#158 Building Human-Centered AI Experiences with Haris Butt, Head of Product Design at ClickUp

Hi, I'm Maggie, VP of Curriculum at datacamp, and I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

I think there is definitely value in the prompt. Still, I don't know that design will kill prompt engineering, but I do think that the marriage between prompts and design are going to really unlock the next like, productive interface. Hello everyone.

Welcome to Data Framed. Hi, I'm Adele, data evangelist and educator at datacamp, and if you're new here, Data Framed is a weekly podcast in which we explore how individuals and organizations can succeed with data and AI. I think it's safe to say that every organization today is trying to figure out what to build with AI. More importantly, they're trying to figure out how to embed AI into their products, systems, processes and workflows. And of course, when looking at the ingredients that determine an organization's ability to ship successful AI products, the usual suspects emerge. Skills, a technical infrastructure, and a good understanding of the limitations and capabilities of AI systems today. However, there is another key ingredient that is often not talked about enough, in my opinion, which is great design. In a lot of ways, the success of ChatGPT is not just due to the underlying model's performance. It's also due to the fact that it's neatly wrapped up in a human centered user experience and interface, which is in this case, the chat interface. So how do you build human centered AI experiences? What is the role of design and driving successful AI implementations? How can data leaders and practitioners adopt a design lens? Here to answer these questions is Haris Butt head of Product Design at ClickUp. In case you haven't seen it, ClickUp is a project management tool that's been recently making a big bet on AI, and Haris plays a key role in shaping how AI is embedded within the platform. Throughout the episode, we spoke about the role of design in driving human centered AI experiences, the iterative process of designing for AI, how to design AI experiences that promote trust, how designing for AI differs from traditional software engineering, whether good design will ultimately end up killing prompt engineering, and a lot more. If you Enjoyed this episode. Make sure to let us know in the comments on social or more. And now onto today's episode, foreign.

It's great to have you on the show.

Thank you. Great to be here.

So you are the head of product design at ClickUp? For those who are not familiar with ClickUp, ClickUp is an excellent project management tool that we even use to manage the data frame podcast, really, for any about any type of project or role. Most recently you unveiled the latest iteration of ClickUp, and as you can imagine if you clicked on this episode, AI is a key dimension of the updates introduced. So maybe to first set the stage for our conversation, har is how do you view the role of AI in transforming how we manage projects?

Yeah, for sure. I think it really kind of boils down to integrating it in a way where it doesn't really feel like a burden on people. It is a technology that we're looking as a way to really bring like process and workflow automation to the masses. Right. So as we've seen it today, you've seen a lot of cool stuff around, like generative images, audio text and all that stuff, but it's really good at summarizing data as well. But the context is the thing that really matters. And so within ClickUp specifically, you have all this context around work, whether it be a project, whether it be a document, knowledge, whatever that might be. And a big chunk of how we see it is how do we use this in a way that makes a workspace feel intelligent, not in a way that makes it feel like there's something else you have to do in order to get the benefit of this technology right. And you see that today across all the offerings that a lot of our competitors and even new companies and point solutions are coming up with. That's roughly how we see it today, where right now it's just like an artifact that is presented to you inside of these new tools. But as we look to it in the future, it really kind of becomes more about how you incorporate these technologies in a way where the product itself becomes more intelligent and intuitive and helps you move faster, work smarter, and all that stuff.

Yeah, I'm very excited to unpack that in a lot more detail with you and think about how design plays a big role in making that technology come to life within these tools. But maybe to further demystify how ClickUp is integrating AI into its project management tool, maybe walk us through what ClickUp AI unlocks and how AI will be driving the next iterations of ClickUp as we look to the future.

