Why introduce an intermediate currency that is not dollars? What does that do to the model? Why couldn't it just be dollars instead of credits?
You can sort of change the rate card, so to speak. We're actually a multi-currency product, as you would expect. We needed a currency that says, okay, well, things cost varying amounts or whatever. And so one of the decisions we made is a credit is a credit. So we have a unified currency across all of HubSpot. And the only thing that varies is that some actions might cost multiple credits instead of one credit. So the idea was, can we keep it simple, come down to a single currency, and then we can sell that currency, so to speak. And these kind of credit-based models have been around for a long time. We've had them in the mobile industry with minutes and things like that, and then that, you know, obviously evolved over the years.
Welcome back to the Get Paid podcast. I am your host Manny Medina, co-founder, CEO of Paid. This week, we're doing something a little different and we are re-releasing an episode we recorded a few months back. When I sat down with Dharmesh from HubSpot, and even now I keep having people asking me just how much they enjoy this episode. So, I wanted to bring it back. We talked about HubSpot's move to credits, the future of vibe coding, and much, much more. This is so relevant right now because they actually executed that move. So, I wanted you to hear what was his thinking before they did it. If you missed it the last time, you're in for a great one. We'll be back with a new episode next week, but in the meantime, enjoy this episode with the legendary DS. And as always, let's get paid.
DS, I'm so excited to have you. I have a long catalog of questions. I have no idea how far this is going to take. I, you know, I had an agenda in mind, but then halfway through it, I was like, you know what? Let's just roll with things as they come. So, let's start like, you know, thematically, let's start from the top. OpenAI released the strawberry, had the strawberry release, I think it was November, maybe December last year. It was the o1 model, right, that in that birth reasoning to the scene. What has gone through your mind since, like, since you saw the model all the way to now? What's exciting? What's worrying? What's the let's see what happens, the full Dharmesh of this.
Yeah. And I'm thinking back, so I've had multiple kind of moments of wow with GPT at large, back to GPT-2, that, you know, Sam gave me access to before that it was a ChatGPT app. So I just had the API and had to code my own minimalist chat app, and I had a conversation with it. And even then, you know, this was two years prior to GPT launching, I mean ChatGPT launching. It was amazing, and you sort of had this, and I was, you know, late night emailing Brian, and I sent him the transcript of the conversation that I had, which I still have for historical purposes.
What was it, the context of the email? Look, look what we can do now.
No, look what this thing can do. It's like, this is an unedited transcript of a conversation that I had with AI, right? Like I just laid it out there. I'm like, this is going to be big. And it took a while for the world to catch up. And I think, and then when ChatGPT launched, everyone sort of figured out what I had seen a glimpse of, which is like, this is real. And then we've all had our kind of magical moments, like, I can't believe software can actually do this, right? And so that was kind of moment number two, is sort of witnessing, and obviously ChatGPT and GPT-3 was much better than 2 was. 2 was very rudimentary. And then so 3 sort of awakened the world to what we might now call this kind of single-shot prompting, you know, no reasoning, but still very, very good at kind of building this world model and a broad range of prompts. What reasoning changed, at least in my head, is that, you know, we had always thought, we collectively, the industry, or at least most of us mere mortals that were not working at the frontier labs, were like, okay, well, I sort of understand the nature of scaling laws that says we can throw more compute and more data at this and it'll get better. And reasoning, what that changed was that, okay, well, not all of this kind of learning necessarily happens, or the compute investment has to happen in pre-training. It could actually happen post-training at inference time that says, okay, if we just give the model, with the existing data that it has, just time to sort of self-reflect and quote-unquote think, it can come up with higher quality answers. And that was groundbreaking both in its simplicity and elegance, and we can talk a little bit about, you know, how it actually works, but then it sort of opened up an entire new dimension of what can be done with these models, right? So now we're not limited simply by how much compute we throw at it in pre-training and train the model for hundreds of millions of dollars and, you know, go on for months at a time. It's like, now we can say, oh, you know, we can throw inference-time compute at it. And what to me that unlocks is that that's something we can sort of dial up and down as well, right? And we're starting to see that in some of the newer reasoning models that says, okay, I want you to overthink this particular coding problem, or I want you to dial up the reasoning, use more tokens.
