Always been the case that if you have a really good inference algorithm, you just need to pump it with a good dataset, and you'll get better results. Most people who are paying attention believe that the future of knowledge work will be primarily done by agents because there's so many things you and I do on a daily basis that you can hand off to an agent. ClickUp has like nine different applications in one, plus we have probably the most complete integrations ecosystem of any platform out there. And our biggest bet is essentially how can we fully leverage the power of agents in order to maximize the productivity of human workers and human companies that are on on ClickUp. Fast forward again, twenty thirty. What's like the future future state? Twenty thirty. I foresee a Alright. We got Jay Hack here, head of AI at ClickUp. Welcome. Thank you so much. Alright. I wanna start with some background on you, because I wanna get into where ClickUp's going and what bets you guys are making on AI, I think it's gonna be really interesting. But I also wanna go back. You were lead data scientist at Palantir. I don't know if lead data scientist. I was a lead on several projects. Okay. But at the time, the company was about a thousand people, so Thousand? Fresh out of college. I feel like everyone knows Palantir, and I feel it may have been less known back then. And in interest to the wide world, I think with AI, they've just been, like, blowing up in the names everywhere. Their, you know, forward deployed engineer was like famously, you know, coined by them. What was it like working there? Yeah. Well, you know, I think to me at the time, it didn't feel like they were everywhere, because I went to school in Palo Alto at Stanford. Yeah. You know, they had this this huge campus. They basically took over downtown, and, like, you would just see people walking around everywhere with a Palantir shirt. So it's, know, back in the media now, especially with, you know, some of the the sort of recent escapades the US is going on, but it did feel like it was ever present back then. Yeah. I mean, my time at Palantir was fascinating. I learned so much, and it really kind of set me on the the course that my career took ever since. My role at the company was essentially we would be doing pilot projects. So we would go to, like, a Fortune five hundred company, maybe it's like a bank, and we would try to sell them on Palantir software, which was essentially a way of assembling a data lake and then building certain custom applications on top of it. We would use Spark to, you know, pull data from all different corners of an organization, put it in one place, and then we'd build them some type of like an actionable workflow on top of it. And so I was the guy that would sort of airdrop in after we had the data assets set up, so the pipes were kinda flowing. And then we would say, okay, what's something interesting we can do with machine learning here? Because that was my specialty. So just to name a couple. My favorite one was in Japan, we were working with a a convenience store chain, and they wanted to determine what are the what is the optimal assortment of rice balls that you can put on the shelf such that you optimally capture the demand, but you don't have any spoilage. And the thing about these convenience stores in Japan is at least at the time, they would deliver them three times per day. And so it's like literally this online optimization problem you're trying to solve is like, okay. Based upon, you know, whatever features we can extract about this convenience store, maybe it's near a middle school, maybe it's near an Apple store or something like that, can we predict what the demand is gonna be at a given point in time, and then like literally run like an integer program, is an optimization, discrete optimization to figure out what's the optimal assortment of stuff. And this is right out of college that you're doing this? Yeah, this is my first job out of college. That's wild. Okay. So then you go from Palantir to you you founded Mira That's right. Beauty. Yep. And then but even before that, you were if if my notes were right, you like, it was in college, you had something called, like, CB Chess, which was, like, basically taking, like, images of three-dimensional, like, chess boards of, like, in real life and turning it into two d using heat maps to predict or to say where the pieces were moving. Is that right? Yeah. You know, this is just a a school project, but it's the first thing I ever did that went viral. And so it kinda set me on my course of like being a little bit more active on social media. And the the nice thing about computer vision is it demos so well, even pre deep learning. So this is in twenty, I guess, twenty thirteen maybe that this came out. You can just show an image and people immediately get what you're talking about. Natural language processing is a little harder to demo unless you have something interactive. The idea behind CB Chess was you could take a laptop and put it next to a chessboard and watch people play a game of chess. And then from that video, the question was could you reverse engineer basically the sequence of board positions that you go through? Nice thing about chess is that you have this perfect grid board, and so it ends up being the case that there is actually a well known transformation, assuming you can detect the corners, but transform that to like a perfect grid that would just fit in an image. So you're finding the straight lines, essentially. You're finding the corners. Corners. And then if you can find enough of the corners, you can basically infer what that would look like in a grid. And then on top of that, you can actually look at, okay, we have certain constraints on how chess pieces move across a board, and so that gives you, you know, a lot of assumptions on like what the possible