So you look at ChatGPT today, I always refer to, like, programming, where anyone can learn a programming language today. But for some reason, it's kind of difficult to do. Talking to computers is tough. AI made that a little bit different. It made talking to computers a whole lot easier. The machine can take human, like, natural language instruction now and do things for you. And so when ChatGPT first came out and we started thinking about how to incorporate this into the product, we kind of looked at ChatGPT, we looked at how other tools are doing it, and the first thing we noticed was, well, there's a huge barrier to entry still. What do you even pay to get a desired output? What can you even use this for? And so first thing that we actually started doing was building custom tools, right? These functions inside of the product that would allow you to do specific things by providing a certain set of instructions. Instead of giving you, like just a little chat box and waiting for you to figure out what exactly to ask it, we created tools for marketing, for sales, for support for product. And what that basically allowed us to do was humanize props in a little bit of a way, right where you go to your mom, you go to any common knowledge worker, and you're like, hey, yeah, come up with a prompt for yourself. And that will basically allow for you to produce a set of inputs and get yourself a magical output that helps you move faster. But it actually created slowness in that coming up with a prompt is difficult to do. So we built out a team and we started putting together a library of proprietary, like, prompts that would help certain functions or roles. And so that was the impetus for our vision, our North Star for ClickUp, where it's like, okay, we have to humanize AI in a way where it speaks your language, and we're continuing to push down that path. So the V1 of ClickUp AI really introduced templated prompts, but we did it in a way where we created, like, an interface that accepted a set of inputs to help you produce outputs more quickly. So really common example would be like meeting notes. So who are you having a meeting with? What are you talking about? What are the desired outcomes? Normally you would open up AI and say, hey, I want a meeting note template that wants to do this. But we went ahead and did that for you by creating an interface for it. And so when you use ClickUp AI now you just click on a button, say, meeting note, say what it's for, and then we use our own prompts to generate that meeting note. Template for you almost instantly. And so that was like V0. Now, as we look to it further, there's a lot of other cool things you can do too, right? You can reference things inside of the product, you can add context related to work, projects or people. So I mean, if you follow that line of thinking further, that's where we're going to land in the coming months as an integral kind of pillar to how ClickUp will work and function in the future. That was kind of just a way for us to start thinking about how this stuff kind of folds into an all in one workspace. So if you're inside of a task, might not make sense to offer up some of these tools in those areas, right? So a big chunk of AI, I think falls onto context, what you might be doing, who you might be doing things with within the scope of work, and intelligently presenting that stuff to you. And so the next evolution of that as we continue to pull on this thread was like, we started to think about this process automation, like where is the burden? Where is the time being wasted for like majority of people? And we looked at like what AI could reliably do given a set of inputs. And the next thing we could start thinking about was kind of like, everyone's going down this path of co pilots and chat companions and things like that. But we once again following this line of like humanizing this technology in a way, we started thinking about it more through the lines of like, who is this person inside of your workspace? And we kind of landed on an AI project manager. And that instantly made it a lot easier for us to think about what would this AI project manager do? The status updates, the standups, the weekly summaries, all of that stuff came into view. And then we got really excited. Then we were like, okay, all right, this is, this is getting fun. And so you, you think about, I'll just like poke around at some of the ideas. But like when you think within the realm of an AI project manager, it wasn't so much like let's replace project managers everywhere, not yet at least. But we were like, what if everyone had their own project manager? I say this a lot, but like, I really do believe that like modern workplace productivity remains unsolved in some ways. Lots of really cool tools, lots of really cool platforms, but we still haven't really achieved a tool. Actually molding to the way you and your team work, which is, I think a north star for our product. Org is like, we have an incredibly flexible product, but we still need to make it like work how you work basically right now, the delta between signing up for a tool and getting it, getting it to that point, it takes a lot. No matter what tool you use, it's a lot of effort. And so with the AI Project manager, it became more about what are the things that we could reasonably do right now that would just help you, kind of give you peace of mind, get you moving faster. So we started with thread summaries, right? So we were like, oh, you're having conversations with your team, you're doing it async. Once it gets past a couple of responses, you're just sitting there staring, poking through, trying to find the action items and whatnot. And so with thread summaries we made it so easy to just like click a button and get high level points right away. Action items presented to you, obviously we're following the data and the moment we roll it out it's like boom, right? Everyone wants like the TLDR I wanted to call it actually, but we didn't get to do that. Doesn't really have like a positive spin to it. But that was like a really cool like aha moment where we were like, wow, this is, it's just a button, but it really does simplify so much of your ability to keep up with work and keep things moving. So then after that we were like, okay, this works really well. What else can we do with the data that's unique to people and their projects and so of ClickUp that other people wouldn't necessarily be able to do? And I think this is where a lot of the project management, management excitement came to light, where we have such a vast amount of data inside of ClickUp that is related to people teams, their work. And our atomic unit of work is the task. So you create a list, you start creating tasks and da da da da da, you got like a or so of them after a couple weeks really comes to becomes difficult to just like keep up with that, right? Like project managers are basically the ones that are staring at it, giving you a weekly update, changing statuses. So we started, we had, we actually have like a really robust like task activity data pipeline. So then we were like, oh, okay, cool. What if you had a time interval and now you can just follow this task instantly. And what if we made it so that we placed emphasis on the tasks that mattered to you most? So then we added task summaries. So there's a part to this that's just fine tuning the function itself. Like Adil goes to recording with Horace today and We've got like a long brief and you, you hit it, but then you think about it through the lens of like process automation. We have like a really powerful automations framework too. So now let's just say anything that's due this week, I want an automated summary sent to me. So if you follow this, I know I'm like putting a bunch of different things out there, but when you look at ClickUp as like an all in one workspace, right, you're beginning to create this feedback loop now. So we've created the function and now we're using automation as a way to feed it into your inbox. And basically once we get this to completeness, right, and we add all the bells and whistles, you can totally imagine a world where you sign up, you tell us how you work, what are the things you want to keep up to speed with, and then the flywheel just gets built and then you just go about doing your work and you don't really have to worry about someone keeping you up to speed. Things you're watching, things you're assigned to, things where you're mentioned, all of that stuff begins to neatly be presented to you. So we've got the function parts done so far. We're doing it on the task, we're doing it, we're doing it at the, the person level. And so the next kind of few things that we did, I mean, I'll. I don't know how, how deep you want me to go in them, but the other things that we started to do then were at the person level, killing the standup. Like, what an enormous waste of time. What did you do yesterday? What did you do today? What are the blockers? Get into a meeting. And so we started to automate that recently. We've got that coming out the door pretty soon, which is really, really cool. And then custom fields, so. So your lists and projects are filled with lots and lots of crazy, crazy updates. Now imagine you spin up a project and you set a time interval and every time you go to your project it's just neatly updated. So that's at a high level where we're headed as a company. I don't know how much deeper you want me to go, but it really is about introducing this intelligent kind of framework into the product in a way that like, just lets you be almost like driving a Tesla. Your hands still have to stay on the wheels. They still have to stay on the wheels. But the car does drive itself and it's aware of all of its surroundings. Eventually, hopefully we get to a point where the hands are completely off. But at that point I'm questioning what the point of work even is.