Yeah, you have to think fast, and you have like, go for a long time and go think a lot about this.
Yeah. And so I think what, and from a business perspective, I think the thing that it unlocks is that it can now do reasoning, that means it can do a job at, you know, what I think of as like functional decomposition that says, can I take a larger order task and kind of have the model itself say, oh, here are the things I need to accomplish in order to get that task done. It could have effectively a short-term memory to keep track of what it has and hasn't done and kind of cycle back and can follow multiple paths in parallel to say, here are four ways to solve that. This one seems to make more sense, and it's all doing this sort of quote-unquote in its head. And that unlocks a bunch of business use cases that, you know, previously weren't possible with just a non-reasoning model. So I'm very excited about what that does.
So when you think about like the things that you're excited, things that you're worried about, and the transformation that SaaS is going through right now, and like you're living in your own skin at HubSpot, like just SaaS in large, what is this ability to reason now and inference-time compute bring to bear in terms of jobs, things that we can bring to the market, you know, threats and possibilities?
Yeah, a few thoughts. One, on the jobs front. You know, I tend to lean towards the kind of optimistic side of it. In terms of technology, I know AI is different from prior waves that have come, but there are some similarities. So I think there is going to be some disruption because AI is going to be able to do more and more. We're moving from relatively kind of low-level tasks to we're going to move towards higher and higher tasks. We're seeing this already. It's moving very, very quickly. So I think that's definitely going to have an impact on the job market. The optimist in me says that AI will kind of giveth more than it taketh away in the kind of grander scheme of things, which is, and we've seen this with automation and technologies in the past, that says, okay, well, there were roles now that are just unnecessary because those tasks can now be done by technology, even in the pre-AI world, and we sort of accept that. I think it just sort of elevates humanity at large to say, okay, now we can take kind of our specifically oriented brains and apply them to things that are more creative, less of a grind, and I think increases, you know, human potential. So that's the optimist in me. And what it, you know, as a software entrepreneur, what it does from a business perspective, which the reason we should all be excited is that we can now do with AI using software things that we never imagined software would be able to actually do, right? We've always built tools and we've said, oh, we can increase the productivity of X role by Y% or something like that, and that's great, and we've built, you know, lots and lots of market cap for lots and lots of companies and driven lots of productivity for our collective customers. But now it's like, it's a whole different game, right? It's like, okay, well, it's not just about increasing productivity by a marginal percent, it's like we can sort of reimagine what we can do with software, which is both exciting from an opportunity perspective, but then also, and there are, you know, questions that float around, around, oh, does this mean SaaS is dead, right? Because AI is going to just basically be able to do everything, people are going to vibe code their own applications, why do we need software companies, you know, that kind of line of reasoning. So I actually love your post on this where you said like, what, and I like this one thing where you said the reason we're not doing this at HubSpot, even though we can, anybody can, is focus.
Is focus. Is that focus means that there are some things that you decide you're going to do and do very well, and some things that you decide you're not going to do, you're not going to do very well. So you're not going to code your own HR application or payroll. Turns out that that could be a little hard, and like a lot of lives, you know, depend on getting paid.
Yeah. And even so, let's assume like even if we don't go after the big ones or even the hard, like payroll is a super hard category, regulation, things like that, even the easier ones like expense tracking, project management, pick your thing of choice.
Yeah, yeah, yeah, yeah. Expense management, right, right. Collaboration, any of those things. Yeah, yeah.
