Don't forget to jargon check your AI

This week on Equity, Alex was joined by Nathan Baschez, the CEO and founder of Lex, an AI-infused online writing tool that recently raised capital. Together, we're talking through a few key topics that have been top of mind in recent months: How many AI-powered, or AI-using writing tools can the market support? How far into the generative AI moment we are, and how much we should anticipate in the form of technology improvements? And then we discussed the nuts and bolts aspects of pricing an AI-powered service and other financial matters related to building a service today that leans on artificial intelligence. The last question is far from idle. Recall that back in 2020 there was conversation amongst venture players about the economics of AI startups, with the perspective at the time indicating that while the cohort might have more difficult early economics, that their numbers (gross margins, really) would improve over time. But what about when a startup is using, say, an OpenAI API for its core AI work? Will similar efficiencies bloom? Equity is back into its regular groove now that Disrupt is behind us — more to come! And before we go: Check out the UpFlip Podcast where you get to unravel how great businesses are built, how they are run behind the scenes and how their success can be replicated. We think you'll love episode 79 where they featured this guest who transformed his passion for gardening into a $7.3 million-a-year venture. You can find the podcast on Youtube or where ever you listen to podcasts. For episode transcripts and more, head to Equity’s Simplecast website. Equity drops at 7 a.m. PT every Monday, Wednesday and Friday, so subscribe to us on Apple Podcasts, Overcast, Spotify and all the casts. TechCrunch also has a great show on crypto, a show that interviews founders and more! Credits: Equity is produced by Theresa Loconsolo with editing by Kell. Bryce Durbin is our Illustrator. We'd also like to thank the audience development team and Henry Pickavet, who manages TechCrunch audio products. Learn more about your ad choices. Visit megaphone.fm/adchoices

• 0:11Speaker 00

  Hello and welcome back to Equity, a podcast about the business of startups, where we unpack the numbers and the nuance behind the headlines. My name is Alex and this is our interview show where we sit down with a guest, think about their work and unpack the rest. Today, we are talking to Nathan Bichaz about the ins and outs of startups using modern AI tools in their software projects. You may have heard about Nathan. He's the co-founder and CEO of Lex, an AI-powered writing tool spun out of subscription media company, Every, and he previously worked at both Gimlet Media and Substack, where he was the VP of Product. Nathan, welcome to the show. It's so great to have a fellow technology and media dork on the pod.

• 0:48Speaker 01

  Thank you for having me. It's honestly, I feel like it's a danger for everyone listening that we're both, our interests are so aligned, we could go on a long one here.

• 0:56Speaker 00

  We could go on a long one here and probably just completely off the rails and right into the ditch.

• 0:59Speaker 01

  Exactly.

• 0:59Speaker 00

  But we're going to keep this relatively focused on AI and how founders should think about use of AI in their products and what you've learned so far. But to set the stage a little bit, you and I caught up a couple of weeks ago, I want to say now, time flies, because Lex raised 2.75 million, led by True Ventures, that was last month. So I'm really curious, you know, we talked beforehand, then the news came out. How has response been?

• 1:24Speaker 01

  It's been awesome. Honestly, I wasn't sure exactly what to expect because when we, like the sort of first version of Lex launched almost a year ago now.

• 1:32Speaker 00

  Wow.

• 1:32Speaker 01

  The environment around AI was totally different. It wasn't saturated, especially maybe image generators were like getting a lot of attention then, right? Like stable diffusion was kind of recent and Dolly 2 was still feeling very fresh. And the text side of things hadn't really gotten that much hype yet, which was interesting, right? And so I think when Lex launched, it caught like sort of just the right time where people were primed to feel like AI was getting better, but they didn't realize exactly how or how they could use it. So the response to that was like, Kind of crazy and overwhelming in a way.

• 2:01Speaker 00

  Yeah.