positions you could be in at a given point are. And then you can also look at, you know, basically just what are the pixels that are filled in on any given square, and that allows you to basically back out where the pieces are. Still get comments on GitHub about this. Some people like text me on Twitter and say like, hey, CB Cheth. Have you always just been amazing at math? I think math is my first love, I would say. I mean, you know, the experience of being into math is you're just constantly confronted by people who are like way better than you. And so I definitely don't consider myself to be like a math whiz, but it's what I do for fun. Is like I love three blue one brown videos on YouTube. He does such a good job of explaining different math concepts. Fun fact, classmate of mine in college. Yeah. And you're essentially taking, like, a large amount of unstructured data and making it structured. Right? Like, that's with a lot of these things, whether it's Mira went on to I think what you guys did was pulled aggregates of, like, customer reviews and all these things of beauty products to be able to say, not just use, but, like, visual data and everything to be able to say, is this actually going to do what you want it to do? Is this gonna get you the outcome that you want so that you can purchase something that's, you know, cosmetically appealing, I suppose? Yeah. Would say that is a consistent thread throughout all the problems I've worked on in my career is that, you know, we have these very powerful algorithms on the one hand that machine learning has developed over the course of the last, you know, sixty years or whatever. And the bottleneck a lot of the time is actually the data they have access to, and now we would call that context. But it's always been the case that if you have a really good inference algorithm, you just need to pump it with a good dataset, and you'll get better results. So that's what Palantir was all about, basically assemble all the information from an organization in one place and then predict useful things on top of it. Mira, similar. It's, you know, really targeted this problem of specifically shopping online and how can you assemble a a case for an individual for you should or should not buy this product. And a lot of time that boils down to, like, oh, take a photo of your face and match you to influencers or analyze reviews and ask the user questions about what they prefer. And now we're doing that at ClickUp. You know? And it it's very similar problem at in in essence is basically agents are limited, at this point, basically, by the context they have access to. And if you give it nothing, it's only gonna be a chatbot. If you give it everything, it can do what a human can do and and much, much more. So And the way you got the ClickUp was through you left Mirror. You founded CodeGen. I think you exited Mirror in twenty twenty one, founded CodeGen twenty twenty two. That's right. And CodeGen was was it did you call it ticket to pull? So explain it. I from what I understand is essentially saying, the ticket is go fix this thing, and then you figured out how to get it to essentially before Claude code or any of these things to be able to go say, alright. What's the context of that? Let's go look at the enterprise database architecture or however this the code is put together and go actually build it and give it back to a human to review the package. That's a fairly accurate account. I would say credit where credit is due. In twenty twenty two, Copilot already existed, Microsoft Copilot, which was the or GitHub Copilot, I guess, which is like auto completion in your IDE as you're writing code. And I think to many people who are paying close attention, myself obviously included, it was obvious at that point that code was gonna be the first place that machine learning really hit. But at the time, what everybody was focused on is basically you have your IDE, your integrated development environment, and there's like a sidebar so you can maybe chat with it. It'll autocomplete text as you're writing. But it was really these like very short hops that a human was taking. And my reasoning was it it makes sense to sort of skate to where the puck is going. It's imminently going to be the case that you're able to assign a much larger segment of work to an agent, and we'll go off and it will do things for you, will take multiple actions, edit multiple files in the case of code, then give you back a finished artifact. And so we were the first asynchronous agents. The word agent kind of I mean, the word agent has existed since, like, nineteen sixty. We were the first one to, like, really try and launch that in an enterprise coding context, and the ticket to pull request comes from this idea of you have a linear ticket, which is specification of what you want. We would spin up an agent on it. We'd go off it, write code, and then it would come back and give you a pull request. So that's the genesis of it. So, okay. So you felt like you were one of the people towards the beginning that saw where things were gonna go. Because we're everyone's seeing it today. Right? There there companies or big companies a couple years ago saying you know, a year or so ago saying eighty percent of their code is written by AI, and you're kinda like, really? And then now it's like, why isn't it a hundred? Right? So I guess question is, did you has your point of view changed on where how AI will affect engineering and developers in the past since that point to now? I think that in broad strokes, the thesis that I had at the time has largely come true. You know, a couple things did change, though, that are, like, more of my new points. So one of them that was not clear at the time back in, like, twenty twenty three is the extent to which this would be a focus of the major foundation model apps. There have been a couple