Anthropic

“Ultimately, you need, like, really great machine learning engineers. Like, that's where it starts. And if you don't have that, then you know, you gotta find it hard to build truly, really, truly great things. You know, so, like, what OpenAI provide what Entrop…”

“I mean, I talked to Claude 20 times a day. You know, I just have long conversations with it. I have tons of projects loaded up with all kinds of contexts on different topics. And I literally could not do the job I do now without having a conversation partner l…”

“One thing we're trying to get better at, and that's also letting the product be like educational in a very literal way, which is like the thing we did not do early and now we're changing is just tell Claude more about itself, which was like, you know, it's in …”

“I think prototyping with these models is a thing that is underused. Like, our best PMs do this, where we'll get in some long conversation about, like, should the UI be this or that, and before our designers have even, like, picked up their Figma, like, our, of…”

“And, you know, the stuff that's happening today that we're working on, that you guys are working on, it all feels like magic. 12 months from now, we're going to be like, can you believe we use that garbage? Because it's going to, I mean, that's how fast this t…”

“It's the thing it most reminds me of like from the Instagram days where like Apple, like WWC announcements, you're like, this could either be awesome for us or could like absolutely like cause chaos for it. It's like that, but your own company is the one kind …”

“I mean, not many people could, but like, this is like probably a list of three companies that would have been interesting.”

“And some of these model these evals we've seen, like, even the golden answer, I'm like, I'm not sure a human would say it, or, like, I think that math is actually a little wrong, like, getting a hundred percent is gonna be really hard, because even just gradin…”

“Some early things that we're seeing that we think are really interesting, one is UI testing, which is like, I was, like, at Instagram, we had basically no UI tests, because they're hard to write, they're, like, they're brittle and they're, like, often, like, a…”

“The, you know the most important model that we use is the Sonnet model from Claude from Anthropic. And it is the best model at coding.”

“When the Anthropic dropped the computer use sort of capability. You know, we slaughtered our roadmap because we don't really have an explicit roadmap. We like immediately jumped on it and started building things and we've launched some things around it. We're …”

“They're trying to make them use human interfaces like Anthropix computer use, but those are really expensive and you need to kind of process all those images and video.”

“ChatGPT's reputation with me is not good because it regularly over promises and under delivers, and it does it without kindness or humility. Whereas Claude, I don't know how they did it, at least in my experience, brings kindness and humility.”

“And it's like the reason why cloud has so much more personality and like is more like a librarian. I don't know. Like, I don't know. I am like visualizing Cloud being, like, a librarian. Like very, like, nerdy or something. It's because I feel like it's a refl…”

“I would say, like, Antarctic, I learned from Antarctic that, like, They're much better at, like, focusing and, like, prioritization or, like, very, very hard, like, very hardcore prioritization, I guess, and they need to do it. Like, but I think, like, OpenAI …”

“Smart, small models are becoming even smarter than, like, large models. And that's because of, like, the distillation research. This happened with, like, Cloud Tree Haiku. I was like working on like post-chaining of like Cloudy Haiku, and I realized it was muc…”

“One of the things that I've learned early days at Anthropic was, like, we've discovered, especially with, like, cloud three training, when you taught the model some of the self-knowledge of, like, hey, like, you actually don't have a physical body to operate, …”

“The servers were just melting anthropic. Ran out of GPUs for us. Like Dario emailed me. He was like, listen, we just, we don't have anything more to give you at the times”

“Software is deterministic. When you write code and you hit run, either runs or it doesn't. And that is why that's the key insight Anthropic really had. They just went deep. And this is what they're doing. It's just reinforcement learning. I'm basically permuta…”

“Sonnet was, you know, Sonnet was the, really the first model that flipped the equation because that was really us and all cursor and all these other things. The rapid growth started the second sauna went online, right? Like we actually tried building bull like…”

“It used to be that OpenAI had this, like, massive model lead, you know, 12 months or something ahead of everybody else. That's not true anymore. You know, I like to think we still have a lead. I'd argue that we do, but it's certainly not a massive one.”

“JavaScript is almost, in my mind has always been almost like the English of programming languages. It's a language that if you learn it, you reach billions of devices. So it's not a coincidence that when you ask ChatGPT or Anthropic or Gemini to build you a we…”

“We have custom models that we built that are post trained on top of popular open source models that can make these edits really, really quickly to the code base. And the reason why you would want to do that is it's a faster and be also that model can actually …”

“I love Claude code for developing, even though I have my complaints about it. There's something just very addictive about just telling it what to do, and it can, it's basically something that you run within the terminal. Of your computer, and so you just type …”

“Particularly three, seven, four has less of the three, seven really like to ask follow-up questions and we call it elicitation. And sometimes it'd be like, I don't want to talk more about this with you, Claude. I just want you to like go in and do it.”

