Transcription
Please welcome CEO HubSpot Yamaly Rangan and CEO and co-founder Anthropic Dario Amade.
[Music] Hello Inbound. Welcome to day two. Yeah. How was day one?
Thank you. Thank you. And Dario, welcome. Thank you for taking the time and thank you for having me.
Absolutely. Well, first of all, congratulations on your hunting round yesterday. Thank you. Huge. That is huge. It it Yeah, that's great. So, look, this is a growth conference and you are the fastest growing software company. You went from zero to 100 million in 2023. You 10xed it to a billion in 2024 and we are not even half we're halfway down the year and you're at 5 billion. And to put everything in perspective, it took Microsoft 40 years to get to 100 billion and Apple 35 years. You might end up doing it in a handful of years. How are you scaling and how are you staying mission aligned to your principles?
Yeah. So, you know, it's it's it's it's funny how we think about the exponential, right? Because it's it's almost like there's always two voices in my head where I look at the exponential, it's growing 10x every year. And there's a part of me that says, you know, this is crazy. You've got to use your common sense, right? What this once this gets into the billions, like the money's got to come from somewhere. People don't spend that f like it can't happen. And then at least so far, it keeps happening. At some point, it's going to stop. it's going to to slow down at least at least somewhat. But but you know, I've I've I've found the tension between those two things always to be very interesting because you ask yourself, well, how could it happen? How could it keep growing that fast? Um uh and you know, we face the same tension with investors where on one hand there's a temptation to say, oh yeah, it's going to keep growing really fast, but it sounds crazy. And and and so I I've just started to investors and to everyone saying, "Look, there's two ways to look at this. We don't we don't know you know to keep it growing we've had to scale much faster than a normal company would in headcount in processes in go to market and so we always try to stay ahead of the curve we always try and say how can we how can we do this faster let's imagine where we're going to be 6 months from now in terms of sta you know staying aligned with the mission you know I you know I think succeeding and staying aligned with the mission doing both of is maybe the most challenging thing balancing those two things and you know I think I think the way we try to do it on one hand we have to be pragmatic you can't be rigid about everything but we look at the things what are the core things we really care about in the mission safety security trust both in the you know the the simple sense of like you know being trusted to work with enterprises and being trusted to bring out this technology for society and so keeping those as north stars even as we go through this hyperrowth. I think that that has been what's allowed us to continue to scale and actually I think paradoxically trust in the mission among our employees among our customers has been a unifying force that's allowed us to survive the scaling in in a way that I think you know some other companies that have grown that fast can become fragmented.
Culture is this sort of unifying force. Yeah, we we think about culture as a second product and I think that's very similar to how you think about it. Now you mentioned the exponential curve. Most of us here think linear terms. We can't think exponentially but obviously the model capabilities are going exponential but also growth is compounding. You seem to have a clearer view of what happens in the exponential curve. So what happens with AI in 2 years or 3 years? Where does it go?
Yes. So again, you know, I would go to this I would go back to this dichotomy where we're on this exponential and if the exponential continues again, it only has to continue for another one, two or three years. Absolutely crazy things happen in terms of the model in terms of the model capabilities. And so again, there's, you know, there's like two minds. There's like well common sense this just can't it has to stop. It can't go on versus what if it actually does. Yes. And and I think I am someone whose instinct has always been on the side of what if it does, I think it will. I'm not certain of that, but it's it's where I'd put my bet. And and the way I think about it from the perspective of the models is, you know, today I would say the models we have, they're maybe as smart as a smart undergrad in some areas. They're able to augment what what professionals do. You know, we've seen that in medicine. stories of people, you know, uploading their medical information to AI. They're not yet at the level where they're creative enough to go beyond the frontier of human knowledge. And I think if the exponential continues for only maybe one to three more years, then we'll start to go beyond the frontier of human knowledge and then things really go crazy. Um, like what happens then? Yeah, I mean, you know, I think we're going to get to this world where AI is able to make original scientific discoveries. It's able to work with humans to make those discoveries. Very small number of humans will be able to work with a massive AI agents to build new companies, new types of economic activity. I in particular have almost, you know, I've been the most excited because I was previously a biologist. I've been the most excited about the medical breakthroughs that that AI can make. Its ability to cure diseases that that has always struck me as the kind of clearest positive application, right? If we're looking for how will AI make everyone's lives better, we look at cancer, we look at Alzheimer's, we look at mental health, you know, the these problems that affect billions of people, you know, can can we use AI to to reach in and really solve these problems much more quickly than we could before.