revelations specifically in code generation. One of them is this notion of positive transfer where if you train a model, a large language model, specifically to write really good code, it ends up getting better at everything else. And there's good reasons behind that. It's kind of like, you know, code is like a explicit encoding of human reasoning or logical thinking or whatever. And so if it can do that really well, then it'll be able to write better poetry because there's some amount of, you know, logical thinking behind that. And so as a result, it is basically in the best interest of any of these Foundation Model Labs to make their code, their agents extremely good at coding. And there's also many other things that play into that such as it is probably the easiest place to generate synthetic data. And by synthetic data, I mean, you can, you know, generate a program and then ask an agent, is this gonna run or not? And it's very easy to basically do this. Right? And so that has been a major market force that I don't think I foresaw the extent to which these folks would lean into it. And now you have probably about ten companies that have all raised more than a billion dollars or have at least more than a billion dollars of capitalization directed towards specifically asynchronous coding edge. The extent to which that would be competitive is not something that I Yeah. Anticipated. Okay. So now you're at ClickUp. Yep. Bigger company than obviously, Palantir is massive, but, like, compared to the last couple startups, had a bigger company now. Do you feel like you are able to move at the speed you want to and have been able to in the past? You know, so the thing about, like, speed in twenty twenty six is there's no company in the world that moved at the speed that, you know, a company that's fully leveraging AI today does. So I think that it is, like, a hundred percent the case at this point that code is no longer the bottleneck. Once we align as an organization around this is the feature that we wanna ship, the time from, you know, banging the gavel and saying this is where we're gonna do it until being in front of a customer is crazy short. I think, yeah, if you're at a company of like ten, twenty people or something like that, like, obviously, you can just YOLO something into production and nobody's gonna stop you. There's security controls. We have data privacy. All this other stuff we have to take into account, but it's definitely not the code that prevents it from being, you know, you from shipping very quickly. Also think that ClickUp, more so than any other company I've encountered at that scale, is very much so like a ship fast startup mentality type of company. I don't think it would be around, you know, if if it wasn't. Like, they've just been able to build so much surface area and, you know, ship so many features. So, no, I I I feel like it's actually quite comfortable to the startup experience. I remember for so long, I would see ClickUp billboards everywhere. I used to travel all the time. I would see them everywhere. And I had never used it. I was always on Monday or Asana or one of these other ones, and I get the cadre, and we they were using ClickUp. And so I was like, alright. I wanna finally figure out, like, what's thing can do. It's been really impressive. And like what you said, like, there are a lot of features that feel like they get released without this massive, like oh, maybe I've just unsubscribed from the emails. I don't know. But, like, I don't get this massive, like, this I feel like maybe the days are gone where you're like, there's a product release coming out, and here's all the specific things where, you know, it's just feels like and maybe maybe that is the case still, but it feels different. But it feels to me that it's just new features are just there. You're like, oh wow, this is cool. ClickUp is incredible at that. I think more generally, it is the case that there's gonna be an emergence of these like horizontal platforms that have really good primitives in them. So they have like good data models that represent like chat and documents or whatever. And the interface layer on top of that has never been so malleable. Right. And there are so many instances where it's actually much better to have like a bespoke workflow for whatever you're trying to do just to reduce the number of clicks compared to shipping some static Kanban board or whatever. And so, yeah, ClickUp is a hundred percent lean into basically, like, reconfigure things, see how it goes, like, always, you know, dynamically modify it. And I think what you can look forward to is, fortunately, agents are really good at doing what I just described. And so while it may be the case to say that still there's some amount of, like, product release around, like, a new type of view landed in ClickUp, I think imminently in the future, it's gonna be the case that you'll be able to essentially say to an agent, hey. You know, I'm trying to schedule a bunch of podcast recordings. Here, you know, here are the users, research them, and then make me a bespoke interface that allows me to go through and vet the different, you know, ideas for questions to ask them or something like that. And it's gonna nail it because, as I'm sure you're aware, Opus four point six can, you know, one shot this stuff so well. Yeah. Okay. So you've got you've got at ClickUp, you have this, I would imagine, an internal mandate to say, let's not allow any external party who's using an API to be able to do something with ClickUp data better than we can do it natively. Right? So, like, where do you like, what are the bets you guys are making in terms of where people wanna use AI? In well, you do, like you said, a lot of broad landscape in terms of what people are using ClickUp for. Where do you think the bets are? Well, you know, I think that the the