How a VC and tech founder used AI to launch a brick-and-mortar business in their spare time | Andrew Mason (CEO of Descript) & Nabeel Hyatt (Partner at Spark Capital)

21:42Andrew Mason:There are always still lawyers involved.

21:44Nabeel Hyatt:At some point, you know, there's real estate agents, there's lawyers. But what was really helpful, obviously, is you now hear all these anecdotes of people who, for instance, are still going to a doctor but are taking their health diagnosis and dropping it into LLMs to get a second opinion. I would say the cycle felt a lot like that. I think it was about building a level of confidence that felt like you could lean into the experience because you weren't going to get fleeced. and so we still relied on experts all the way through we still but but i would just sit at night and say like i don't even know we're trying to describe a concept for an experience of this membership club with retail in the front and we're going to do special events in the back and what is the permit for this in berkeley that won't take like two years for me to get done and uh and that's where you get a little bit of confidence as you're navigating through things Andrew would come back and he'd be looking through Berkeley weird like local ordinance websites and copying the pages and dropping them and clawed and getting feedback on them and things like that.

22:45Nabeel Hyatt:And so it really wasn't it wasn't, again, like like I was doing all the work. It was just that I in that instance, I would say I was giving us the confidence, whereas maybe five years ago would you've been like like you would have had the I would have had this fake expression you had, which is just like this seems scary. it seems Byzantine, like the inertia can't even get started. And so I'm just going to go do the other thing. And this will still just be a thing I complain about. And when I fast forward that, it's just the thing that makes me excited is not just that this exists. It's just like, literally, this wouldn't have existed without AI.

23:19Nabeel Hyatt:Two, as Andrew mentioned, there's so much grunt work in getting something like this open that, again, it couldn't have happened as a side gig. It just couldn't have. It would have had to have been a full-time job. And then there'll be this new, interesting, hopefully break-even place that people actually like in the local community. And if you fast forward that, like, obviously, we are in AI and we have some coding backgrounds and we're very highly technical. We're dropping into things like Cursor and Windsurf, which is not what the average local business person is doing today. But if you fast forward three or four years, you can imagine the multiplicative effect as these tools get broader and broader and broader.

23:55Andrew Mason:And for communities, that means you get interesting businesses and places to go and new people to meet and ways to connect that could not otherwise exist. So I'm very much an optimist that the more we can build, the more amazing things we're all going to get to experience. This episode is brought to you by Persona, the B2B identity platform helping product, fraud, and trust and safety teams protect what they're building in an AI-first world. In 2024, bot traffic officially surpassed human activity online. And with AI agents projected to drive nearly 90 % of all traffic by the end of the decade, it's clear that most of the internet won't be human for much longer.

24:37Andrew Mason:That's why trust and safety matters more than ever. Whether you're building a next-gen AI product or launching a new digital platform, Persona helps ensure it's real humans, not bots or bad actors, accessing your tools. With Persona's building blocks, you can verify users, fight fraud, and meet compliance requirements, all through identity flows tailored to your product and risk needs. You may have already seen Persona in action if you verified your LinkedIn profile or signed up for an Etsy account. It powers identity for the internet's most trusted platforms, and now it can power yours too. Visit withpersona.com slash howiai to learn more.

25:18Andrew Mason:You know, you mentioned Cursor. I just can't get out of a How I AI podcast without somebody talking about Vibe Coding. So Andrew, I'm going to kick it to you to talk about how you actually built an AI concierge for this gaming place. And what is it called? I forget. We didn't mention it. Tabletop Library. Tabletop Library. In Berkeley. Tabletop Library in Berkeley. We forgot to promote it at the top of the show. But how did you build this AI concierge so people could actually access this and have a great experience? Yeah. So one of the ideas that I was quite excited about with this was something that could help organize on-demand game sessions with other cool people that took away a lot of the social weirdness of setting these things up and the social risk in the way that nerds in Silicon Valley are prone to worry about.

26:22Andrew Mason:and so the the mechanism was going to be that you would have you would have this way of saying i want to play a game at some specific time see if you can find a group of people form a quorum and then if you can form that quorum you can reveal us to each other and we'll book it otherwise nothing actually happens and we started out by thinking at the very beginning like okay we'll just find a SaaS software package that gets as close to this as we possibly can. Let's look at table management stuff or other brick and mortar reservation systems. Then we started looking at co-working space management systems.