I can't imagine I I I that's why we talked to you because I think you are you have much clearer picture of what happens in medicine, what happens in business and you know just a couple of years. Uh I want to talk about you know the business application of AI and it's super interesting. You have a thriving app business but you've said that majority of your revenue actually comes from APIs which means that you're betting on business use cases. Is that actually true? And then talk to us about how you came up with cloud code which is the first really big business use case. What are the next couple bigger ones that are coming up?
Yeah, we are absolutely betting on business use cases. Um, so you know, if I if I think about it, there's a dynamic and I was, you know, thinking about it kind of as these early technologies were were being developed early in the history of anthropic in 2021 and 2022. Often when there's a new technology, the first way of thinking about the new technology is kind of an analogy to previous technologies or too much like previous technologies. And so I think it's it's not an accident that the chatbot emerged as the first application, the consumer chatbot because it's like social media or it's like Google and people are using it as a substitute for those things. But I think as often happens with new technology when it's when it's used as an analogy or drop in replacement to previous technologies that's only a tiny fraction of its power, right? And so I think the ultimate power of AI which is taking a little longer to emerge because it's a new thing um because it's a totally new way of of using the technology is is kind of enabling the businesses and the economic activities of the world. Yeah. Um and so I think in contrast with some of the other major companies in this space which are consumer companies or are clearly signaling presence in a consumer direction, Enthropic wants to be the company that serves enterprises and serves business use cases.
Love that. Um and and so you know we we are scaling up in this area. Um, you know, I think I think as you mentioned, um, you know, code is maybe the first area to to show to, you know, to kind of show the the, uh, h magnitude of business use cases. I'm on one hand incredibly excited about code. On the other hand, I actually think the lesson from it isn't that code is a a uniquely compelling application of AI. I mean it's really compelling but I think what's going on is that coders people who write code are are people who adopt technology quickly and people who are close to the LLM revolution. So the diffusion of the technology has been faster. But but what we're but but what that shows to me is that the incredible value we've seen in code we could see in all of these other areas. It's just there's more friction and we have to get over this friction. And so a lot of what we're thinking about is yes how to how to expand on the code side um but also how to reduce the friction in areas where there's huge potential right areas like financial services areas like sales and marketing areas like um, you know pharmaceutical and and pharmaceutical and biomed areas like uh legal productivity um, it it really manufacturing it really runs across the board and I think were held back not by the power of the technology to change these things but but by needing you know just needing to solve all the practical problems that comes up when there's a big enterprise that does something other than AI and somehow has to mesh or integrate itself with the new AI technologies.
Is there like a moment where you were like uh you saw something in cloud code and the way developers were using it that it became the big thing and do you see that type of pattern emerging in the next couple of big use cases?
Yes. So the moment was actually before the launch of cloud code. So um early in 2025 um I had kind of given a directive to the company to say hey we should try to use our own AI to make ourselves more productive right we should try to accelerate ourselves with the technology and so you know I I didn't want to tell people exactly what they what they should build because that doesn't work very well um you know I said I said experiment with things and so there were a number of experiments and one of them was you know one of them was this internal tool I it had some other name that I've I've I've forgotten by now, but it was basically this command line tool that you know you could you could use to write to for anthropic engineers and researchers to write code and I noticed that you know this tool was presented and then within a week like a hundred employees within Anthropic used it and then like within two weeks like you know 500 like half of anthropic was was using it and I'm like people really seem to like this it's being adapted really fast that we were then about to release cla 7 um and you know I looked at it and I was like hey can we just release this publicly and you know there were various reasons it was like oh you know we'll have to and I'm like this this seems really compelling you know uh uh uh you know I I think in general you shouldn't have a prescriptive view on you know you shouldn't have a road map and say follow the road map if something is getting product market fit including internal product market fit like that's the best signal and you know nothing else nothing else matters right like you know that that's the thing that's that's the thing that's going to succeed. Now, if you're worried about it from a safety and moral perspect, that's another thing. But if it if it appears safe and it has product market fit, even internal, you should make it go as fast as possible. Um, and so that's that's what we did with quad code and the external reaction was very much like the internal reaction that we saw to I mean I I will say that our uh R&D team they're using it every single day and just in one year our engineering productivity has gone up 42%. 42% more code commits this year compared to last year. So thank you for what you do. It's really incredible.