“To really have agency and have these agentic use cases. Like one way you approach it is computer use, but computer use has a bunch of limitations. The way I get way more excited about everything is an MCP and our models are really good at using MCPs. All of a …”

“I don't think we're gonna buy a cursor. Cursor's very big but we love working with them.”

“Where my go-to product strategy partner is Claude, and it has been basically for that full year where I'll write an initial strategy. I'll share it with Claude basically, and I'll have it, you know, look at it. And in the past, it's pretty anodyne kind of comm…”

“We had our, one of our co-founders, Jack Clark is our, you know, head of policy and he got connected to Rick Rubin. Cause I think he's been thinking a lot about Coding the future of coding and creativity, and they've stayed in touch and, you know, Rick got exc…”

“Like there's still set, we call it overhang, right? Like the Delta between what the models and the products can do and how it's been, they're being used on a daily basis, huge overhang. So that's, we're still like a very, very strong necessary role for product…”

“Like one is pushing the frontier of the models and then having a repeatable process. This actually goes back to our summit conversation, like a repeatable way to evaluate how well your product is serving those use cases and how well, if you drop a new model in…”

“Hey, I'm in the discord, the, you know, the, Cloud anthropic discord. I'm in the user for I'm on X and I'm reading things and like, here's what's emergent. That's step one. Models can do that today. Step two, which the models probably can do today. We just hav…”

“We really rapidly became bottlenecked on other things like our merge queue, which is the sort of sort of get in line to get your change accepted by you know, the system that then deploys it to production. We had to completely re architect it because so much mo…”

“We're limited more around like making sure the context flows through all of that. So we have the right access to those things more than the model's capability to reason and propose.”

“I came in, everybody expected me to just like go super, super hard on consumer and make that the thing. And I again would make the other mistake. Instead, I spent a bunch of time talking to like financial services companies and insurance companies and like oth…”

“He's like, I think we'll be at 90% by the end of 20, 25 or something like that. And sure enough, we're at about 72 now with the new models and. We're at 50% when you made that prediction and it's like continued to scale pretty much like as predicted.”

“I think being willing to build more at the edge of the capabilities and basically break the model and then be surprised by the next model.”

“I like to say internally, like Claude has no chill. Like if you put Claude in a Slack channel, it will chime in either way too much or too little.”

“It would be very dangerous to over optimize on like Claude's likability, you know, because you can fall into things like, you know, is Claude gonna be sycophantic? Is Claude gonna tell you what you hear? Is Claude going to like prolong conversations just for p…”

“The functional unit of work at Anthropic is no longer like take the model and then like go like work with design and product to go ship a product. It's more like we are at like, we're in the post-training conversations around how these things should work. And …”

“The team that works in the most futuristic way is the Claude code team, because they're using Claude code to build Claude code in a very self-improving kind of way. And, you know, early on in that project, they would do very line by line pull request reviews, …”

“At this point, I would be shocked if it wasn't 95% plus.”

“I don't have a great answer yet, but I think that like, it's not the web two O or even the social media days, like engagement metrics, you know, it should hopefully really be around Like, did it actually help you get your work done?”

“I actually had Mike on the podcast, their CPO, Mike Krieger, and he pointed out that Dario, every prediction he's had so far has been right over the past few years.”

“And Opus four has it it's really wild. And I think that's super important because it opens up all these use cases where you might want to use a language model as a judge.”

“Claude code, everyone inside every, that's basically what we use. If you're building something, you're using Claude code.”

“We've done, we've testified to Congress about how models can do biological uplift in terms of, you know, making new pandemics using the models, and that's an A-B test against Google search. That's like the previous state-of-the-art on uplift trials, and we fou…”

“Get used to it because this is as normal as it's going to be. It's going to be much weirder very soon.”

“And so the way I think about it, I think like my best granularity of forecast for like, could we have an X risk or extremely bad outcome from AI is somewhere between zero and 10%.”

“We published a computer using agent reference implementation in our API only, because when we built a prototype of a consumer application for this, we couldn't figure out how to meet the safety bar that we felt was needed for people to trust it and for it not …”

“If the model is in a box trying to improve itself, then it could go completely off the rails and have these secret goals, like Resource accumulation and power seeking and resistance to shutdown that you really don't want in a very powerful model. And we've act…”

“Where we've seen evidence in the wild of deceptive alignment, for example, where the model will appear to be aligned but actually has like some ulterior motive that it's trying to carry out in, in our laboratory settings.”