27:07Andrew Mason:So there was the just standard book a table, straightforward with your friends. And so the idea was that you would be able to use this software to either just book a table with your friends or try to do what we're calling a lfg looking for gamers mechanism where you could say like go out and find this group for me and then book a table and we were playing with all this sas software and it's like it often is where nothing's quite right like you want something so much more custom and then at some point we realized like all we're talking about building here is a relatively simple set of tables, a relatively simple database.

27:49Andrew Mason:And then you could just have like a chatbot living on top of that. Why don't we just build it that way? It wasn't, getting back to that mindset thing, it wasn't at all obvious to us when we started this. We spent a month going down this. In retrospect, I'll show it to you. It seems like a no-brainer, but it wasn't obvious to us at the time. So I'm going to give you a quick demo of what we built, and then I'm going to show you a little bit of how it works, because the one thing about it is we're still building it, and it's kind of slow. So let's say I want to organize a game this weekend, and I can go like, hey, can you set up a game for me any time this weekend?

28:30Andrew Mason:Maybe a deck-building game. So now that goes off to the agent. And I'll show you how we built all that now because it takes a second. And then we'll come back and see its actual reply. There's two parts to this system. One is the database itself, which we built in Airtable. And the other is this workflow from a company called N8N. And let me start by walking you through the basic things of the database. So there's a table where we have all of our users. What's really interesting here is things like availability. So it's going to eventually go and try to recruit other members. Normally in a database, you would store that as a bunch of related tables.

29:16Andrew Mason:But here you can just have free text. This is where you put your gaming preferences in. So it's just like a free text list of the games that the agent will update as stuff goes. So structuring these databases becomes so much simpler. Then we have like the tables that are available in the shop. We have games that we stock. And then we have actually the reservations once they get made. I would like to pause here and just call out you're using Airtable instead of another kind of, I'm sure there's lots of databases that you could have done on the back end if you wanted to. Could you just tell us why Airtable and why you chose to do something like this for this particular project?

30:03Andrew Mason:Yeah. So what do you think should have been in the consideration set before I answer that? Oh, I was just wondering. I mean, I would have just spun up a Postgres database. And if you had a cursor, I would have just gotten straight to just an empty database and a schema. And that's where my developer mind went. So I'm curious why you started with a tool like Airtable? I would not have gone there. I would have been more like, should we be using Notion or some other kind of something with a heavy GUI? Because the nice thing about something like this is we can also build human interfaces for interacting with this directly.

30:41Andrew Mason:So if we have somebody that wants to call up and book a table or something like that, we can build an interface for somebody to just go directly into it or manage events or something. So it just makes it much easier. And it means that people who are going to be, you don't have to be an engineer or build a front end for every command that you want to do.

31:01Nabeel Hyatt:It's a good prompting though, Claire, because this is, again, this was probably four months ago. It's very possible that four months from now, the right answer is you can one-shot a web interface to all of this. And then there's no, that it looks exactly the way you want it to look like. Like in this world, Airtable, one, I don't have that confidence in Windsurf, certainly not four months ago, as we're getting to Cloud4 and beyond, maybe so. And second of all, this just has all of the sorting mechanisms and filtering mechanisms and a million other things that come along with the human use of typing in 100 games or doing other work in those workflows.

31:34Nabeel Hyatt:It's just helpful.

31:34Andrew Mason:Well, and one call out for folks out there thinking about building something like this is in past lives, I would have thought, oh, my God, I don't want to work with the Airtable API. I don't want it. But so many of these things come pre-integrated now with these workflow builders. So much of what you use with these AI tools can figure out the integration points for you. So you can pick the thing that's easiest for you to use. And when you need to integrate it, I think it's very accessible. So it makes sense to me for a business like this. I was just curious if that was yours.

32:05Nabeel Hyatt:I mean, a ridiculous amount of these cells, to be clear, came from that cloud project we talked about earlier. and then were MCP connected into Airtable and it was just injecting data directly in.

32:15Andrew Mason:Can you talk about that flow really quickly? So you told us how you sort of set up these tables for how many physical tables you have, what games you have, who is interested in what kinds of games. How did you actually populate this data through MCP? I'm really curious about that. Yeah, let me give you an example of that one. So like this is, we have a small retail section in the front and we had this idea of like, what if instead of just putting things in alphabetical order or by like a traditional category, we merchandised and curated these categories. You can see some of them here, like brain burners, silent strategists.