Um, you know speaking of which I want to talk about how every big tech cycle starts with a platform that unlocks like huge waves of innovation. We've seen this in mobile app store created hundreds of mobile apps and I'd say AWS has enabled an entire generation of SAS applications to be built on top and we think at HubSpot that there is a new AI operating system that is getting formed because of all the work that you're doing and the foundation model companies are doing is do you think of anthropic as a platform company?
I think I think we think about it exactly that way. How do you is it like mobile? Is it like AWS? Like how do you think about it? You know, I I think the closest analogy might actually be cloud um in part because it literally runs on cloud. But you know, if you look at a business like AWS and some of the the other clouds, some of the other clouds are the same. You know, they're they're kind of they offer different things. They're a bare metal platform that anyone can build on and that's very fundamental and general to the economy, right? Everyone needs cloud computing and I think we're heading towards a world where everyone needs AI on top of cloud computing and so you can just like you can buy the instances in this total bare metal way like just I'd like some CPUs I'd like some I'd like to rent some compute you know I think of our kind of API that way. Now if you look at the business of someone like AWS there's various kind of you know there's various kind of tooling and kits on top of it. You can you can think of that as things that are in the middle layer like clawed code maybe maybe you know fits fits in that uh fits in that somewhat the engine behind claw code which we're increasingly starting to um, you know expose more and more more and more uh uh pieces of to people and then in some cases if you look at the platforms like AWS it makes sense for them to make build apps themselves on top of the on top of the basic platform and sell them and you know we are very thoughtful about when we choose to do that typically it's in areas where we feel like it's an area we know really well so you know I think the reason we made claude code is we ourselves are writing code right so we have the skill there we can assess product market fit I think across the ecosystem it is our intention much more often to be a platform than we are to kind of directly build things ourselves um, you know, we're not going to run clinical trials for drugs, right? You know, we're not going to start up a manufacturing company, you know, we're we're not going to do uh, you know, we're not going to do marketing like um Oh, good. You're not going to do marketing. Thank you. Thank you. You heard it here. uh uh uh uh but you know we we we we we think about you know where is it that we have some that we have some unique advantage that we can do something different and you know I think in the rest of the areas you know we really want to kind of enable our customers and there will be customers of all kinds you know we see startups we see enterprise customers right um uh uh and uh, you know, we're you know, we're kind of agnostic right we think in general there's going going to be you know a question of in every area as it's being revolutionized by by AI are the best players to you know, to um uh to capitalize on on on that are the best players you know, the incumbents with all their kind of built-in advantages or, you know, startups who can start from new in the new paradigm I think it will differ in in in different fields and in different parts of the field and I think our view is we, you know, we just want to help everyone succeed right we want to help everyone one be more effective.
Yeah, that that makes a lot of sense. You you mentioned like startups and enterprises and enabling them. Uh if you look at AI, the consumer adoption of AI has been just incredible. Never seen anything like this. 800 million people using AI every day. What does it take for small medium businesses to have that type of adoption of AI? Do you see things that are accelerating it? And do you see things that are blocking it?