• 2:02Speaker 01

  And I was kind of like, okay, like, what's going to happen when we announce the fundraiser? Because, you know, the product's been out for a little while, of course, like very nascent. And we made a lot of progress with it. But yeah, it's been awesome to see. I feel like it's kind of shifting to a place where it's no longer just a demo. And people are starting to kind of like, take it seriously and, and adopt it, you know, as an individual or in organizations and stuff and making progress.

• 2:25Speaker 00

  So one thing I always hear from founders is, you know, we're, we're raising money and we're announcing this capital raise because we want to increase our visibility so we can hire, or we're looking to, you know, call our customers that we're going to be here for a long time. Or there's usually kind of a standard, kind of bushel of reasons why people talk about their financing publicly. So in the case of Lex, did you have an expectation going in that you were like, all right, we're going to announce this and then hopefully, I don't know, 10 developers send us an email or whatever?

• 2:49Speaker 01

  Yeah. Honestly, that was the main one is just hiring, right? Cause you never know, I think, who in your network is actually thinking, you know, maybe I kind of want to think about something new, right? Especially something as sort of crazy, you know, as an early stage startup. It's not always the most rational career choice in a lot of ways. You have to do it because you just can't imagine doing anything else. There's a lot of folks I've met over the years who I think would be awesome to work with and you never know who, who might reach out, like, you know, when you announce something like this. And I think, yeah, just to get the chance to share the story with y'all was amazing also, just so. That was, that was a fun part of it too. But yeah, definitely engineer hiring was a big part of it.

• 3:26Speaker 00

  This actually reminds me of something that you told me when we were talking and I, I want to bring this up because it's an interesting job classification. You said you were looking for, if my memory serves, a couple of founding engineers.

• 3:36Speaker 01

  Yes.

• 3:37Speaker 00

  Is that right?

• 3:37Speaker 01

  Correct.

• 3:37Speaker 00

  So I've heard about staff engineers. I've heard about various gradations of like, you know, engineer seniority in the software world, but founding is certainly an interesting way to go about this. And what it sounds to me is you're looking for co-founders, but you've already raised money, so it feels a little bit late. So for founders out there, what does that mean?

• 3:54Speaker 01

  Yeah, so founding engineer is a title that I've seen crop up recently a little bit more. And I think it does a good job of kind of, you're on the founding team, right? Not necessarily a co-founder, because yeah, we already raised money. We already have got a decent amount of users and usage and all that stuff. But I mean, really, the business has not been built yet, right? Like, we haven't launched paid plans. There's still a lot of stuff to figure out in the product. So if you're hiring like a staff engineer or a senior engineer or something like that, it kind of implies there's like an organizational structure that you've got some place within. And there's a business that you're just kind of helping move along, right? Here, we're creating the business together.

• 4:28Speaker 00

  Right.

• 4:29Speaker 01

  Still, it's still very, very early. So for a certain type of engineer, that's exciting. That's exactly what they want. They want more ownership over the product and over the direction of the company, both literally in terms of like equity and in terms of just if you think something's a good idea to do, there's very little standing between you and us doing it. And that's just really appealing to a certain type of person. And they're totally willing to put up with the fact that, you know, right now, 99% of the code was written by me and it's kind of janky in a lot of ways.

• 4:54Speaker 00

  I wasn't gonna say it, but like, I don't, I mean, I've known you and your various guises. I've known whatever he's been doing, so I've talked to you in that context. And like, you know, yeah, when I think of someone who I think is very good at slinging words, I don't immediately go, I bet they're also amazing at code. I bet you're giving yourself not enough credit here. But I can also see how going into a codebase that was written by one dude.

• 5:15Speaker 01

  Yeah, exactly.

• 5:16Speaker 00

  Is not going to be like working for Google.

• 5:18Speaker 01

  It's funny, there was an engineer who I was talking to and it was, it was the funniest burn because he was like, you know, seeing your code, like, it's actually kind of inspiring in a way. Cause like, yeah, it doesn't actually need to be good code to be a product that's good and like that people like.

• 5:31Speaker 00

  Oh, man.