So there's a lot to unpack there. And there's a couple of things that you mentioned at the beginning that I think I really want to tease out. Right. You mentioned two keywords here that I want to focus on. First one is interface. Right. And second one is focusing on, you know, humanizing AI.

So I'm really eager to explore the.

Value of AI and reinventing project management, really removing the drudgery of day to day work. Right. But I really want to focus with you, given your experience with design, given you the way you're thinking about the integrating AI into the ClickUp experience is the role of design in creating a beautiful user centric, highly human AI experiences. Right. In a lot of ways, right. What made ChatGPT so successful, what made it explode was that simple chat interface that really democratized a sophisticated large language model in the hands of the masses. Right. Like even my parents asked me about ChatGPT and they've used ChatGPT. So maybe to set the stage. How would you first define a user centric experience in the context of interacting with a large language model?

You know, I spent a lot of time thinking about this, but like, I really think that when it comes down to it, there probably isn't anything better than natural language to leverage technology. Right. And what's really cool about when you think about AI through a design lens, right? It's like, I think about like how you meet a stranger. Almost like AI in itself, in and of itself is like a black box. You really only get where you're able to ask it. Same as when you meet a stranger. You're like, it's tough to come up with questions and like, you know, carry the conversation a little bit. You can have a list of topics to talk about when you, when you have it, but in the moment. Yeah, list of prompts. Exactly. It's kind of difficult. You, you, you don't really, you don't really know what you can get out of a conversation, right? Both in a person to person perspective, but also like person to technology. You have to be really good at asking questions, right? And so if you follow that, you look at Bing Chat, they've incorporated it into the browser. A lot of project management tools have put like a little chat window there. I can't help but feel like most of it is going to come from designing the other side of the conversation. And what I mean by that is like, right Now I have to come up to you and say, hey, I'd like a template for something, or I would like to do this. But when I think about that interaction, I actually want to flip it in a way where AI is asking you what you'd like to do. It's presenting options to you based on what you've done in the past within the context of things. And right now, it's kind of. It's scoped to your specific applications and platforms and such. But you can totally imagine a world where this is just on your machine, right? It's on your laptop. It's totally local to your entire computer environment. And when you think about that realm of design, what do we have today? We've had Siri, that did great. And I just. I just accidentally activated on my phone. We have. We have Alexa, we have Google, we have all these things, but they just listen. You're never in your house, and the home assistant's like, hey, it's 90 degrees outside right now. You want to turn on your air conditioning, or you left your car unlocked. Did you want me to lock it for you? Small things like that, which give you the confidence and knowledge to understand what these things are able to do for you, right? So then your way out the door, when it's hot outside, you know that it can do this. You turn it on right now, you're like, hmm, I wonder what this thing can do. You ask a question and it's like, sorry, you have to go to do something else. And so within the realm of design, specifically, I really think about that experience a lot, where today we just add buttons to everything. It's like AI this, AI that. And you see this across all the projects. AI comments, AI. And I think of it as. It truly is experience design in the sense that when you're somewhere inside of a product or when you're doing something, it's that proactive nature of the technology that can sense what you're doing and bring your attention to something that may be of value to you. And of course, from there it comes to customization and molding it specifically to how you work. But that's how I think about it through the lens of design, is that like, the natural language interaction part is probably really good thing. As you mentioned, it opened it up, it humanized it. It really did. It felt like texting your friend. And the really great thing about it, I think, was that it created more of a feedback loop between person in computer. Whereas normally in programming, you write a function, you test that function like Ash, like not something work, something didn't work, something broke. And then you rack your head up thinking about how you fix this function. But now you're just a continuous eval loop where it's just like, hey, can you give me the weather? Where do you want the weather? Like, here's the weather based on what I did. Oh, no, I meant weather in London. Oh, okay, cool. And you just keep on like, probing at it until you get the desired output. And I get really excited by the idea of an intelligent workspace in the sense that, well, AI just doesn't. You don't really need to talk about AI. You want to move in a direction where the interface of an application just evolves in a way that it accepts these inputs and presents these options to you in an intuitive way to help you get things done. So you're setting up a project for the first time. You should be able to instruct AI. Not in the sense that AI comes up as a chatbot and says, hey, what would you like to do? Kind of thing. But you're kicking off a project and you can present a set of inputs and have that be generated, and then from there some kind of experience that allows you to tailor it further. And the real bulk of the benefit there is from like the fine tuning, basically. Right. If you want to run marketing briefs, if you, you want to run your engineering team, it can sense that. I don't know how far we are from that. I'm still kind of like exploring, but I can't help but feel like most interfaces will need some kind of overhaul or evolution that kind of accommodates this. Because the back and forth between an application and chat, I think, is less than ideal right now. It's not integrated in a way that actually helps you be that 10x human being.