Yes. So um I think there is very much the potential for this and in fact it's already happening in some areas although it is not yet happening in other areas. So I think within code which we've been talking about for a while the adoption curve is actually if anything even faster than it has been in consumer. Um if we look at the growth of clawed code it's gone from basically you know zero or nominal before we gate it in you know 3 to 5 months it's gone to from you know it's gone to you know uh a half a billion and you know is is kind of well on its way well on its way to a billion. We have the largest coding product, but the mar the market is bigger than just us and it continues to go very fast and it's it's you know growing the fastest in in kind of small small startups but also developers for companies that that don't do this. Um I think for other areas of the of the enterprise that exponential either has not yet happened or is it just starting to happen. we're we're early in it and and the way I would diagnose it is is what I said a little bit before which is that the underlying technology is capable of that kind of 10x per year growth y that's even faster than consumer um the I think I think the key barriers are one in areas other than AI y less familiarity with the technology right like you know we still kind of you know do demos you know in say the insurance space or the financial services space, people of course have heard of AI, they've heard of the hype, but if you actually give them a demo of like say in the insurance space, here's how AI can do claims processing. Here's how AI can help with underwriting. They're like, oh my god, I heard the word AI so many times, but I didn't know that it could do this specific thing in in my industry. And so that knowledge has still not propagated. There's I think another you know another challenge which we're working you know we're working very hard it's I think you know it's a very classical enterprise challenge is when I go to you know big companies but this is also true in some ways of small and medium companies um and say look you know you know we'd love to work with you on this almost always the CEO people one or two levels below the CEO are very excited they fully understand the technology but the company let's say it's a pharmaceutical or a bank. It's made up of perfectly smart people, but people who the thing that they're an expert at is not AI. So, they have to learn new things. So, that kind of enablement and getting them to understand what AI can do across the company. Um, that takes time. Yeah. And so, we are we are working on the process of how to make that happen. And it's really starting to happen. I think in the last six months it's really started to pick up often driven by the coding use cases at those large enterprises right that's the beginning they say oh our developers are really clamoring for because the developers are very familiar with the technology um but then once they've adopted people elsewhere in the company start to ask well you know what's this what's this code model we're doing what's this company we're working for um and and then it starts to drive it so I'm very optimistic that the use across the across both large enterprises and small and medium companies will will you know it's it's kind of starting on that on that exponential now.
Yeah. And and part of it is also safety and one of the things that I think Anthropic has done an incredible job is keeping safety front and center. You have pushed for interpretability of models. You have pushed for controls in terms of regulations. What do you think is like the biggest risk and uh how how do you make sure that you are maintaining that safety and trust especially for small and medium businesses to continue to adopt the technology?
Yeah, absolutely. So, you know, I you know, I think for there's there's a number of these different you know kind of kind of you know safety issues that we you know we think about and try and be thoughtful about them. And I think for each version, you know, for for each issue, you know, there's there's there there's kind of a duality where there's the kind of, you know, there's kind the kind of grand and principle, you know, worries about as the how these models get powerful, what we worry about, and then there's a practical version of as we deploy them today, what are the practical problems people have? And you know I I I I kind of think of them as continuously connected to each other where where one is, you know, a kind of intellectual articulation of where things are going and the other is well what part of it is is is kind of happening you know happening right now where does the actual rubber hit the road. So, you know, one issue I worry about is kind of the alignment or controllability, yeah, of the of the AI models. And, you know, I think I think in the in the in the long run, there's there's issues of, you know, we have these very powerful models that are smarter than any human. How do we keep them kind of under human control? But I think the ver the short-term version of that that we're facing today is as we as we look at these models, how do we make sure that they do things that people, you know, that people want them to do? If I'm an enterprise and I'm using these models, you know, if I give them access to my enterprise data, you know, how do I make sure they don't, you know, vent that that that that that data, you know, publicly or just give access to it? How do I make sure that people can't use, for example, prompt injections to, you know, compromise security of the model? How do I handle privacy? And I think there are brand issues as well, right? you know if just to take a not to pick on anyone in particular but if I look at you know the Grock model or something like that right that's an example of not handling this well right if I want to use a model in the enterprise it's probably undesirable for that model to talk a lot about Hitler it's probably undesirable um uh uh and so you know I think I think we generally strive for that we generally strive for the opposite of that um I think there's you know lots of uh concerns about abuse of the model. Um because our model is so good at code. Um a concern we've had for a while is that bas basically people will use the model to hack into things for for for cyber crime. And indeed we we recently shut down we've been monitoring and we were able to shut down very quickly. Um someone state actor was trying to systematically use Claude to you know conduct ransomware attacks. Um, folks in North Korea were trying to use it for kind of employment scams to evade sanctions. Um, so we track this stuff down. We shut it down really quickly. We're always really transparent about it. I I worry that you hear about this from us because we're tracking it so much and because we shut it down so quickly. But the fact, you know, that but but you know, I I don't think that means it's not happening elsewhere. That's true. I I think it means we're the only one who's who's looking at it. And so we're always kind of very careful to make sure that particularly when our enterprises offer their own products, we can help them to prevent these things from happening and to to kind of shut them down quickly because no one wants their product to be abused in this way. By the way, like you know, I talk to customers a lot the there is a big concern in terms of how they can maintain the privacy of their data and which models to work with in order to do that. So I think like you continuing to talk about this and push it is really important.