• 5:33Speaker 01

  He was like, don't take me wrong, like, it's not horrible. You know what I mean? Like, you're cutting corners in all the right ways, but like, you know, there's corners cut.

• 5:41Speaker 00

  This is like, this is like telling someone you go on a date with. You're like, you know, for someone with your natural level of attractiveness, you've really done a lot. I mean, the accessorization, you know, your buttons, you know, that's, that's brutal. Did you, did you take that person to hang up the call immediately? Cause I would have just cried.

• 5:56Speaker 01

  No, yeah, no, I love it. I don't, um. I mean, I think it's a point of pride in some ways. Also, there's just a lot to learn, right? Like, and I don't have a computer science background. I have more of like a just self-taught person writing code to build the products that I wished existed type background, which is, you know, different type of thing.

• 6:14Speaker 00

  Yeah. I mean, I'm not going to have you do those horrible like whiteboard logic tree tests you get at a big five company, but that also I don't think means that you're good at scrappily building a product. And I think honestly, if janky code is your first code, jank away because it's more important to get it out there and learn. Though I'm not going to start quoting leading startup ideology here, but I do think there's something to be said for rapid learning in the marketplace.

• 6:36Speaker 01

  That is for sure.

• 6:37Speaker 00

  And this is where I want to start because one thing that I've been just blown away with in the last, since basically chat GPT came out and I began to very late give a shit about generative AI and LLMs is how fast things are changing. I mean, from GPT three to 3.5 to four felt like a weekend. I know it was longer than that, but it just felt very quickly. New models from Facebook, major companies being built, so much going on. And so, given that you do have generative AI built into Lex, how do you balance your time between learning about the new state of the art stuff and actually deploying stuff into your code? Because I feel like you could spend all your time learning and no time actually writing code and make no progress. So I'm curious about how you balance that as a founder.

• 7:23Speaker 01

  For me personally, I have no idea if this is the right answer for every company, of course, but for me personally, it's like 80% just building features, looking at our data, talking to users. I mean, okay, so it's 80% that and then the other 80% right now is building, building the team. So talking to engineers, you know, doing projects with folks, because I think there's no better way to sort of learn if you're, if you like to work together than to actually work together, which is a philosophy we're borrowing from another true ventures company, WordPress, Automattic. They're sort of famous for their, for their project-based interview process. But yeah, and then I spend very little time, honestly, like, I would say, Once a week, I have like a moment where I'm looking into some of the research-y type stuff, but I don't think it's the most important thing for us right now. At some point, it for sure will be, and someone on our team will need to know a lot about

• 8:08Speaker 01

  that stuff. But I'm not a machine learning person, or PhD, or anything like that. And so at some sort of like intellectual level, it's just fun, right? Like, I can do it for fun to try and learn how transformers work, or what's new with, maybe you've got like a big machine learning model that gets edited by a little machine learning model that is more personalized to each user. And that's the way you can do more efficient fine tuning of large language models or whatever. Like, there's all sorts of like ideas that float around and are interesting.

• 8:34Speaker 00

  Yeah.

• 8:34Speaker 01

  But it's ultimately too kind of academic to be practical right now. Because the most important thing is, the capabilities are already so much better than what most writers are actually using or taking advantage of.

• 8:45Speaker 00

  Yeah.

• 8:45Speaker 01

  So, the bottleneck is not, we don't have powerful enough large language models. The bottleneck is, it's not been built into products that people can actually use in the right way, and marketed to people in a way that gets people on boarded and understanding the value quickly.

• 8:58Speaker 00

  So, the reason why my question was flawed, hearing your response to that, is I was thinking about needing to learn what's new, so you can properly deploy it inside of your product. But it sounds like what you're saying, and then recalling our prior conversations, if you're using, for example, an OpenAI API, you don't actually have to spend that much time going upstream, because you've already got an endpoint you can just link into. And so if they make progress, you get better stuff, but you don't actually have to know what happened over in someone else's public cloud cluster.