32:53Nabeel Hyatt:Almost like the way your favorite independent bookstore merchandises their books in the front, right?

32:59Andrew Mason:Yeah. But there's this puzzle of how do you pick the games and how do you make them mutually exclusive so that you're not merchandising the same game in multiple categories. And that is the kind of thing that probably people have the idea for all the time when they're opening something, but it just ends up on the cutting room floor because of the amount of time it would take to do it. Or you do it at the beginning and then you're like, it's just not worth it. And you cut it and it just makes it trivially easier to do it. So it's like games about cats, for example, you can do something like that.

33:35Andrew Mason:You just give a brief description of what your category is, and then it'll go off and come up with the games automatically. So it's nice in Airtable that they have these built-in AI fields and stuff like that. Got it. So you're using built-in AI categorization by Airtable to create a category and pull in relevant that looks like records or games that might fit. Right. So that's the way that we use the existing tools. Nabil, what were some of the things that you populated with MCP? Is there a tab I should switch to here?

34:10Nabeel Hyatt:Well, probably the easiest one and the most wonderfully nerdy is if you hit games on that tab and then go click to in library. So one of the things we wanted to do, given this is called tabletop library, is to think about this place. If you're starting your board game navigation, and maybe if you've ever been inside of a board game cafe, you've had this feeling, you walk in and you've maybe played Catan and maybe your friends had you play Wingspan and now there's just like a wall of games and you have absolutely no idea where to start. Well, this has been solved. In libraries, you have a card catalog library system where all of the French history books are next to each other and so on and so forth.

34:46Nabeel Hyatt:And so this is the kind of thing that would be like literally impossible without AI, but we basically have a bunch of design prompts in Claude back in those projects that help to categorize into a kind of a Dewey Decimal system of board games, all of the games in existence and then looks at each game and applies them. So if you look at here, this is a list of games that are in there. And if you look at the bottom, you'll see the TLCS code, like 420.5 on this game, Nemesis Lockdown, right? And that means it is a category name in cooperative. That's the 400 section is cooperative. The 200, the 20 section, 420 is adventure co-op games in adventure.

35:24Nabeel Hyatt:and then the dot five is the weight of the game, one being the lightest weight and five being the heaviest weight as in complexity of the game. And so now you can imagine if you're interested in playing an adventure co-op game with your friends, you walk over to the 420 section and they're all nicely grouped together and then those colors, those yellows and reds are how hard that game is to play and so on and so forth. Impossible, literally impossible without AI.

35:50Andrew Mason:We are really taking people on an adventure here. It's truly the most ridiculous idea that if he had proposed this and we didn't have AI, I would have gone to the mat like this is such a crazy thing that nobody's going to want and it's going to be such a waste of time. But because we do, we can indulge in this kind of stuff and it's really fun. And maybe for every – there's five of these. One of them will be the best thing ever and maybe this is it. Okay, so you've shown us your Airtable. It's hooked up to this workflow, kind of like a Gentic system on the backend. What does that expose to end users?

36:32Andrew Mason:So let me show you this now, right? So it looks like it already set this up when I tried this demo earlier, and it found that. So it's already set up a game for me this weekend with Slay the Spire, a great dick building game. I've scheduled it for this weekend and created a request. And so now what's happened is it's gone out on its own and it's messaging other people who like these kinds of games and it's trying to build enough people to schedule a game. And this whole flow we're using N8N for. So basically what happens is you send a text to this phone number. We hooked up Twilio. It comes in and just does a little bit of basic formatting.

37:12Andrew Mason:then it looks up my phone number in the Airtable members table here and it finds out who I am and it figures out if I'm a user or not if I'm not a user it has a whole flow and this is the agent you can see there's these long prompts that basically do all of the routing that we've learned that says what to do if it's not a user I'm not going to focus on that one I'll focus on the if they are a user. So if they are a user, we have this like central agent and these are all the tools that it has access to. So these are all Airtable MCP tools. So you have something that can just like get a list of records, something that can create a new record.