“I think progress has actually been accelerating where if you look at the cadence of model releases, it used to be like once a year. And now with the improvements in our post training techniques, we're seeing releases every month or three months.”

“I think AGI is kind of a loaded term. And so I tend not to use it very much anymore internally.”

“And I think honestly, through our safety research, we have a big opportunity to do things that no other company can safely do. So for example, with computer use, I think that's going to be our huge opportunity, basically like to make it possible for an agent t…”

“And if you look at something like sycophancy, I think Claude is one of the least sycophantic models because we've put so much effort into actual alignment and not just trying to, like, good heart our metrics of saying, like, user engagement is number one, and …”

“Like we're definitely not slowing down on hiring at all.”

“Instead, I like the term transformative AI because it's less about like, can it do as much as people do? Can it do literally everything and more about objectively, is it causing transformation in society and the economy? And a very concrete way of measuring th…”

“And then another piece that's come out is constitutional AI, where we have this list of natural language principles that leads the model to learn how we think a model should behave. And they've been taken from things like the UN declaration of human rights and…”

“So initially we thought that it would be sort of one or the other, but I think since then we've realized that it's actually kind of convex in the sense that like working on one helps us with the other thing.”

“We have seen internally that our legal team and our finance team are getting a ton of value out of using cloud code itself. We're going to be making better interfaces so that they can, they they'll have an easier time and require a little bit less jumping in t…”

“Anthropic has incredible talent density. One of the things I love the most about our culture here is that it's very egoless. People just want the right thing to happen and I think that's another big reason that the mega offers from other companies tend to boun…”

“I'm pretty sure it's real. If you just think about like the amount of impact that individuals can have on a company's trajectory, like in our case we are selling like hotcakes and if we get You know, a five, a one to 10 or five percent efficiency bonus on our …”

“And in terms of software engineering, our Claude code team, like 95% of the code is written by Claude.”

“If you look at the scaling laws, they're continuing to hold true. We did kind of need this transition from like normal pre-training to reinforcement learning, scaling up to continue the scaling laws.”

“One of the things that people really loved about it was the character and the personality. And that was directly a result of our alignment research.”

“I think ChatGPT at this point has Like at least a 10 X difference on MAU.”

“Most enterprises I'm talking to mention Gemini. Yeah, or Microsoft Copilot. So I don't hear much about Anthropic, to be honest.”

“I would say I've always been very, very impressed by Anthropic. Like, I think Anthropic takes a very principled view about what they do and don't care about. And how they want their models to behave in a way that feels a lot more principle to me.”

“That being said, if you look at, like, anthropics constitutional classifiers, it's much more difficult to get, like, CBRN information out of clawed models than it used to be. But humans can still do it in, let's say, like, under an hour and automated systems c…”

“This group was able to hijack Claude Code into performing a cyber attack, basically.”

“I think it was Anthropic hired a bunch of National Lampoon comedy writers to help them train models, and so they're working on it.”

“My advice generally is don't try to optimize. Don't try to cost cut at the beginning. Start by just giving engineers as many tokens as possible.”

“The thing that happens is sometimes people try to use a less expensive model like Sonnet or something like this, but because it's less intelligent, it actually takes more tokens in the end to do the same task. And so it's actually not obvious that it's cheaper…”

“We at this point have like pretty sophisticated technology to understand what's happening in the neurons to trace it. And so for example, like if there's a neuron related to deception, we can start, we're starting to get to the point where we can monitor it an…”

“So we actually, one of our PMs used that as a way to evaluate Cowork before we released it. And I think at the point where we hit where Cowork was able to do like, 48 out of the 50, they were like, okay, it's pretty good.”

“And so I think over time, this is going to become more and more normal where the models are running for a very, very long period of time and you don't have to sit there and babysit them anymore.”

“Over the last year, like, since we introduced quad code, we probably, I don't know the exact number, we probably, like, forex the engineering team or something like this, but productivity per engineer has increased 200% in terms of, like, pull requests.”

“You know, at Anthropic, we're starting to see some engineers that are spending, you know, like hundreds of thousands a month in, in tokens.”

“You know, like a long time ago, we weren't sure if the model was just predicting the next token or is doing something a little bit deeper. Now I think there's actually quite strong evidence that it is doing something a little bit deeper.”