• 9:31Speaker 01

  Yeah, of course. And there's also stuff, like OpenAI launches things, and so we need to learn about that. And that's sort of a different level of, you know, like when they launched function calling, there's a feature we built on top of that, or that we're building on top of that, almost immediately.

• 9:43Speaker 00

  Mm-hmm.

• 9:44Speaker 01

  And there's a lot of useful stuff that they launched, it's more kind of immediately applicable to us. But yeah, I guess I misunderstood the question a little bit. I don't, in terms of new, I guess there's a pipeline, right? There's like new in the research sense, and then there's new in the like, it's something that's practical for a startup to use immediately.

• 9:59Speaker 00

  Okay.

• 9:59Speaker 01

  But even then, honestly, a lot of this stuff, I still don't feel like it's the most important thing for us at the moment. It's good to be aware of, like even the new stuff that we could just currently use, like not research type things, like Llama 2, right? That's something that people are setting up hosting services for. It's pretty easy to like, get to a point where you can just swap it in, almost with the exact same API calls as you're making to OpenAI. For now, the difference is relatively small, and just the most important thing you can do is like, Better understand what problem you're solving for people, build features in a way that work in their lives better. Onboard people into those features in a way that helps them understand it more quickly. All the most important things we're focused on are things that pretty much any other SaaS product always focuses on.

• 10:40Speaker 01

  And the truly new bit is like, I would say relatively small, but of course important to stay up with as things go, but relatively small.

• 10:48Speaker 00

  Okay. No, that's incredibly helpful for me as a way to kind of ground my understanding of the problems. So what you're saying is, look, it's not that we're having a hard time making delicious pies. It's getting those pies to the right people at the right time. We have many, many baked treats. Okay. That's interesting.

• 11:02Speaker 01

  I think we're kind of, so making the pie is still hard, but there's like really great supplies of like fruit and stuff, right? And like, we don't need better fruit. We just need like to put it together into the pie in the right way. Like, we need to actually find out, oh, like, let's make a crumble or whatever. And like, people like that better. Does that make sense? Like if the LLM metaphor, part of the metaphor is the fruit, like stuff?

• 11:22Speaker 00

  I was thinking, I was thinking that OpenAI was making the pies and you were just using. I feel like this metaphor is not the best approach for this, but I think, I think the last couple of months have clarified my understanding of how you go about as a founder, building an AI powered product, thinks about what's new and where to spend their time.

• 11:39Speaker 01

  Yes.

• 11:39Speaker 00

  And there's, there's three buckets. There's, you know, just regular software focused work. There is applying recent innovations from major providers and open source models. And then if you want to, a small fraction of your time is spent more in the future researchy. You know, what's coming in two years versus two months type stuff.

• 11:56Speaker 01

  Totally. I do think there is actually something really important that the sort of like pie metaphor got to though, which is this idea of like, is the thing that OpenAI or any large language model provider, is the thing that they provide basically the whole deal? Or is it an ingredient in the deal?

• 12:12Speaker 00

  Yes.

• 12:12Speaker 01

  There's this term that people love to use called GPT wrapper, right? Which I think is really funny because it's kind of like an insult, right? But calling a, you know, a really robust like product that has AI built into it, using OpenAI or anyone else, a GPT wrapper is kind of like calling, you know, the iPhone a CPU wrapper, right? Or calling a laptop a CPU wrapper. It's like the CPU, the GPU, the LLM, whatever you want to call it. Call it like the whole point is it's an ingredient that is important, but it's not useful on its own at all. It needs to be built into a product.

• 12:42Speaker 00

  Yeah, the way that I would think about that, and by the way, we're talking about rapper with a W as in an encasement versus rapper with an R, which would be a artist. Not everyone who listens to the show, English is their first language. So I try to occasionally, but I'm like, okay, that word's gonna trip people up.

• 12:56Speaker 01

  Yeah, yeah, yeah.