37:58And the instructions

38:00Andrew Mason:for how and when to use that are all in this big prompt that we've put together, which you see gets quite long and it's just been like iterated on over time and it's crazy because like you'll see there's parts in here where it's just saying you know use the use the create record tool to to do x y or z and you just reference it in plain text and it knows how to go and find that

38:31Nabeel Hyatt:yeah you're not calling a function or anything like that you're just just literally writing down what you think it should go do and then it'll figure out how to find the function and use it.

38:39Andrew Mason:And so this agent available to your customers via text, so super simple interface, but is routing to your source of truth data for your business. And then it's engaging with other customers saying, I got somebody who wants to play this game this weekend for an hour, an hour and a half. If you're in, let us know and schedule it and it brings those people together. and as somebody who has put together children's birthday parties just even that seems very very useful and i can imagine we've got to turn this into a startup no third startup that's when you when you you've joined my club is when you're you're doing three but um what i imagine as a business owner though like it's cool as a customer that's nice it's a nice experience it's fun as a builder.

39:29Andrew Mason:But as a business owner, you are probably then getting much closer to, you know, your gentle break even goal of higher utilization, more bookings, more people with lower sort of person to person kind of grunt work that has to happen to make that happen. I'll just say, like, I guess so. We never thought of it that way, though maybe we should have. But it's just the nature of this enterprise that that's not the calculus that's typically going through things on i i will say though that um it's not about replacing a human job there's no like a human couldn't do this type of work at this level where you're going out and like contacting three members to see if they can play the game and when one of them says no you contact another it's like maybe you could do something like that but it for organizing a board game night.

40:25It just doesn't make sense.

40:27Andrew Mason:It would be too slow. So this is just, like Nabil was saying earlier, an example of the kind of experience that just couldn't otherwise exist.

40:39Nabeel Hyatt:Yeah, I would say a lot of our experiences started with a situation where if you had a, in that first group of things, it's the business advisor, the co-pilot, which you're right, if we had people that had done this work before and you surrounded yourselves with a bunch of advisors, those are human tasks that get you a couple of business plans. And we built up Python capacity planning tool and a bunch of other stuff that like way outside of even the scope of what we can talk about here that we just built. The second phase of this is it's just enabling experiences that you just never would have done with humans, period.

41:12Nabeel Hyatt:It's not replacing anything. We just wouldn't have offered these services. You would not have categorized by hand 200 plus, 300 plus board games into a category system you would not have set up a bespoke way that i can natural language say a board game category not even a game over text and say like i want to play an adventure game this weekend who likes adventure games that can play between six and eight o 'clock and then have that just like manifest itself into a space and then a table get booked very unlikely any of these products would even exist and and frankly the whole project wouldn't have exist without AI in the loop.

41:48Andrew Mason:And then before we wrap on this section, I just have to call out, looking at everything you've shown, it doesn't seem like you wrote much code to make this happen either. It seems like you wrote a lot of prompts and you hooked a lot of tools together, but I'm not seeing you pop open cursor in a repo. This all looks like kind of no code stuff that folks could use in their day-to-day, even if you're not a software engineer. There's a little bit of code in the N8N stuff. So I think with N8N, you still need to be a little bit technical with the way that things exist now. But that's just a matter of time before this becomes more and more something that anyone can do.

42:29Andrew Mason:We have repos for the website that we built, and Nabil built a whole ratings system for members to rate their games and get recommendations and stuff like that. So there's that, too. It's just we figured that was a little more boring.

42:46Nabeel Hyatt:Yeah, we also have a Python application where we took those personas that we talked about earlier. We fed them with an event programming and then a pricing plan. And then that goes into a Python application that uses Pulp and CBC to basically build out an integer linear programming application to figure out our capacity planning to basically figure out whether we can break even dynamic revenue business model planning. So we can set up plans, map it to personas, and then it just says, oh, if this is the utilization rate of this place, it'll max at 250 members, and then we can loop that back in.

43:23Nabeel Hyatt:So there is some heavy programming that goes on in some portions of this. But Claire, the vast majority of it starts from prompts and artifacts and projects and then document creation and that loop back and forth between producing a new document, adding it into the context window of like this is a new canonical truth of this project and then iterating and brainstorming from there.

43:45Andrew Mason:Okay. Let's wrap. This has been so fun with lightning questions. One no AI and one AI question. So the non-AI question is, what's your favorite game at Tabletop Library?