“And so for Cloud Code, for example, we released an open source sandbox, and this is a sandbox they can run the agent in, and it just makes sure that there's certain boundaries and it can't access like everything on your system. And we made that open source, an…”

“And so over 10 days, they just completely used quad code to build it. And, you know, co-workers actually, there's this very sophisticated security system that's built in, and essentially these guardrails to make sure that the model kind of does the right thing…”

“But like I said, it's just growing much faster than Quadco did in the early days. So I think it's starting to break through a bit.”

“So we actually think if you look at private repositories, it's quite a bit higher than that.”

“My take is, I think for people that are using that are using quad code, that are using agents to code today, you still have to understand the layer under. But yeah, in a year or two, it's not gonna matter.”

“So on the quad code team, everyone codes, you know, our product manager codes, our engineering manager codes, our designer codes, our finance guy codes, our data scientist codes, like everyone on the team codes.”

“Actually, something that people don't really remember is Cloud Code was not initially a hit. When we released it got a bunch of users. There was a lot of early adopters that got it immediately, but it actually took many months for everyone to really understand…”

“And Opus four was our first kind of ASL three class model that we released back in May. And we just saw this inflection because everyone started to use quad code for the first time. And that, that was kind of when our growth really went exponential. And like I…”

“All it is, is we inject one sentence into the model's prompt to say, please don't write any code yet. That's it. Like there's actually like nothing fancy going on. It's just the simplest thing.”

“Honestly, for our team though, we're just focused on solving the problems that users have. So for us, you know, we don't spend a lot of time looking at competing products. We don't really try the other products.”

“And my prediction back in May of 2025 was, by the end of the year, you might not need an IDE to code anymore, and we're gonna start to see engineers not doing this.”

“The modern framing that I've been seeing in the last six months is a little bit different, and it's look at what the model is trying to do and make that a little bit easier.”

“And in general, what we see is maybe scaffolding can improve performance, maybe 10, 20%, something like this. But often these gains just get wiped out with the next model. So it's almost better to just wait for the next one.”

“We were building the models in this way that kind of fit our mental model of the way that we built SafeHEI, where the model starts by being really good at coding, then it gets really good at tool use, then it gets really good at computer use. Roughly, this is …”

“Here at Anthropic, Claude reviews a hundred percent of pull requests.”

“Everything that's happening right now around, you know, just like quad code becoming this huge, you know, multi-billion dollar business.”

“And for me, this is actually a pretty important product lesson, right? It's like you want to under-resource things a little bit at the start.”

“All of my project management for the team co-work does all of it. It's like syncing stuff between spreadsheets and messaging people on Slack and email and all this kind of stuff.”

“There's this there's interesting thing that happens also when you when you underfund everything a little bit because then people are kind of forced to clodify.”

“So my read here is like, I don't think Chad is ever going away because this opened up this like new way of like infinite ways to work with the model and to sort of like talk to the computer that we just didn't have before.”

“I think this person is really interesting right now because I think most companies are just hiring like senior talent, like folks that have done things before, are super experienced, but given how much the roles are changing and what we're expected to do is ch…”

“I don't think Claude is there yet. I don't think Claude is there yet in terms of a designer you would hire. I think it is not yet the strong generalist or the deep specialist. Or the crack new grad. I think it's pretty good at a first pass and at presenting a …”

“I think as a designer a few years ago, I would say, like, maybe 60 to 70% of it was, like, mocking and prototyping stuff up, and then spending, you know, the last 20 years, some of the last 20 or so, like, doing the sort of, like, jamming with engineers, consu…”

“Like overall, it was like 10 days to get it from what we had internally. To something that we were ready to ship externally. So we'd been building it for a while, but we weren't really sure about like the actual form it was going to take.”

“Because I think the way that you really lose trust around quality and releasing something early is if you release it early and then nothing ever happens. That, that is what, something that degrades a brand.”

“And I think that what will probably happen here is that a lot of those UIs will be generated more and more often by the models, as opposed to something that we're like hand coding each instance.”

“Part of the role of the designer, at least at least at a frontier lab at Anthropic is kind of spotting the ideas that are illegible and trying to understand what's there and how to, Take that and, like, transform it, whether it's through storytelling or whethe…”

“Right now Claude is like writing a lot of the code and it's like not always, it's not always like picking up stuff in the design system and whatnot.”

“I think Firefox just a few days ago, maybe last week said that they'd, they'd done a release, which was assisted by Anthropic. Anthropic had Discovered a hundred like potential vulnerabilities in Firefox and responsibly reported them to Mozilla who then fixed …”

“Both Anthropic and OpenAI spent the whole of 2025 focusing all of their training efforts on coding.”