• 12:56Speaker 00

  No, I appreciate that. I think the the diss GPT rapper to me comes from when people like take a fully baked thing and then sprays some whipped cream on top and call it a new product. But in my experience, I have not seen that much of it. So it's almost like a pejorative aimed at a missing group of startups. Because I don't think Lex is a rapper built around GPT three. I think it's a modern writing platform that has AI help built in and, you know, from our conversations, mostly about helping people get unstuck, helping them move forward and so forth. And yeah, if you call it a rapper, I just think that would be dismissive of of why it's good.

• 13:33Speaker 01

  Right. I mean, that's that's how I think. And I think there's a lot of people building. I mean, it's not just about Lex. It's about anyone who's using open AI APIs or using large language models that they themselves didn't train. I just think it's a huge misnomer that it's like, oh, yeah. All the value is going to be created at this sort of large language model layer because it's so technically complex to build and so expensive to build. And it's true that there's some economies of scale there. It's hard to do something that's technically sort of at the frontier, but it's not always the case that those are the layers of the value chain where most of the profits accrue. Like, I mean, it's really technically hard and complex to build like a smartphone touchscreen, but the people who supply that to Apple like don't make very much profit, you know?

• 14:14Speaker 00

  And also, what is the high margin product today will not always be the high margin product down the road. I mean, like television used to be this like massive capital purchase. And now if my baby threw a block through our current television, I'd be like, good, I fricking hate that curved monstrosity. I've been meaning to buy a new $200 plasma because we figured that out to the point in which it is now just a commodity.

• 14:32Speaker 01

  Exactly.

• 14:33Speaker 00

  So. AI competence, you mentioned how I think it was Lama 2 and GPT-3 for your needs are kind of the same. How fast do you think these models are going to keep improving? And then as a founder, does that rapidly unlock new features for you? Or are you more going to be just using the improvements to these models as we understand them as a way to make things you've already built better inside the product itself?

• 14:56Speaker 01

  It's both. It definitely unlocks new features. For sure, existing features get better, but I think it's hard to understand unless you spend a lot of time or you kind of go through this cycle of, oh, this would be cool. And then you try and get the AI to do it and it's just not quite good enough.

• 15:10Speaker 00

  Yeah.

• 15:10Speaker 01

  So we don't ship that. It's hard to understand how much that matters. And even the difference between certain features that are sort of viable with GPT-4 versus features that are not viable with GPT-3.5 or whatever, there's a big gap there. And I don't think very many people appreciate this. Yeah, I'm excited about future even larger models that could come out. I have no idea about the timeline or how much better or whatever. Like, all that stuff, to me, feels kind of like outside my control. And I don't know. And I'm excited and I want to stay on top of it. But yeah, I think it'll unlock new types of features.

• 15:40Speaker 00

  All right, we're going to take a quick pause here. But when we get back, notes on what GPT-4 can do that 3.5 cannot.

• 15:51Speaker 00

  All right. So Nathan, the thing that you just said that blew my mind is that there's things you can do with GPT-4 you can't do with GPT-3 because it's not quite there. And to me, looking at a release of 0.5 seems relatively modest, but it sounds like there's quite a lot there. So could you give us an example of something that might be possible with GPT-4 that's just not possible with 3 or 3.5?

• 16:11Speaker 01

  Yeah. So there's a feature we're working on, which is funny because it's probably a terrible idea to talk about something that we haven't shipped yet on the TechCrunch Equity Podcast.

• 16:17Speaker 00

  It's just you and I talking. I don't see anyone else on this Zoom call. You know, it's fine.

• 16:23Speaker 01

  So basically, it's called checks. And it's like, you know, there's spell check, there's grammar check, of course, but you can also check, you know, using large language models for higher level things. Like, one thing I have a lot of trouble with when I write is putting in a lot of like, maybe, probably, I think, you know, like, kind of hedging, right? Not in a way that's necessarily like just, you know, whatever, being accurate, but kind of it just gets in the way of what I'm saying. So you could tell a large language model, hey, like if I say any of these things and it doesn't seem like 100% necessary, just fix it, right? So it's like confidence check almost. You could have jargon check. If I'm writing and I tend to slip into like, you know, earlier I said GPT rapper, right? Not everybody necessarily knows what that is, especially in audio, the word rapper you normally hear referred to, you know, Drake or whatever, rather than OpenAI based products or whatever.