Yeah. So there's this idea that I think we're circling around at the moment, which is, I think over since the release of ChatGPT, there's this concept of prompt engineering, Right. That has been quietly, quite, quite publicized. A lot of people discuss, is this the next big job? Right. I'm of the opinion that I think design will kill prompt engineering at one point in time.

And when I mean that, I totally agree.

And what I mean by that is like a great interface. You know, you mentioned here, like ClickUp a press a button, really kind of. AI actually is proactive about presenting to you what type of capabilities it is. Right. Think a marketing specific task. Right. There's a slider here somewhere, right. That makes it specific to a marketing Task, Right? So where do you see that skill set of prompt engineering evolving? And do you think design, and how do you think design will kill prompt engineering in the future?

Oh, man. I actually love this question a lot. I think there is definitely value in the prompt still. I don't know that design will kill prompt engineering, but I do think that the marriage between prompts and design are going to really unlock the next, like, productive interface for human beings. Like, when we first worked on the AI tools it was born out of. Honestly, I was showing chatgpt to my dad and yeah, I was like, hey, check this out, right? So I'm like generating snippets of code, I'm generating QA checklists for features, and he's just like, what the heck is. Like, what, like, what are you talking to? What is doing this thing? It's totally confusing. And I kept on thinking about that, I was like, most people aren't techies. When I used to work at the Apple Store, right? Like, people would be like, how do I set up my. You're buying a $3,000 machine and you're like, how does email work? And I always think about that through the lens of, like, designing products. And so I was like, I don't want to put a chat screen into ClickUp as, as first thing for AI. I, I, I, I don't think it, it, it just like, sets off a light bulb for you.

Yeah, it's not a problem.

That was like the, that was like when I first started thinking about, like, what is the interface on? And all you heard about was prompts, prompts this, prompts that. Every influencer, every person had their own prompts that they wanted to get out. How do you pick, like, which one's the right one, which one's good and which one's bad? There's just so much to it. So I wanted to simplify that. And I was like, what if it was just a simple form? So we hire a prompt engineer, we battle test them ourselves, we make a list of things that we would use AI for and we templatize it. We kind of leave the whole prompt terminology behind and we just build a tool. There was a lot more to it that I wanted to do that we haven't quite gotten to yet. But when I shared that with my dad, it's a normal form. It was like, you want to put together an sop, there's an SOP button right there. Select it. And I think when you think about how prompts are going to evolve, I think it's going to mostly be the presentation of the prompt. I don't know that the form was the right thing, right? But it's actually accepting a series of inputs, putting it through a series of like, conditions and rules, and then get. Getting you to an output. A really cool example of that. I think there's a. What was the app called? It was. There's a really cool app I tried out this week. It's called like Gamma app. Have you heard of it?

No, I haven't heard of it.

It's like a pre. It's like a presentation. It's like a slide deck software. I was working on a. I was working on like a pretty long document and I was like, how do I. I was like, now I have to go turn this into a deck so I can share this with the team. And I remember that I had seen this app somewhere, but it was so awesome. I just took this doc, it was formatted perfectly and I dumped it in and it accepted like, it asked, how. How do you want to present this? And I said, well, I want to keep the formatting of the document and I want to share this with my design team and I don't want it to be more than 10 slides. And it took the contents and it just perfectly, like, literally perfectly spit out the ideal slide deck. And not only did it do that, there are parts of it that kind of COVID like diagrams and visuals. It went and mapped visuals to the sections and laid them out inside. I was shook. I was just sitting there being like, now imagine if we brought this kind of experience to ClickUp, where you have all these knowledge bases and documents. You want to present your new vacation policy. Imagine that there is like a tool for turning something into a presentation with a click of a button. I think there's a part to it where the. There's a base layer of prompt that achieves a specific output. But I think there's another part to it that is like. It like, adds to the prompt in a way that is unique to you. So right now that's like re. Prompting, right? But it's. It doesn't change the underlying prompt. So I could ask it to do something and then I say, hey, no, I want to refine it further, do this, do that. But then you go back and like, all those changes that you made don't converge, they don't compile. And so I think when it comes to how design is going to influence that, it's really going to be like, how do you create that loop? Right? Is it learning? Is it is the AI learning, I'm not entirely sure, but giving that feedback to you as you wanted to put together a brief to talk to me today, and you're like, no, those are not the right questions. This is not the right tone of voice. Are you going to do that every single time you want to talk to someone? Or is it learning and giving you feedback, asking you, hey, do you want to modify the questions you're asking so that the next time you do this, it's easier to do? That's where I think design is really going to overtake a lot of this prompt stuff. Like, it's not. It obviously matters, like, what's underneath the hood. But I think creating the feedback loop between, like, man and machine is really going to be the thing that makes AI actually useful instead of, like, it's kind of helpful.