“Both OpenAI and Anthropic have specialist security models That they will not release to the general public because they can be used to break into websites. So they have like invite only, like registered security researchers can apply for access and they've bee…”

“The reason OpenClaw took off is Anthropic and OpenAI could have built this and they didn't because they didn't know how to build it securely. If you're an independent third party, you don't have that restriction. You can just Build something and put it out the…”

“He said, I'm the only PM that he's hired from cold email.”

“It's delivering results, right? Like, and it's like, you can push it, press play with it. And it's like, it ultimately prints money where I'd say that the win rate is like, I would expect a senior PM to do better. Like I would say like, this is like a junior P…”

“The nineteen billion number you quoted is from the end of Feb, so that is also out of date.”

“I basically you know, one of my manager, Ami Vora was, I think, a podcast guest of yours, right? So I say, hey, based on what you know of Ami, both publicly, she's written extensively about product, and then internally, and then our discussions, what is everyt…”

“And so for us, that top line objective of this just like, this needs to go well for humanity. That is something that we are happy to take a significant commercial hit for. And we've done that time and time again, right? So like we have, you know, like back the…”

“One of the biggest Problems in the industry is capability overhang, where the models are just getting better so quickly. And it, the real challenge is on the product side of how do we start to diffuse those benefits to people”

“I think we actually believe in this stuff more strongly than we say externally, so, like, it is just such a Key part of how we think internally that sometimes like we just like re, you know, reword your statements because it's like, people might just think we'…”

“It's clear that while PMs and designers are getting more leverage from AI, engineering is getting the most. Leverage right now. I look at tools like Cloud Code and like they, the amount of leverage engineers are getting from them is higher than I think the amo…”

“With Cowork, you have this, the Slack MCP, and you can tell Cowork you can tell Cowork to say, basically, look across Slack. You know, the projects that I'm working on. These are the things that are top of mind. Go and find me areas of potential misalignment r…”

“Everyone has these notebook channels where you kind of have like your own like Twitter feed in a way where you're just, Talking about your thoughts about things, and so you can go and, like, join the Slack channel, the notebook channels of people on research a…”

“Something that Ben Mann is one of our founders on the vodcast had written. It's dated in 2021. I think a few months after they started the company and it was like, here's why we should just focus on AI coding. And this is like, you know, this is like five year…”

“We're roughly maybe 40 people now”

“Probably the things that are maybe different is that we, I think that we index a lot more towards larger swings as opposed to smaller optimizations. Like if I think about a traditional growth team, I would have probably done maybe 60, 70% of my time on small t…”

“That, that 10 X year on year revenue growth trend has been there since the beginning. I think, twenty-twenty-three was zero to a hundred mil. Twenty-twenty-four was hundred to one. Last year was one to roughly 10.”

“If you're in the game and you're a leading player and what you're doing is working, you can influence people in, who are also in the game to take the right actions. And so that is like the core of it. Like if we just pack up and go home, you have no influence …”

“The product value that we will deliver in two years time is probably like a thousand X, a hundred to a thousand X what it is today.”

“I think then the, probably the big differences is I would say that like roughly 70% of what I spend my time on is, is what we internally refer to as success disasters. And that is where like things have gone so well that other things are breaking now.”

“The frame that we have is that if a project is less than, is two weeks of engineering time or less, then the engineer is on the hook to Effectively be the PM for that. And so that means things like talking to security, talking to legal, talking to cross-functi…”

“In this world of engineers are mini PMs that the engineer who is like highly product minded is a unicorn is an absolute unicorn. I think the same thing is true for PMs where it's like, okay, now, you know, this, in this ratio of the designers really stressed a…”

“One of the biggest things you come into Anthropic is you need to understand that probably, yeah, 50, 60, 70% of how you operate in the past, just throw it out the door. It's not going to be relevant.”

“Anthropic had a version of Claude. We had a chatbot before ChatGPT was launched, and we had ultimately chosen not to launch it for safety reasons. I think the team didn't want to kick off Effectively like an AI global arms race”

“Probably 70, 70%, maybe 6070, 80% of what we ship does not have a PRD.”

“We are starting to look at how do we automate growth, which I think is like a really interesting area. So our growth platform team we have, we're very lucky. We have like Alexei Komisaroka who teaches growth engineering at Reforge, and he's just like the guy o…”

“At that point, you're probably not going to run everything locally on your machine anymore. There's just like not enough RAM to do it. And so we're thinking about. How do we make it easier for you to manage all these? These will probably run remotely.”