Yeah, and I think I put this in the LinkedIn post. I was proud of this line, so modestly I'll put it back out there, which is, you know, companies should be focusing all available calories on things that create customer value. Period. Right. So if you start with that principle, it's like, okay, well, yes, we have the engineers, we have the wherewithal, LLMs are definitely good enough to build, you know, relatively robust applications now, but just because we can doesn't mean we should, whether it's a low-level app. It's like, okay, well, if there is an existing SaaS application that mostly meets our needs for whatever thing we're trying to get done, we should 100% of the time be using that application. Now, if it doesn't exist, I mean, there's definitely reasons like, oh, we need this particular thing automated, this thing, this tool, that it's too small of a problem, there are no real legitimate software companies out there that solve it, fine. But if a software application that you can buy off the shelf that gets you even like 80% of the way there, I think it would be foolish to try and replace that.
I think every startup out there is just cheering you on, like, go, Dharmesh, say it louder for the people in the back.
Yeah. So on a related note, we should, because it's, you know, just talking about the business aspects and the impact of AI, there's been kind of discussion around, will AI, you know, kill SaaS as another kind of related theme. And I haven't written a post about this, so this should be next in my lineup, is that...
And I can see it in your head, so let's get out that post.
It's definitely in my head. And so here's, if we go back, and you've been around enough to see this, when, you know, when Salesforce came along and introduced the idea of SaaS, built a massive company, super successful. And all credit to the team and to Benioff, because it was truly visionary and it was interesting. So at that time, you might recall, they had this, they put on buttons and everything, like the no software thing, right? They had the word software crossed out with the big white button, red slash going through it. And both Benioff and others knew that he didn't literally mean no software, because obviously Salesforce was a software company. What he meant, and that the shorthand was software as we knew it, as it was sold, as it was deployed, no longer makes sense. This is the new world. SaaS is the new world, right? And he was exactly right about that. So I think what's going to happen now, so by the way, and what that changed was, oh, instead of running things on-prem, you're going to run it on the cloud. As a result of that, we're going to have multi-tenancy. As a result of that, the business models can actually be different. We can charge differently for software than we may have historically done with a, you know, large enterprise license and a 20% maintenance trailer, right? There was all that that existed, and so there was a set of kind of dominoes that fell together as a result of that kind of movement, movement to the cloud. And so we all know that, and it's been a great run for software companies, but the reality is with the introduction of SaaS, what did not happen was the end of software, right? What happened was the blossoming of software. It was even better times for people to be building software companies. It just needed to be built on this new model.
Can you make the argument, Dharmesh, that SaaS brought software closer to the customer, or did it really do nothing?
Oh, it definitely brought us closer. I'll give you one tactical example, you know, because I've had, you know, my first startup was, you know, pre-cloud, pre-internet, and the way we had to sort, and it was in the enterprise software category. So we would build products, we would send them on CDs and floppy drives even, you know, back then. And the hardest thing as a software entrepreneur, as a developer, is that you could ship a product out there and you had literally no idea what percentage of your customers had installed the latest release that you had fixed, you had what the usage was. It's like, does Soa use it, you know, five times every day or never? Like, we had no...
The software didn't call back home to report itself.
Exactly, it had no way to phone home, right? And then we got some instrumentation back over the fullness of time, but that was effectively the kind of the standard model. And once we went to the cloud, it's like you're shining this, you know, bright light onto your customers and what they're doing and what they're struggling with. And we now have such kind of rich visibility into what our customers are doing. So just by that alone, I would never trade SaaS for the older model. Put everything else aside. Let's say we didn't have multi-currency, didn't have, you know, but just that alone was worth a shift, and it was transformative for us as builders.
I think that point of understanding how people are using the software and being able to see it with your own eyes, I think that is probably the most important contribution of SaaS to bringing software closer to the people. So what happens with agents now? What happens with AI? How's that changed?