Yeah, I couldn't agree more here. And I think we can dedicate an entire episode just to discuss this particular topic. Right. But you mentioned the earlier thinking around the ClickUp AI and how you're thinking about developing the interface for that intelligence layer and the ClickUp app, maybe. ChatGPT was released late 2022. Right. Walk me through that journey of iteration and that kind of the different ways that intelligence layer changed in ClickUp and how your thinking evolved up until release today.

So obviously, I mean, the first part, like, with these tools, it was like, you get the tools in, okay, you give it a series of inputs, produces an output that was V0, we added re prompting after that, produces the output, modified the output. And as we kind of, like observed, like, the tools were helpful. Like, obviously there's like the generative part, which is like, hey, write me a blog post. Do that. We baked it into docs and we baked it into our task creation experience. And we kind of let it sit there. We were like, okay, let's like, see where that takes people. It didn't catch, like, wildfire obviously enough. And we were like, why? Why is that? And that was like the light bulb for me personally, where I was just like, even if it's there, people don't necessarily reach for it. You know what I mean? Like, you're. It's very tough to change behavior. And so the chat interface and us evolving that into more of like a tool slash form was cool. I still think we have a lot of improvement to do there. But then it became more about, okay, this actually has to exist more functionally inside of the product. It can't just be like, who talking to almost like a person inside the product. That's where we started moving into like, okay, this needs to obviously be baked into the product experience as a whole. Now we could definitely blow the entire application up and go like, all on AI and everything's all intelligent, but obviously we can't do that. We have to, like, we have to be thoughtful about how to kind of weave this into the product experience where it's just like a new affordance that shows up and you're like, what does this thing do? And that function, that feature adds value in a way that it creates supply wheel and that conversation with AI. And so that's kind of where, as I was mentioning earlier, we started doing like killing the standup, getting automatic task updates, getting automatic thread summaries and things like that. Where now instead of like, hey, do you want to write a blog post or do you want meeting notes and things like that, like the tables turned a little bit where it was like, hey, I'm offering up this information to you about the thing that you're looking at. And that still kind of remains the goal today. We're not quite there yet. We're still in prototyping mode in the sense that there is obviously AI, in the sense that everyone's using it today, which is you're chatting back and forth, but thoughtfully creating moments in the application. I like standup a lot. Like, standup is one of those things that we all racked our brains on for a long time where we were like, damn, every single person on every single team does this thing every single time. And then you have a person employed to hunt those things down and make sure they get updated. And then you have a manager that's responsible for reviewing and all this stuff. It's a lot of time, you know, larger and larger an organization grows. And so touching back on the idea of like an AI project manager inside your workspace, the central theme for this is trust. In our case, if you have a project manager, you need to be able to trust this project manager, right? Like, that is the underlying crux to AI. And that kind of makes it really important to be objective about the information being presented to you. And so based on where you are or what you're doing, having the ability to get that information and being able to trust it and then continue to probe or get more information is where the bulk of our attention went to. And we're still going down this path. As I mentioned, we. We want to get to a place where this is like, virtually automated, but from like a design evolution. It was like, put AI everywhere. Cause everyone's talking AI, like, let's just go crazy with like, AI this, AI that. And my thinking has completely gone 180 on this in the sense that, like, I don't want, I almost don't want to talk about AI. I don't want to say that this is AI. I want it to feel. I keep referring to that, but it's like I want it to feel like driving a Tesla, I really do. And that like all, like, slowly over time, you sit in a normal car versus a Tesla, like all the belt, like all the knobs and all these things kind of like disappeared. And I know a lot of people say that's not a good thing, that bad UX and all these things, but I think the intention behind it is get to a point where, like, you are hands off, right? You're not thinking about these things anymore. And it's creating a new foundation for experience. And so a lot of, I think what we did in terms of like seeing ChatGPT come, right, and then they, like, ChatGPT allows you to create different threads. Most of the AI apps allowed you to create different threads. Next you're probably going to see people wanting to organize them or save them or share them. Like all these things naturally kind of emerge. It's the same for ClickUp. It's like all these new AI needs are going to emerge and there's two ways you can look about, look at it. And I'm sure every single, every single company is probably thinking about this in a way that, well, there's. There's one world where there's a parallel AI experience inside the product and there is a separate world which is AI is just perfect. AI is the product, it's just the next version of the product. And things will slowly begin to, like, be simplified, organized in a way where it's designed to take an input and produce the output for you. It's designed to learn what your next step is. So before you can even think about it, it's presented to you. I don't think we're quite there yet. I think we have quite a bit of work to do. I think a large chunk of it kind of boils down to us, like, what parts of what we've designed so far are we going to have to undo to accommodate this? And that's the design challenge today.