mandate, as you phrased it, is not let's stop people from doing things with ClickUp data. I would definitely say if somebody is able to build a better feature external ClickUp than we can do internally, we would use that as a source of inspiration. Yeah. And we would say, hey, you know, maybe because of, like, various latency or software engineering reasons, we can do a better job at this because it's first party. But we definitely enthusiastically encourage people to build cool stuff external ClickUp. In terms of the bets that we're making, so ClickUp, the platform for people who are not aware, I think of it as, like, nine apps in one. It's like a super app. In fact, it may be one of the most ambitious front ends in the world because even something like G Suite, which has, you know, you have like sheets, you have docs, all this other stuff, they're on different pages, whereas ClickUp has like a chat thing. It's like Slack. There's tasks. There's whiteboards. There's documents. There's a Zoom competitor baked into it. All this type of stuff. So all in one application. And we, you know, essentially power nontechnical knowledge workers. I think that most people who are paying attention believe that the future of knowledge work will be, at least as we know it today, will be primarily done by agents because there are so many things you and I do on a daily basis that you can hand off to an agent, we will do a very good job at it. And so our biggest bet is essentially how can we fully leverage the power of agents in order to maximize the productivity of human workers and human companies that are on on ClickUp. And so to that end, a couple different initiatives I can point to. One of them is what we call super agents. This is essentially an agent with all of the tools. So you just airdrop it into your ClickUp workspace. It can do anything a human can. You can run it on automations. You can ask it to do specific tasks for you. So I mentioned this deep research thing a second ago where you say, go find me a bunch of podcast guests, make a task for each one, propose a couple questions for it. That's like a bread and butter use case for us. But also, I have one that, you know, every day we'll monitor our GitHub. And anytime somebody edits the prompts of one of our AI applications, it'll DM me. It'll say, hey. Just so you know, this guy over here modified this prompt to do that. Just wanna flag this for you, that type of thing. So it's sort of like the things you would maybe historically hire an intern for, but I think a lot of the power behind it comes from the fact that, you know, like I mentioned earlier, ClickUp has like nine different applications in one, plus we have probably the most complete integrations ecosystem of any platform out there. And so you can have one agent that basically spans all of the platforms that you would interact with on a given day in your browser. Saying it back, what you ultimately want would be for people to not use agents through anything external. Use them through ClickUp. Allow ClickUp to have the first party data of your, you know, project management and whatever else you're using it for, and allow it to have these external connectors to whatever other data and tools you need in order to bring in. You know, it's an interesting question around how do we feel about people using external agents. I mean, the mission of ClickUp basically is is to replace all software. So we think there's gonna be essentially a single platform in the future that you use, and it it'll meet all of your needs. And I don't think that that precludes people using external agents. You know, we have an MCP. Anybody can use it. I think you can be quite efficient working in ClickUp using, you know, Cloud Code or something like that externally. We see that the people who adopt that by far and large are actually technical users thus far. And ClickUp is very much so targeted, like, non technical teams, and so it hasn't had the same level of uptick. But I also think that if you extrapolate outwards, there's just no way that somebody will be able to make an external agent that is as cost effective, efficient, knowledgeable, dialed in as something that we would bake first party. You know, it's integrated into the very foundation of ClickUp, has access to all of our data models, you know, we can optimize how the tools interact with it, etcetera. So I I foresee a a multi agent future where, you know, everybody has their preferred client. Maybe it's ChatGPT, maybe it's ClickUp Brain, something along those lines. But also, the majority of things that actually get done in your company are done asynchronously in the background by cost optimized, efficient, and company specific agents that are run within ClickUp. I always like to ask about the future. Twenty thirty is a good mile marker. It's not too far away, but, you know, I'd be here before you know it. But it's also far enough with the advances that are happening right now that things could be drastically different. I guess the biggest question I have for you is, do you think everyone you know, we talk about you hear about sass apocalypse and everything can be recreated. Sass mageddon, yeah. Yeah. Yeah. One of our our in order for engineers to get an interview at our company, we are typically asking them to do this test where they have to rebuild, like, a major application that we give them in two hours and do it live. And they can't do it, I was like, I can't get an interview. So with the ability to build SaaS and replicate SaaS so quickly and you guys wanting to be the one SaaS, where do you think the future could be for ClickUp when it comes to you know, productivity sounds like the metric. That's the metric you wanna go after. Does it really matter if there's this massive interface, or would it just be a chat interface that is, like, listening to you work, taking in your