Effectively a similar thing, right? And we may not be running around with labels that say, you know, SaaS with a red line through it, but effectively I think what's happening is that we're going to be working at an even higher level of abstraction as builders, right? So before we've had, you know, and the languages have gotten better over the decades that I've been, you know, I've been at this for thirty-plus years now, and you see, you know, we went from assembly to C to C++ to 4GLs and all, and we've had this kind of movement where we can work at higher levels of abstraction. AI takes that up a massive, massive notch, right? So as builders, we can say, and I mean builders now more broadly, doesn't just have to be programmers, it can be builders of any sort, can sort of take the ideas that are manifested in their head and with a relatively high degree of fidelity translate that into like working product for some definition of working, right? It's like, okay, that wasn't possible before. So now we can prototype quickly, we can iterate more quickly, we can produce products, launch products more quickly than we had before, right? There's a step change already. Even if, let's say, let's say tomorrow something happened in tech, this is a complete sci-fi scenario, and it's like the models are not going to get any better. What you have is what you have. Everything's locked down now for whatever reason. We've all collectively decided we have enough technology in our hands as builders and software companies and people of the world to last us another decade, if not longer. We definitely have just to absorb that impact.
I agree with that. But you were talking, so the scenario you're talking about of like the higher level of abstraction where English is the new interface for building sort of, you know, speaks a lot of vibe coding. And you just posted on LinkedIn that there will never, there has never been and will never be a vibe-coded piece of software in the HubSpot core code. Is that a point-in-time statement or is that a forever statement? Like, feels a little declarative and final.
Yeah. No, it's, so here's a thing I actually meant. I'm going to have to go back through and edit it because it clearly came off unintentionally wrong. So I'm not suggesting, so we have vibe code or, you know, code built through AI coding agents in HubSpot today exists. To make it more accurate, at least ninety-nine percent of HubSpot engineers right now use, you know, vibe coding agent to coding on a daily basis, right? So this is not something that we're, it's like, oh no, this could never work, even including our most senior engineers. So everyone's using it, and it's part of the HubSpot product. The point I was trying to drive at is that this kind of idea of, because, and I prefer agent coding, but we use vibe coding, it's more popular, is that because vibe coding can produce code with increasing levels of accuracy, increasing fidelity over longer runs, you know, semi-autonomously, and that's all those things are true. The thing that I would posit, though, is that code is a small part of the actual solution of actually putting something in customers' hands, and especially if you're now, there's multiple dimensions on which you can measure the kind of viability of vibe-coded things. For instance, if I were building an application just for myself, vibe, absolutely perfect. Now, as you move to larger and larger user base or whatever, I think it becomes less perfect. Not that it can't ever do it or whatever, but that's kind of one criteria. If it's a temporary app, this is I need this one time a quarter or something like that. I'm not going to run it. It's not like running my entire business or whatever. A better candidate for vibe coding than other things. So the thing I think we're further away from is to say a single person is going to vibe code an entire software business and do all things necessary in order to make that software business successful because that's going to involve the coding. Yes, it's going to involve design. Yes, it's going to involve UX testing. It's going to involve customer discussions and customer feedback. And a lot of these individual pieces we may actually, you know, have AI help with. But I just, I'm not a believer that you can sort of vibe code your way into an iPhone app, and I just don't think it's sustainable. Yes, we have existing examples where, and you see this on the social media channels all the time, where so-and-so built a million-dollar app on the App Store, you know, vibe coding something, and they've never written a line of code before. Sure, we will see those examples, but as a baseline, I don't think that's going to be the predominant model. What I think will happen is I think we will have the equivalent smaller, like software teams, people that say, I'm doing this for a living for this particular category, and it's going to be in categories that historically were too small in some cases to actually pursue because the economics didn't make sense, but it's still going to be small teams of people that are devoted to building an application that they will use vibe coding tools for. But it's not just a kind of one-and-done. I can move to the beach and have martinis or whatever, and just the money keeps rolling in over Stripe, right? I don't think that's what's going to happen.
Let me, let me, let me get a little, let me turn the knob a little to the controversy side. So HubSpot has been patently a little late on the adoption of agents in their core offerings. A little, not so much. How do you introduce this new technology that is fast to build, fast to release in a company that's twenty years old? Is it part of the core fabric? Is it a separate team? Like, how do you accelerate? Do you think about this a lot? Like, how do you accelerate the release and the usability and the adoption of agentic-type software for a company that is already successful with a ton of users that has a culture that has already sort of like cemented it? What's the right approach?