“There's products that we ship that aren't as polished as I wish they were, but you know, our top goal is to help empower professional developers. And if a product isn't successful, as long as it's not blocking the core use case, it's okay because we'll hear th…”

“So for example, we have a really tight process between engineering, marketing, and docs. So when engineers have a feature that they feel is ready and that we've dog fooded internally, they post it in our evergreen launch room. And then Sarah, who leads our doc…”

“It wasn't designed for third party products, which have different usage patterns than our first party ones.”

“I think now with AI and with how much that has accelerated engineering and with how quickly the model capabilities are improving, the timelines for a lot of our product features have gone down from six months To one month and sometimes to one week or even one …”

“We've been moving pretty fast for Several quarters now, so I think it, it's not fully Mythos. Mythos is an incredibly powerful model. We do use the models internally, and I think this has increased our rate of shipping a little bit, but I don't think it explai…”

“And this, if an automation doesn't work a hundred percent of the time, it's not really an automation. And that last five to 10% does take more time. Also building the automation is often a lot slower than you doing it yourself.”

“Even coding is easier because you can verify the success. Whereas crafting the character requires a very strong sense of conviction and what, who Claude should be.”

“So in the past, we didn't actually want to do something like power up because we felt like the product should be intuitive enough that you can, that you don't actually need to go through any tutorial. And over time, we've just realized that there's just so man…”

“Yes. Yes. It's a process failure.”

“We realized that this was the result of human error. There is a human working with Claude to write a PR. This was just an update to how we release our packages. And it actually went through two layers of human review. And so this was a result of human error an…”

“A lot of the changes that we make with a new model is removing features that are no longer needed. So a lot of times we add features to the product as a crutch for the model because it's not naturally doing itself.”

“He's our tech lead. He's very much the product visionary. And he is great at setting like, this is what the product needs to be. In like three months, six months from now. This is like what the AGI-pilled version of the product is. And a lot of my role is figu…”

“I think part of what makes a great coworker is this positivity, this like bias towards action, this ability to give you like earnest feedback, not just agreeing with every single thing that you say. And so we try to imbue this into COD because we think it make…”

“And actually almost all the PMs on our team have either been engineers or ship code here on code code.”

“I think one of the most important things for Anthropic is to grow the number of users that we're able to reach. One of the ways that we're able to do this is with the cloud subscriptions with our first party products. And so we just very much want to double do…”

“Like, I think one of the things that quad code has really unlocked for our entire company is it really lowers the barrier to making any custom app that you want. And so we we've seen this like surge in personalized work software that people are building for li…”

“The CLI is our initial product surface, and it's also the one where our features often land first. And so it's the most powerful of all the tools.”

“I think the reason why an engineering background is particularly useful, at least for the next few months, is if you have an engineering background, you have a better sense for how hard something should be. And that's often a factor in what you choose to build…”

“This code review is so good that our engineering team relies on this code review to pass before we merge PRs.”

“And so we do see the token cost per engineer or like per any knowledge worker increase every time that there's a model jump or like a substantial product improvement. I think it's still much lower than what the average engineer salary is, but we see the percen…”

“So for Cloud Code, what we do is we actually ship almost all of our features in Research Preview. We clearly brand this when we ship something so that users know that this is an early product, this is just an idea, this is just something that we're trying to g…”

“With Opus four and later models, we realized that we didn't need to force it to use this to-do list. It would like naturally use it itself. For the earlier models, we had to keep reminding it. Hey, did you finish everything on the to-do list? You can't finish …”

“We have this list of team principles, and this includes who our key users are, why those are our key users. And the reason that we articulate all of this is so that everybody on the team feels like they understand how our business works. They understand what's…”

“For features that are like particularly ambiguous, it does help to write out just a one pager on what the goals are what the delightful use cases are, what the failure modes currently are that we need to fix. And there are occasionally some projects, especiall…”

“So the way that I split the products in my mind is if I'm building something where the output is code, I'll use code code or desktop or code code on mobile. And if the output is anything that's not code, I'll use co-work for it.”

“I think we also tend to hire people who have been in the industry for a while and have experienced lots of ups and downs and have a good sense for what gives them energy and how to maintain their energy over time.”

“I still think it comes back to product taste. Like as code becomes much cheaper to write, the thing that becomes more valuable is deciding what to write.”