fireflies, and telling you you have a task without even showing you a board. No. I mean, this sort of gets at what I mentioned earlier, is the idea of generative UI. I do think that assuming you have the data asset in place, the right data models that represent, you know, the primitives that people use for productivity, the interface layer is something that is, like, infinitely configurable. There's always a better UI for a given workflow that you're going through, and I do think that it'll be dynamically generated in a lot of scenarios. You know, we we actually have a product like this internal to ClickUp. We'll be launching it soon. It's called VibeUp. It's essentially vibe coding within ClickUp, but That's cool. The interface is connection directly to the underlying ClickUp data. So it'll automatically sync, and you can make like a, you know, podcast interview interface or something along those lines. It's amazing. You know, what we see is that people will create one, and they don't really need to modify it. So it's like you kind of saturate your personal workflows at a certain point, but I can certainly imagine a future in which it will fully dynamically spin up UIs. And that would work with dashboards and everything as well. Totally. Yeah. Yeah. Come in and say, hey. You know, I'm the CEO of company x. How's my business going? Tell me about it. And this thing, agent goes off, just go bing bing bing, calls a bunch of tools, and then whoops up a dashboard for you, shows you, you know, this is the conversations people are having at the company. Here's how our supply chain, which is managing ClickUp tasks, is doing. Here's a summary recently of what's happened, a little card down here. Yeah. I think that, you know, there some people would tell you that, like, no code's failure, which is objectively the case, no code or or low code at least Yeah. Didn't end up panning out really. That implies that this idea of generative UI is not gonna work in the future. I think there's still a order of magnitude difference between how easy it is to say, I'm the CEO. I wanna learn this thing, and boom, you get a dashboard versus having to learn some bespoke, like, graph connecting language for no code or low code. And it's so so convenient. Well, what's cool about the idea of what you're you're saying is it encourages people to give more first party data to the platform that's going to to do this because you want it to have as as much context as possible. I have a background in marketing, I'm notoriously the type that, like, I don't want no cookies, no anything, don't serve many ads. I have no ads anywhere. I don't watch anything with ads. But here I am, and I'm I'd be willing to wear a necklace that just recorded my whole day Yeah. Because I kinda want the LLMs to have that that context so I don't have to give it as much context. And I'm just like, hey, what did I do today? What's what follow ups came up today that I should have action on? The more data that people are using if people aren't using your chat feature because they're using something else, this is incentive to use it because you're feeding it more context to be able to answer the questions. Right? I think that's absolutely right. I think one of the biggest taxes on productivity in the world today is Slack's MCP is so limited. To your point, you know, you you can do stuff like you can only search for messages within a finite time horizon. You know, they don't let you actually index the data yourself, and so you can't do stuff like, hop through a Slack message to a Google Doc or something like that. Basically, it means that anytime your agent wants to go look at Slack, it has to like go off on a side quest for a while and then come back. And there's a much more efficient way to do that. So yeah, I do think that there's this convergence of all the different data sources, and the people who can get the most context that represents what they're trying to accomplish in one place are gonna be the best able to leverage agents. Last question. Fast forward again, twenty thirty. Right now, you've got ClickUp using its own your dog food. Right? You're you're you're testing this thing out where they could you vibe code your own thing. What do you think you'll be vibe code, or what do you think you'll be testing in dog fooding at that point? What's, like, the future future state? Twenty thirty. I am very bullish on Neuralink and basically brain machine interfaces. I think there's a number of ways that we can get there. I think that is like the true step function improvement. You know, an interface is basically using like electromagnetism in the form of visible light to communicate information inside of a computer into your brain, I don't think that's necessary, to be honest. I don't think this is, a ClickUp program that we would run I think we'd probably piggyback off of, like, Neuralink, but I can totally imagine a situation in which you are, you know, your brain is manipulating, like, the ClickUp MCP in order to make modifications inside the program or, you know, update the status of tasks or communicate with other people, all this type of thing. You really believe that that's within four to five year horizon? I think that there will not be mass adoption of it in four to five years. But I think if you're saying, what is the type of thing that you and I would be sitting here getting excited about that's like demonstrating results on Twitter, and like you can get it for ten thousand dollars? Yeah. I I I do think there's a couple approaches. Maybe not full readwrite access for a brain machine interface, but I think definitely, you know, something along the lines of like, you can initiate a Google search and then get an intuition about the results that seems reasonable in that time frame. Awesome. Alright. Jay, thanks so much for coming by. Thank you. Appreciate your time. Likewise.