Yeah, and there's one thing that you mentioned here as well throughout your, Your answer was that was very resonant, was just trust, right? Like, how are you going to trust this AI project manager and I think there's a big risk. Of course, this is quite widely discussed, right? Like the risk of hallucinations, of bias of fabricated output from a particular model when you are, let's say you're in the ClickUp app and you get a summary of thread or a stand up note or like with all of the action items, right? What is the role of design in maintaining trust and ensuring and minimizing these potential risks of a hallucinated action item? Or even worse, like a forgotten action item that didn't make the cut in the summary, for example?

Yeah, for sure. There's actually a really good example for this. So we actually have one of the only universal search products in the productivity space. And so we were like, okay, we have access to all this data, we can do this ask AI experience, right? So you open up the spotlight and you're like, hey, what is our office PTO policy? And so the V0 of it was just like. And this is, this was even for summarizing docs and things like that. We were like, oh, if we have this knowledge based product, you can just ask your workspace anything. And based on the data we have, we're going to give you an answer and you load everything up into the vector DB and you index it all and you're good to go. And the V0 of it, I think I actually asked like, what is our vacation policy? And it gave the most bizarre answer. It was ridiculous. It was ridiculous because people use docs for all kinds of, like every team is using. Like, it's not scoped to me, it's not scoped to the workspace. There's all kinds of random things inside. And so I asked something, I swear, I think it was like some kind of quote from Star Trek or something like that. And I was like, okay, so this is not helpful. So from there it was kind of like, how do we build an interface that doesn't make you freak out? Like, is this completely artificial? Does this exist in a doc somewhere? So a really simple example of that was like we just like incorporated sources and building out the functionality to specify sources so you know that the experience is contained to a set, like a defined set of knowledge or information. And I think for ClickUp it's a lot easier because we'll definitely move down this path where the information is scoped and relevant to you. We'll build out the stack in a way where if like Horace has a meeting notes doc and he says this is, this is a doc that is I want in my ClickUp Brain, and there is our design team handbook, and this should be in the ClickUp brain. And there's like our product development framework. So, like, I will move in the direction where you specify the knowledge, right? So you know the knowledge exists, it has value to you. And when you ask for information as you need it, it's presented to you in a way where it says, here is the information and here's where it's coming from. So you know its sources, its citations and things like that. And I think if you're familiar with, like, Perplexity, they've done a really good job at this. And I think I took a lot of inspiration from that. And you have Bard and you have chatgpt, but Perplexity, I think, took it one step further in that I don't know that they were focusing on trust, but it makes me trust it more because that context is there for me. And I think about it in a very similar sense in that that's just knowledge. But you also, within a collaborative workspace, want to know who said what. So if I summarize a project, it's not like, hey, this is what's going on, right? Then I'm like, wait a second, like, who made that decision and why? Like, why is this being said? That whole, like, citation and context, we still need to add that in. There are still times when I'm like, I'll do a task summary. I'll do a project summary inside of ClickUp. And I'm like, oh, sweet. But then there's still those moments where I'm just like, I know this isn't a hallucination, but I don't know exactly where this is coming from. Then it's not useful anymore, right? Because then I have to go right back in. I have to comb through like a dozen or so comments and like, like, oh, okay, there's my peace of mind. It's not just, like, fabricated, it's actually just not included. I would say that's like, probably as far as I've thought about trust so far. I'm sure there are other things around, like letting you know this certain information is secure. Like, I don't know, maybe you had to submit some kind of form where it's asking for your Social Security number and your private phone number. You can just submit it and then it does something with it. But I would personally appreciate some kind of like, hey, this information isn't being stored anywhere. It's not being shared with anyone. There's a lot of that, I think that still needs to be Incorporated on a larger scale. Right. Like so much proprietary information that companies have that prevents them from using AI technologies and such, simply because of the fact that you're training these models on sensitive data. And so we have to build that trust, letting people know that their, their, their workspace data is safe. With ClickUp, we have to create that feedback loop in some way. And I think just starting with the sources is probably a good place.

So in a lot of ways, what you're mentioning here around hallucination and the risk of hallucination is very unique to large language models and to that particular type of AI. I think traditionally design teams and engineering teams didn't have to grapple with these types of issues. Where the set of interactions exists in an infinite space where you don't necessarily know what is the output a large language model can have, whereas traditional software exists in a finite space. One plus one equals too, right? You know for certain what is the output you get given a certain type of input. So maybe given that. Walk us through the unique challenges of designing for large, large language models in AI, unlike traditional software or traditional design in this particular circumstance, for sure, I.