“We're sacrificing product consistency. Historically, when code was expensive to write, you would carefully plan out everything in your product suite, how every product relates to each other, what the use case for every single one is, how they integrate, and yo…”

“Dario is that she wasn't, oh, he wasn't just lucky once at OpenAI. He did the same thing twice and it was successful twice.”

“They didn't actually implement what's called now the long-term benefit trust until their Series C. But they had the right and the intention to do it in all their legal documents from inception.”

“They turned down a two hundred million dollar contract and bore the wrath of the world's largest army and government.”

“Dario does not have dual class shares the way that Mark Zuckerberg or Larry and Sergey had.”

“Anthropic being the fastest growing company of all time, just like absurd, breaking all records is a public benefits corporation should make you feel okay about doing this.”

“Anthropic has directors on its for-profit board who are appointed by and are accountable to an outside group of trustees who are AI safety experts who do not have equity in Anthropic. So they do not have a financial incentive in its growth.”

“Most of the best companies today are using this structure. Like all the major AI labs are incorporated as BBC's Anthropic most famously of all.”

“OpenAI for a while was, I, in my opinion, like, very, very behind on this, and then in my opinion has surpassed them recently.”

“Like, I think that Cursor's cloud implementation is better than either OpenAI or Anthropix and is more advanced.”

“And so then you can kind of get into calling those races where, again, it's like being in 1997 and saying, well, is it going to be Excite or Yahoo? And the answer was no, generally.”

“I don't think the fact that you run AI lab suddenly gives you, or rather, and if you're going to use argument from authority, then it should be relevant to the field. So, like, I'm interested in Dario's opinions on where models are going to go in the next six …”

“But at the time, Dario was saying, you know, 90 to a hundred percent of all our code's written by, you know, our, you know Claude. And, you know, we just monitor it and watch it.”

“The other thing that we found interesting on the Cloud Code team is after a while we felt it could start being a lonely experience cause we all started just working with our agents so much. So recently we started maybe like a pairwise programming lunch.”

“When I first joined Cloud Code, we had really great kind of like product generalists. And then I realized, oh, we were missing folks with a systems background. And so that was definitely an area that we needed more folks with kind of like systems and distribut…”

“Well, definitely. I think most commits are Claude assisted.”

“So I think we'll, we'll move more towards that kind of like asynchronous style of working.”

“Like now I just have a routine that automates all this for me. And then there's also, it's almost like before I would, you know, be able to kind of like, you know, generate some prompts, but now with routines, it's almost like I'm, you know, having an agent he…”

“The other shift that I've recently done is I actually have a quad code remote session that I enlist in all of our repos. And so this way I have full visibility into the work that everybody's doing. And this instance, he also has access to all our Slack channel…”

“I think we didn't even have cloud code reviews last year. And so speaking of bottlenecks, that, that was a really, really big bottleneck of, you know, the human reviewers.”

“Like now, not only engineers with, we also have designers, PMs, everybody on the Clark code team checks in code.”

“Along these lines, there's just this tweet that you guys put out yesterday where you showed here's the tweet, anthropic engineers on average of eight times as much code per quarter as they did compared to 21 to 20, 25.”

“That was an engineer on the team of, hey, we should maybe track square words. I'm like, oh, that's a great idea. I remember it. I had just, you know, joined when we were having that really fun conversation. Yeah, so it's again, like, it's very interesting to k…”

“Bad is like a very bad, irrecoverable error. And sad is something that's kind of like a pain point recoverable, but it's interesting when you stack up sads, it could, you know, generally go to bad, but even having a, like starting with a high level framework l…”

“On the manager front, I think giving time for managers that join to be able to go deep and to do that before supporting people. And then they actually like end up building really great rapport with the team. Like, because if not, I think sometimes as managers,…”

“There was another saying, I really like don't forsake motion for progress. Because if you're measuring like, you know, like tool user usage, then you're measuring the action, but is it really making whatever the end outcome of yours like important?”

“Outside of metrics, especially if on teams, as you're having more people adopt AI tooling, I would say definitely go on a listening tour for the leaders listening to a podcast, especially, you know I love to focus on the senior engineers as well. Like hear fro…”

“I think six months was too long. So now for sure some projects will take more than a month, but we try to do like a month planning, like really lightweight. Actually, there's not even dots. It's really just Also hunting on a little spreadsheet of what we think…”

“So when we do have a model, we actually, for every model, we do have a PRD. Less necessarily for our researchers, but more for our growing product surfaces, for our engineering teams, for our stakeholders like legal and safety and others as just a source of tr…”

“One thing we say a lot on the team is you need frontier products in order to have frontier models and for people to feel the magic. Of frontier models.”