Yeah. And to answer your kind of first question or one of your first questions, I think about this a lot. It's almost all I think about, right? It's like as a founder, and this is not just AI, as a founder, the thing that keeps me up most nights is, are we doing everything possible to achieve our ambition or to maximize the probability that we achieve our ambition, right? It's like, okay, well, and although I'm an optimist, the paranoid side of me is always thinking like, what decisions are happening in the business today that are going to seed the plant, the seeds of our own destruction? We just haven't seen it yet, right? It's like going to happen in a year, two years, five years from now. So that's, those are the questions I ask myself late at night. And so the way we're tackling this, and this is not necessarily new, so we have, you know, the entire core team, and we did like others have done and have talked about publicly, you know, in Q1 of '23, we threw out the roadmap we had at the time and said, okay, well, this is not a drill, this is not small. And so we had believers through the exec team, so it was not a lack of conviction on AI at all. And we moved, I thought, pretty quickly, but, you know, opinions can vary as to how quickly we launched things and got things out there. So, but the way we approach it is that there are the things that impact the core platform and the core product, and we should do those things and are doing those things and have launched a bunch of AI features. We have agents out there, you know, that are built and supported by the core team, like our customer agent, which is a nice, I'll call it mature for just on a relative scale as a category. But then we have, you know, projects within HubSpot that we internally call like these HubSpot Next projects, and these operate effectively as a startup within, and it's not an innovation lab per se. It's like literally a startup. So we make an investment decision around, are we going to put a seed bet down on this particular startup? It's going to have a founding team. It's going to have, and it is not bound literally by, I'll say any, because it's close enough rounding up to any of the constraints that HubSpot has. You don't have to use the brand. You don't have to use our billing system. You can create your own Stripe account. You can say you can build on your own stack. You can use whatever tools you choose. You can say, I want to go after this market. So we have a limited number of guardrails. One of which is you have to have, just like a startup would need to have if it were out raising capital, a thesis about the world as what do you think is going to happen in order for you to accomplish this. That's kind of one of the guardrails we have. The second one is that the would be huge if true. So we don't want these bets to say, oh, well, we're going to have a, this is going to be successful and we have an eighty percent probability of adding fifty million dollars ARR to HubSpot's bottom line. We don't care. Like that's not the charter for those particular projects. Those projects are, if it works, you know, it will be a multi-billion dollar business. It's got a chance of being a bigger business than HubSpot itself is, like those kinds of bets. So we have a slew of those that we fund.
And we're not sure, I buy the concept of a startup, like how's the org structure change? Who do they report to? They still have a boss, right? Like they're still in, like where's the startup end and where does the startup begin?
Yeah, it reports directly. So we have a, HubSpot org structure has gone through a change this year. We have a new CPTO, Chief Product and Technology Officer, that's head of product, engineering, UX, and these HubSpot Next projects. So it all effectively rolls up to him, and I'm biased, but he is freaking awesome. But anyway, of course I'm going to say that, but it's true. And yeah, so that's how it fits kind of structurally. But effectively the startup, so we have...
So that person will be like their board, if you would.
So we have a three-person board for these projects. It's me, our CEO, Yamini, and the CPTO are the three people that are on the board that make the decisions. And similar to how you might have, and so when we kind of put it together, and this is the first time I actually talk about this, but it's fun to kind of talk through in real time, is that, and I've been around, you know, startups and venture and things, made a bunch of my own investments. So I sort of know the mechanics of how that works. And so one of the heuristics that we use is that when possible, a HubSpot Next project startup should work like a venture-backed startup would. And the only reason to deviate from that is if we have a very good reason, like, oh, in this case it's different, so we should do it this way. And so for instance, it's like, okay, well, you know, we have a board. It's like, yep, we have a board that kind of oversees. So let's think of us as the GPs in the HubSpot Next kind of fund, so to speak. You know, we sort of make a decision around if there's going to be like a thematic area, and ours is, yeah, we still want it. We don't want to start pizza chains or, you know, going into hardware or something like that. It's going to be related to HubSpot. Doesn't have to be our existing user persona, right? But it's like it has to be go-to-market kind of related, but it doesn't have to be exactly our customer base. In fact, we'd sort of prefer if, in some ways, for some of the bets, for it not to be. And then we have a, what we'll call a, just like we would have in the VC business, there's a partner that's effectively sponsoring the deal, right? That's the lead partner on the thing or whatever. That's going to be more involved.
