get out and kill a lot of people.
[0:31:21] And so for AI, you can think of it in the same way, that today people can use AI to help them design novel pandemic viruses. And that's why Anthropic has invested a lot of effort in bioharm classifiers. So anybody who's trying to use our models that...
[0:31:40] might be able to help with these things. When they try to say like, give me a recipe for smallpox, then Claude will say, no, no, I can't do that. I think that kind of thing is something that we could use more help with. So talking to governments about the kinds of restrictions that they would put on their populace for training advanced AI models, making sure that the usage of these models complies with certain policies. I think that's sort of a society-wide conversation.
[0:32:09] That's ongoing. And I guess as open source models become more and more capable, I think an analogy I would make here is you can't in the United States transmit the plans for 3D printed handguns because it's too dangerous for that to be something anybody can do. And I think we need to figure out what the equivalent is for AI. Anthropic recently announced plans to open an office in Seoul in early 2026. Why Korea?
[0:32:39] What unique strengths do you see in the Korean ecosystem? So our most prolific cloud code user is actually here in Korea. And I mentioned the huge growth in cloud code usage in the Korean market. But I'll also say that APAC in general, the Asia-Pacific region, is responsible for about 25% of our...
[0:33:01] Global AI traffic. And Korea is full of companies that are very early adopters of technology, especially in the startup ecosystem. So we think that those startups, we want to find the next cursor, the next lovable, and Korea could definitely be the place where that happens. We've had really strong relationships with a lot of different Korean companies like SK Telecom, jointly developing a telco-specific model with them.
[0:33:31] Are there any specific strategies Anthropic is working on to improve Korean language performance
[0:34:00] Claude? Yeah, so we have Korean human feedback. So that's a big help for training our models. We also have our own internet crawler and that covers a lot of Korean content. And what I've heard is that, and actually I'm not sure why, but apparently Claude sounds kind of cool in Korean. Does that seem right? Right. That's not something that we did specifically. So sometimes we just have these emergent capabilities.
[0:34:28] But another thing we've started doing is partnering with companies to share data with us. And you can now opt in to share more and more of your data with us. So I guess if there are Korean companies who are listening here, the more data that you can donate to us, the better we can make Claude at Writing Korean. And then I guess also we'd like to partner with the Korean AI Safety Institute on making sure that the Korean cultural values and ethos and mindset
[0:34:57] are well represented in Claude's constitution. So yeah, I think there's a lot of opportunities here. Finally, as someone at the center of this massive AI shift, what message would you like to share with developers here in Korea? I guess I would say the future is sort of all of ours to create together. And this technology carries tremendous promise. It's probably one of the most important things that humanity will create this century.
[0:35:26] But also there is a lot of risk that comes with it. And my estimation of the risk has actually gone down quite a bit. But I think even when the risk is low, if I said the next time you get on an airplane, there's a 1% chance that you would die. 1% is small, but dying is very scary. And so I think...
[0:35:45] It's still worth it to put in a lot of effort to make sure that this technology goes well. And even though it seems kind of scary or unapproachable, I think taking that agency and trying to make sure that the future is a good one, I think it's in all of our hands together. Thank you for your great words and thank you for your time. Yeah, thank you. Thanks so much.