“One thing I asked the team frequently or how I think about when we're building a product is let's say Claude eight comes around. What do you, what changes in what users do? And then what should, what does that mean for how you're building today? Is it going to…”

“I don't think it's necessarily just about raising the IQ of like experiences we build, but also I use it a lot and actually like prepping for how to have better conversations in the moment during like crucial conversations. So I love that book. And so I actual…”

“So I've used Claude to help with things like, are we making the right pricing decision on the next version of Claude? It's a little bit meta, but using a research version of Opus, asking it to figure out how is your price and being able to come out with better…”

“Here, you have to sweat the tokens as much as you sweat the pixels. And so one activity we have on the team is reading the transcripts. And understanding, ah, what was the trajectories that failed?”

“I think there's like product overhang and user overhang, like to maybe put it in our PM language, even on today's models. And I think there's like a lot that we could be exploring on like our current opuses and definitely with like Fable, for example”

“I think even at the founding of the company, researchers were talking about how do we get clawed to you? Use a computer. How do we get AI to, like, navigate a screen, right?”

“That to me was like one of those, like maybe hidden inflection points of we were starting to find our identity, that we could build products, build experiences that were different for what our competitors had seen, what was already out there.”

“And what I saw was something like 80% of what people meant in the early days for this failure was Claude would not write the right JSON.”

“As you add in more compute and data, what's called loss, aka the loss from next token prediction goes down. And so it's a very smooth linear curve of, like, the models get more intelligent as you scale them up. What's actually also interesting in that paper is…”

“I think, you know, software engineering has been really transformed by AI. I think there's areas like biology, life sciences. These are all things that we're just kind of at like the foot of the exponential on. Like, maybe software engineering were on the expo…”

“I think Opus Four Five wouldn't have had that moment without a product like Cloud Code. And Cloud Code, I think, wouldn't have had that type of adoption accelerated without Opus Four Five.”

“People are starting to use code these models, not just for code, not just like code autocomplete, but actually writing long form code. And it's that an opportunity for us to train, you know, Opus three to be better at. And it ended up being a relatively smalle…”

“If you bring that to a researcher and you say, please fix Claude from being hallucinated, it's not very actionable. And so part of the time of the team is understanding, okay, what's the trajectory of why that user gave that feedback?”

“When we have versions of Claude, we actually run that eval and just check. I think at this point it's always a hundred percent or like 99.9. And so it's no longer a pain point.”

“Our like leadership, our chief scientists, our heads of like fine tuning and like RL folks are actually really close to the training runs and actually look at things like how the training run is going, evals, looking at the underlying data.”

“I think judgment is one, is an area where it's accumulation of so much nuance and so much experience, and these systems haven't experienced as much as humans have. And so I think that hard earned, like, judgment Is a area for product leaders and just generally…”

“So I joined in 20, 23. Like you said, we had five product engineers. There was one engineer for the entirety of our API business, if you believe.”

“We actually have a saying on the team of evals are the new PRDs. Cause in order to deliver that user value it's not that exact artifact that people used to write in the last like one to two decades. It's a new way of working.”

“Definitely when we were training and testing Opus three, I think that was the moment when the company, I think we were less than 200 people still at that point.”

“What you don't want is like a AI that just agrees with you, right? What you want is this technology to actually augment and grow and like get to a better outcome. And so sometimes it's Having Claude push back makes me better, and so that's great. Like a cowork…”

“We actually on my team have not changed our hiring loop for three years now. So what we actually look for and the traits and how we evaluate generalists like PNs, generalists like research product managers have actually been the same.”

“If you thumbs up or thumbs down on any of our product surfaces, if you contact your salesperson with feedback about the model, it will make its way to me. We actually, with every, like, research model, I actually get pretty close into understanding favorabilit…”

“Second one I've already mentioned, which is like stop bothering engineers with how does the code base work and actually just go and talk to Claude and understand it, which is really helpful. You know, I used to say early on in my career that the goal was alway…”

“The CTO of Workday is just a member of technical staff at Anthropic, CTO, CTO of Instagram, Box, the CTO. Super.com CTO. They're just like engineers now at Anthropic.”

“The sort of concept of the model that's too dangerous to release it kind of conflates marketing inference capacity, and then also economic considerations, like, do you want to externalize your competitive advantage or use it to make yourself better?”

“Despite the, like the sort of scary concept of this, like, you know, supremely intelligent model, that's totally proprietary to one company. The, so far the kind of industry trends haven't played that way at all.”