Think it really just boils down to context. Like just to think about it in the most simple of terms, I just think about it in terms of context. There's like billions of permutations and parameters. Maybe trillions, I don't actually know, but as you mentioned, it's a black box in the sense that like the output is near infinite at this point, right? Where it's just like you can put something in, something's gonna come out at some point, or that's just how it works. And so I think the difficult thing is like thinking about using AI within a, a certain context and making that like, intuitive meaning that like, if you're interfacing with a doc, you're interfacing with a task, a person, a project, the whole workspace as a whole. The interfaces we have today don't really make that obvious. It's just a blob of text that gets presented to you. Sometimes it's structured, sometimes it has citations. It's not obvious to people what part of the black box you're talking to. Right? It's not obvious like how this kind of information comes together inside the product in a way that helps you trust it. So I'm personally not probably qualified from a technical sense to speak on it, but I think just from a product experience perspective, it does come down to trust and context to me in the sense that it has to feel like second nature to you in the sense that when you are interfaced, like I use Siri as a good example for this. Like, you just don't know what you're going to get out of it. Right. One day it might answer your question, the other day it might tell you to open up your phone and go to the browser and do it. Do it yourself. With large language models specifically, I think specifying the information within the context of what you're trying to do is a challenging things to think to do. Because I think just as easily as you can have information presented to you, it can just as become a daunting task to go figure out how and why this particular output was produced for you. Right. And I think going back to the prompt engineering bit, this is where it gets really fussy because you have chat GPT 3, 5, you have anthropic Claude, you have. What is it? Chat GPT, what is it? 4, 0 or whatever. Like these model continue.

Yeah.

And there's no, there's Pro. Like when we first did our AI tools, we built it on top of 3.5 and that, that accepted a specific set of instructions in a certain kind of way that were optimized for what we were asking it. Right. And then it was like, oh, okay, let's switch it to 4.0. Tons more parameters, tons more creativity. It'll like, it unlocks a whole lot more. Might be a little bit slower, but it'll be more powerful. But what we actually found was the model doesn't work the same. I don't know why, but it doesn't work the same way. And now you got to go back and like all 200 tools that we wrote, now you gotta, you gotta create that feedback loop again. You don't know the diff. And you think about that further. This is just, you know, we're, we're at the bottom of the beginning. Yeah. This hasn't even started yet.

Right.

Imagine when like multimedia and all these things begin to be incorporated into the models. Prompt engineering is going to be like Amazon Turk. Like, you're just going to need to keep going and going and going until you get like, keeping up with this is going to be crazy. And so my mind wanders in that area where it's like, are we going to build our own model? Like, is building model going to like. It's very costly right now. Even the hardware required to produce it or create it is costly. I don't know what the answer is there yet. I think this is just, this is just going to be the nature of the game for a little bit where the space is going to evolve so fast. And I think this is where design becomes really powerful actually, where if we let the evolution of the models be what it needs to be, we focus on fine tuning the experience. So regardless of what the information is that's going to be passed through, we just think about exactly what the interaction between you and the model is going to look and feel like, how we're going to bake it into the product experience. So even if you change what's underneath the hood, maybe the outputs we can continue to fine tune, but focus on the experience so that at least the data is available to you, the information, the usefulness is available to you as you need it, and then let the large language model wave go and see like what it allows us to do.

That's really great. And in a lot of ways we're currently in the wild west of AI.

Oh, 100%.

100% the, the ecosystem and how it's evolving. And as this world wild west that we're discussing evolves. Right. I'd be remiss not to talk to you about where you see the future of productivity heading. Right. Human models get stronger, better, more resilient. Where do you think workspace productivity will be in a few years?

Oh, geez, I wish I had the answer to this one. It's really tough to say, to be honest. There's a part like obviously there's like the doomsday approach, which is like productivity is no more like you don't need to go to start a farm or pick up a hobby or something like that. But I, I kind of see productivity across the timelines as like a series of like expansion and contraction. Right. Like there was a point where we were punching in punch cards to talk to computers, right. To just run a sum function. Knowledge workers were writing on like sheets of paper and that was compiled into notebooks and presented on someone's desk. I feel like it's just going to allow people to do. It's just going to unlock a new generation. It's going to just redefine productivity. I don't know what that looks like yet, but I can talk through my own experience in that. Like I run both product and design and it's like left brain, right brain a little bit. It's takes a lot of, it takes a lot of my time. Right. Outside of like hiring an assistant or bringing on an additional project manager or moving waste more slowly, there aren't really any tools that make my day to day really easy. It's just a lot of mundane repetition process. I very like, I, I'm working on the tools that are going to take that all off my plate, right? And then I think about what is that going to allow me to do? And it's going to unleash my creativity, it's going to give me my brain back. Like it helps me be strategic, it helps me be creative, it helps me start thinking ahead more and there isn't really an outcome tied to that. But you can see that if I get my time back, which is ultimately our North Star, right? Like we're all about saving people, teams, companies, time. When we actually achieve giving like actual time back, you can imagine people are going to start spending it in other areas. They'll have new tools where these kinds of things are more generative. Now instead of, hey, I need to spend a whole day writing a project plan, it's written for me. I get to focus on how I'm going to communicate this. And now there's a tool that communicates it for me too, right? Like it, it produces that and it's like you just start moving faster, right? You, you see, see this like across the evolution of like agriculture, you had humans picking every single, you had humans picking every single fruit and vegetable. Now you have robots growing or roaming across like crops just picking things off. This is ultimately what industrialization does, what automation does. And so yeah, I think a lot of mundane jobs and such will be eliminated. But at the same time I think it's just going, you're going to have this new set of tools that you can weave together in a way to, to do incredible things. And hard to say what that looks like right now, right? You see, you see glim glimpses of it with like, I don't know if you've seen like Runway ML where you can like do completely generative videos. There's like Dolly and stable diffusion where like artistry is. There's just like a new class of artistry evolving. You see people making music with other artists voices and merging it. It's. You said it perfectly. It really is the wild wild west in some ways. I liked it. I, I mean there's a he, he passed away recently. But like I think it was like Virgil Ablo or something like that that said like everything like a remix of something else. I think we have like the most powerful remixing tool ever now, right? Like just like literally God mode remixing. What you want to mix a photo with some music, like remix it, you know, and it's cool because you just have like the easy button. Like, try again, try again, try again. And hard to say where that takes you. Right. Like, is all of our music just going to be completely generated at some point? Is all art just. Yeah. It's crazy to think about, but I think within the realm of productivity, specifically, I think it's just going to be redefined in the sense that, like, it's obvious that these things are going to mold to you and how you work. And I think work is going to get a whole lot more unique. I think it's going to get a whole lot more efficient. I think it's going to get a whole lot more like. What's the right way to put it? Like, tailored to you. I worked at a company called Detour back in the day and they, they actually build Descript now. I don't know if you're familiar with that.