43:57Nabeel Hyatt:The next one is my answer. Like part of what I love here, and I know that's not a real answer, Claire, but part of what I love here is I actually really love exploring games. I think of it like a movie opening. And we have a culture at one point where you'd really care about the movie opening and you'd think about it and you'd think about the director releasing their next thing. And that's kind of where I've gotten to with games. And hopefully others can get that excited too.

44:19Andrew Mason:Andrew, do you have a favorite? it. Nabeel is such a big shot in the game world that he couldn't actually name one. He'd just alienate too many stakeholders. I like this game called Sky Team. It's a two-player game where you play as the co-pilots landing a plane. And it's really fun to play with my wife or my kids. So that's both a recommendation and an answer to your question. Okay. And I saw a little of this in your prompt. So Andrew, I'm gonna ask you my rap question first, which is when AI is not listening, how do you get to listen? And I will say I saw a lot of like all caps, caution, exclamation, exclamation in your prompt.

45:03Andrew Mason:So do you have any tricks that you fall back on when you want AI to really listen to you? I saw a really good video that I think Y Combinator put out recently. And they were like, when you get one, one change that I did as a result of that is when you start getting stuck in a rat hole and you can't get it to do the thing that you wanted instead of just like keep on hammering at it just revert back and try something different Nabil do you have a trick you have a fallback I think I I've certainly gone through all of the

45:32Nabeel Hyatt:stages of grief and anger and all caps to try and get the model to do what I what I want my more recent one is to just try and add context I like I think my most common thing now is to try and throw tokens at the problem. So a lot of times what I will do is copy the problem I have, throw it into another model, and then talk back and forth for 30 seconds to try and stretch out and elucidate the problem and add a lot more detail and then copy that prompt back either as a, it might be at that point as big as a PRD or something really detailed or so on and so forth, but just basically like it's confused and I need to give it a lot more direction.

46:08Nabeel Hyatt:And so let me then partner with a different AI to figure out how to tell what's really supposed to be doing.

46:11Andrew Mason:Okay, this has been such an adventure. I can't wait to bring my kids over and see what other cooler games we should get into. Nabeel, where can we find you and who should reach out to you?

46:24Nabeel Hyatt:Any founder building a wonderful AI product should reach out. That's where I spend all my time nowadays. You can reach me, Nabeel, N-A-B-E-E-L, on Twitter, wherever it's called nowadays, or my first name at sparkcapital.com.

46:37Andrew Mason:Okay, Andrew, and what about you? Where can we find you and how can we help you? First of all, if you happen to live in Berkeley or North Oakland, you should sign up for this thing or check it out. Go to tabletoplibrary.com and sign up for it.

46:53Nabeel Hyatt:Thanks for doing the business job, Andrew. That's very much appreciated, sir.

46:56Andrew Mason:But also I have to say, yeah, I'm the CEO of Descript. And the way this whole thing started was my head of marketing was savvily like, oh, there's this great new AI show. You should go on it. and I like sabotaging my press like normally like I normally do was just like I'll go on it but I want to talk about this other thing that I'm doing that I think will be more interesting but Descript is doing a lot of cool AI stuff too we just launched the first video editing agent so like cursor for video and I'll say like here's my like founder tip this thing this whole side project has been like the best thing in the world for rewiring my brain around AI.

47:42Andrew Mason:When you have your ongoing gig, you're kind of stuck in your existing patterns and people you have to rely on. Starting something from scratch that you actually care about where you couldn't do it without AI, it's the best. I completely agree. I have done the exact same thing. It has been transformative in how I lead in a bigger company. And I will do your marketing person a favor and know this was not paid, but we happily use Descript to edit this podcast. So we have been happy and it's how we get the great quality that you're seeing today. So Andrew and Nabeel, this has been so fun. Thank you for sharing all your tips.

48:20Andrew Mason:Good luck with the many businesses and have a great day.

48:24Nabeel Hyatt:Thank you. Thanks so much.

48:26Andrew Mason:Thanks so much for watching. If you enjoyed the show, please like and subscribe here on YouTube or even better, leave us a comment with your thoughts. You can also find this podcast on Apple Podcasts, Spotify, or your favorite podcast app. Please consider leaving us a rating and review, which will help others find the show. You can see all our episodes and learn more about the show at howiaipod.com. See you next time.