It depends. We don't have it down to a science yet, and I think it varies across venture firms. Our thing is, yeah, exactly, to really encapsulate the thinking. The way we've approached it right now is if one of us feels strongly enough about it, especially at the seed stage, and this is, you know, most of them are kind of seed-level bets right now. If it's a seed-stage thing, we don't need unanimous consent, we don't need forms in triplicate or anything like that. We just write the check and say if there's someone willing to go do it, we go do it. As long as we have, you know, a founder that's kind of passionate about it and the idea even obliquely makes sense, we don't want to second-guess it too much. Once we get to kind of a, and post-a rounds, then I think, and we haven't explicitly, you know, it's not the bylaws or anything like that, we have no bylaws for this, but my sense is that like a two-out-of-three vote would win. It's like if two of the three of us disagree on something, then it's like, and the third one maybe if they're passionate enough can kind of convince the other two or at least one of them to kind of squish their vote. But yeah, we haven't had to deal with that yet. It's a very collegial team, which has always been the HubSpot way. Yeah, we know what we're trying to do.
What other forms did you consider before you came up with the HubSpot Next?
I'll say this only partly in jest, is pretty much all of their forms. You know, this idea of trying to do innovation in a larger company is one that is not new to HubSpot. It's a problem. It has been for decades, and books have been written about it, and lots of talks have been given. I've listened to a lot of them. So we've tried variations of it. We've tried the incubation model where we say, hey, we're going to have this company. We're going to carve out one of the core product teams that's closest to the idea that we're pursuing, but we're still going to give them some constraints like, okay, you're still kind of flying under the HubSpot banner. You're going to have these things that has to move within a year or two. Numbers that we actually currently track that like make sense. It's going to move the needle out of numbers that we already measure. So it's, which, you know, these are, you know, what some might call like Horizon 2 kind of bets for those that follow the Horizon framework. So we've tried that variation. It's funny though, it's like I don't know that there's any real perfect answer. This one seems to, you know, having been at this for a while now and been very close to these kinds of projects at HubSpot, seems to be working the best. And it may be just a sign of the times that maybe this is the time to be doing those kinds of things because we're going through a transformation. But everything's, you know, the deck's being reshuffled in a way anyway. Yeah.
So if a CEO of a hundred million dollar company were to come to you and say, hey, I got to move faster. All my agentic projects are stuck somewhere in the product roadmap in Jira tickets. They tell me they're coming. They're just not coming. Like, what would be your advice to all those SaaS CEOs out there who are trying to like get back to growth?
I'll say this as founder CTO. My advice would be you don't solve that problem with an innovation bet or HubSpot Next or a startup within a startup. You solve that problem, and it might be painful, is that you might be looking at a team change or a leadership change or an org change or a culture change or all of the above, right? It's like, okay, if the team is not bought in, if there's passive resistance, if there's, and this is, you sort of have, and this is, we don't like to do kind of mandates and edicts at HubSpot. That's not the HubSpot way, as historically has not been. But there are times where like we do make a top-down decision that says, okay, this is what we're doing. This is not an experiment. And it's like, I think what happens, and I'll, and this is actually a good story since we're having this conversation. So the last time we had this level of, I'll call it an intervention, is probably overly strong a word, but like that is when we got into CRM. So when Brian and I made the decision...