We use it on the podcast.

Oh, you do? Okay, cool. Yeah. So descript was originally like an editor. We used to build Detours. But it's, it's so funny to see. I, I left the company before it became Descript. But it's so cool to see when we were solving that problem of editing audio tours with text, it was less. It was like a very, like, clunky tool. But how with AI now, it like records your voice, right? And it's like a full daw and you're just editing your voice with text. And as crazy as that sounds, right. Like, never would I have thought that that's how your job gets easier. Like, it has transformed your ability to edit scripts, put things together and just shoot it out. Whereas otherwise you're in like, logic, taking snippets of audio, shuffling it around. It's truly magical. So I can't even imagine what that's going to do for an AI project manager or for AI standups. Right click of a button, you have this information to you, automatic prioritization. It's like, I think at a high level, the most important things and most relevant things to you are always going to be front and center. Right now there's still a lot of ping ponging back and forth between apps and projects and people and things like that. But when you, when you have this ability to take very, very large chunks of data and scope it to exactly what you need it to be at any given point in time, I think it just like unlocks so much.

Yeah. And I would underestimate as well, when it comes to the realm of productivity, really the potential upside when it comes to our well being at work. When you take away a lot of the drudgery that.

Oh yeah, for sure.

Much of our time. Right. Like, I don't think anyone.

You go on vacation and you're, you. What is it, the Sunday scaries or something like that? Yeah, you're. You're sitting there being like, oh my God, I just took a week off for vacation.

I have a mountain of work coming back.

A mountain of work. Like how, how. That's actually a really good point. I forgot about that one. But yeah, we were actually just thinking about that the other day. But like, inbox, your notifications like fill up a lot really quickly. But like about a vacation mode where it's like, what if you just press a button and it's like, hey, actually these are the only things that matter right now.

Forget about that. Exactly.

You know, exactly. But yeah, that's a great example.

Yeah. So, Haris, as we close out our conversation, do you have any final call to action before we wrap up today's episode? I'd love to hear your final thoughts on the role of design and shaping the future of AI.

I don't know if I have any final comments. I have a lot of threads to pull on, but my mind like literally wanders. It truly is. I think it's such a unique opportunity for designers out there to, to humanize the problems we see every day in a way that is like completely abstract and uncoupled from an actual interface. I think this is like one of the first times I've ever felt this way in that you have a set of problems and you try and cobble up some kind of interface or product experience based on some components and a design library. But this is the first time I think you can cut all of that out as a designer and think about the experience as a whole and like an end to end sequence of interaction that doesn't really require an interface. You can add. You can add an interface if you want. You can get really clever and novel in some kinds of ways, but you really have to think about the individual, you really have to think about the context, you really have to think about the audience. And I just think we've barely scratched the surface here. We're still stuck in this land of like chat and buttons and all this other crap. But I think for designers specifically, I think it's just such a powerful opportunity to just start everything anew. I've been doing this for almost like 15 years at this point. I can't think of a single time I've felt this way before. And that in and of itself is enough to say that this is, like, an incredibly unique point in time where we just have a new foundation to build on. It's quite possible. It's. It's like, totally okay to say that maybe there's a point where there is no application anymore. I think about that all the time. All the time. I'm just like. It's possible that I just hit the delete button on our code base, and he just said, we just have a model, and you tell that model what work is for you, what you need to do. And this day, like, you. You just have a thing, but it's just like an abstract thing in my mind at the moment, where it's like ironcanny or, like, born identity kind of thing. That's what I would have to say is that, like, I think for designers specifically, it's. If you're not thinking about it in this way, I think you're. You're selling it short.

Yeah. In a lot of ways, like, a parallel is the introduction of a new medium in media. Right.

Absolutely.

You know, printing press came along, changed foundationally how information is disseminated. Video came along, changed foundationally how information is in a lot of ways now. AI is changing foundationally how information is being processed and how you interact with a computer. Harris, it's great having you on the show. Really appreciate your insight.

Thank you. Appreciate it.