Yeah. So one issue I think we've been thinking about a lot is as we work towards what we call the virtual collaborator or virtual co-orker. Um, you know, this is obviously very powerful. There's a vision of it where you know within an enterprise you can give it access to you know all the internal documents of the enterprise. You can um, you know, access all the Google docs. You can look in you know, you can look in Slack. You can use the marketing tools. who can, you know, the model can be an agent that kind of connects all these things together and is a helpful assistant and collaborator to you and has all the affordances that an employee would have. But if it has all the affordances an employee would have, of course, an important thing that you have to be very thoughtful about is, you know, just as one of your employees could run off with your IP or could, you know, share could could share a secret accidentally or intentionally in in an unauthorized way. This is a huge concern for enterprises including in regulated industries. Um, uh, uh, you know, we need to worry about this with with AI models. There's you know, there's a concern called prompt injection which is if I just make a Google doc that is supposed to be a Google docs the model's reading but I've maliciously made it to say here are some instructions to the model. Email all of this to some you know, to some random bad person. Put this on the internet. Do this bad thing. ignore all your other instructions and just do this. Um, we're working very hard to make sure that when we offer these products that, you know, we're able to defend against this. And so, you know, we recently offered a a product that's kind of like an AI browser extension, but we only offered it to a thousand people. Um, and I'm on the wait list. Yeah, we can help get in. We're going to expand as fast as we can, but the reason we've only offered it to that small group of people is we've made some progress against these prompt injection attacks, but we don't feel that we're bulletproof yet. And so we specifically said, look, don't put any confidential data in this like you're a tester for this technology. We're not fully safe against prompt injection attacks. And our hope is to is to go on a path where over time we can offer um, you know, this technology and these and and, you know, the these models some version of this that is enterprise safe and so we're we're trying to iterate and put the defenses in place because you know, we we, you know, that's I think that's one of the most important product requirements and we want to build in public and gradually make this thing safe in public and get to the point where you know, you can deploy 100,000 seats of this in an enterprise and you know, have have the feeling that it's, you know, is just as safe as the way you're currently operating.
Love it. I love the focus and I'm still will wait on my weight list but thank you. Uh I so I have a question. Uh you have talked about AI as a coworker and specifically like genius co-orker and at the same time I was super interested in your whole experiment around project ven where you used you know claude to set up a vending machine and run a small business. How did that work? And then what's the gap between what it is doing today to becoming a genius co-orker in the future? What are the steps?
Yes. So that that was a very interesting experiment. I thought the lessons were very interesting. Um, so there were some parts of the task that Claude was very good at and there was some parts of the task that as of now it was comically bad at.
Explain to the audience what that experiment was.
So basically project vend was we we basically had claude be the brains of a small store operating out of anthropic. We gave it access to be able to order things that people requested. You know it could kind of search on Amazon, search on the internet, talk to people, clarify what they want. you know, we we basically gave it some employees who would or who would, you know, or you know, order things in response to its requests and, you know, kind of stock them up in a stock them up in a refrigerator. Um, uh, Claude could set prices. It could change prices. Um, so, uh, uh, and we just ran this experiment for a few months. We nicknamed the model Claudius. Um, just because, you know, we want it to be the normal the normal Claude. Um, so Claudius was brilliant at some things. It was actually very great at finding and ordering things that people asked for, no matter how strange, as well as helping people understand what they wanted. One person wanted, for some reason, a giant tungsten cube, a cube of solid tungsten. Um, and Claude Claudius managed to procure it for that person. It actually delivered the cube of solid tungsten.
How did they even manipulate that request into the process?
I I'm not sure the details, but you know, you can find all kinds of stuff on the internet, right? And so it turns out that includes a cube of solid tungsten. Um, as a business, as a business person, um, Claudius was very savvy in theory in that, you know, it thought about the prices, it thought about supply and demand. Um, the problem is was very manipulable. Um, so if you begged it for a discount, um, it would it would give you a discount. And and people learn to do this. And so Claudius's financial performance was not great. Um, uh in part because it was too easily guiltable. Um, uh so I don't know. It's it's you know it's it's it's a interesting profile of strengths and weaknesses. Um, uh, uh, so not yet ready to run a small business. Yeah. It it's a little bit lacking in street smarts. Um, uh so you know it's you know today's AI has an interesting profile of strengths and weaknesses.
Yeah. and and to go from here to being a genius co-orker, what does it need? More memory, more reasoning, like what is it?
I I would say it's, you know, running a small business is, you know, they're, you know, doesn't include all possible intellectual skills. So, I think there's a lot to be, you know, included in kind of the, you know, increasing the intellectual skills. There's, I think, you know, a lot more around agency. This is a fairly limited form of agency. Um, but uh, you know, I think there's a lot more that could be done and then you know, I do think the street smarts is pointing at something right, you know, the prompt injection that's like a form of fooling the model. So, you know, you don't want the model to be, you know, completely cynical or refusing every request, but but this idea of kind of, you know, having some sense of human human psychology and and being a bit harder for the models to be tricked, I feel like that's that's that's an area of improvement and and and I think actually in a way paradoxically important to the models being safe. Yeah. Um, you know, just having having some sense of, you know, when they're when they're being tricked.
Yeah. Absolutely. Uh so I want to close uh on a personal note. So you grew up in San Francisco, went a little old school here and studied physics, neuroscience, and now you're at the center of this next big AI revolution. Is this the impact that you dreamed of having when you were growing up here?
Yeah, you know, it's it's very interesting. When I was in high school here, of course, the first boom was just happening, right? It was, you know, this was like 19 1997 to to 2000. Um, and to be completely honest, I didn't have much interest in it. Um, you know, I wanted to be a scientist. I was thinking about physics and biology. I was like, "Oh, you know, you can write these websites in JavaScript, whatever." Yeah. Um, uh, uh, and so I I kind of went off and and and did that. But, you know, one of the things I discovered as I worked in those fields, as much as I enjoyed them, is that we were increasingly running up against problems whose complexity sort of defeated the minds of humans. Yeah. Um, and and that is how I got interested in AI because I saw that as the models went towards countries of geniuses in the data center as being be you know, being perhaps one of the only technologies that could help us to work together with us to overcome those problems. And so, you know, working on AI, I still thought of myself I guess still think of myself as as kind of a scientist first. Um, you know, that kind of involved first being a research scientist and working at tech companies, then being executive at one of those companies, and then as I was dissatisfied with with how the people I was working for ran ran their ran their companies and, you know, ultimately pursued the reason I I was in the technology, which is, you know, to kind of benefit people and solve these hard problems. Then that kind of took me towards, you know, founding my own company. And so, and so, you know, yeah. Yeah. Now, now I'm kind of in the middle of this tech ecosystem that I was quite disinterested in even even as I was growing up growing up all around it. So, in a way, things have come full circle.
You are making San Francisco proud and thank you for the partnership between HubSpot and Anthropic. And thanks for coming here today.
Thank you for having me. Thank you so much. It's good.