• 17:07Speaker 01

  And so you may want to at least define it, right? And so if you're using any jargon that's sort of like the average reader may not understand, put in an explanation. So there's some checks that work really well with GPT 3.5, but it's much better, like much to the point where it basically, it only really works if you do GPT for especially for some of the more kind of advanced ones. And you know, that's an issue potentially because it's expensive, but also I just look at it as, yeah, like we're sort of figuring out what are the right ways that AI can help writers? It's interesting because there's a lot of businesses that got in a lot of trouble over the past. I should say startups really more so than businesses. They got in a lot of trouble over the past 10 years for kind of hoping that their margins would improve later on down the road and their cost structure would improve.

• 17:48Speaker 01

  But a lot of those were based on doing things in the physical world, like having a scooter that you can rent or something like that.

• 17:53Speaker 00

  R I P bird, man. There's still birds, scooters around Providence sometimes, and I'm like, I know your financials. I will not see you in six months, but thank you for your service.

• 18:01Speaker 01

  Exactly, but I do think that any startup founder should be kind of cautious to just sort of hope that things will get better. But I do think that everyone I talk to who knows a lot about what's driving costs in serving large language models and training them thinks that intelligence is going to continue to get very cheap, very quickly, like faster than Moore's law type progress. And I mean, we've already seen OpenAI cut the price from GPT three to GPT 3.5 is like, you know, 10x cheaper or whatever. And so I think we'll continue to see, or we'll continue to ride that curve is my bet.

• 18:31Speaker 00

  I hope so, because one thing that people forget is that earlier on in the days of AWS before was this globe spanning behemoth that everyone uses like fucking water or electricity. They used to cut prices all the time.

• 18:42Speaker 01

  Yeah.

• 18:42Speaker 00

  Like things would get cheaper and they would cut prices for storage or compute, whatever. And like, it was this crazy deflationary power giving amazing. Thing. And now because it generates nearly all of Amazon's profit, depending on the quarter.

• 18:53Speaker 01

  Yeah.

• 18:53Speaker 00

  They have to juice it a little bit more, but I do hope that we do see a similar price sort of declines for AI powered services. Ah, that's just, it's so interesting to me that they're so much better. Now narrowing this down to Lex, though, I want to do a jargon check. And I'm writing the document. Do I push a button that says, you know, Lex jargon check? Or do I say, at chat, please run a jargon check on this? Like, do I, the user, have to come up with the idea of that? Or do you build a feature that does the asking or the prompting for me? And I'm asking this because I think founders probably want to know the best way to add stuff baked in.

• 19:27Speaker 01

  Yeah, I think it's good to have both the sort of really flexible ways of interacting with the AI where the users can kind of like request whatever. And the AI can just come up with something useful, like in our case, you can leave a comment in the document, at mention the AI and ask it to reword it in any way you want it to, or answer any sort of research question or brainstorm any sort of idea related to it. So we have that version of it. And then we also have the stuff that's more optimized on the rails and kind of like pre-built for you. So the jargon check will be something that you click a button and it's running the jargon check for you. Or really what you'll do is you click a button and it'll say, okay, what all do you want to check for? And it'll remember your settings and you can create even new checks if you want to. For instance, like, hey, we always want things to sound a little quirky for our brand.

• 20:09Speaker 01

  And we want it like, here's some good examples of quirk and here's some bad examples where you might think it's quirky, but actually we think it's just cheesy or whatever. And if you give the AI that and you run it over, you know, every paragraph in your document, it comes up with really great stuff, especially on GPT four.

• 20:21Speaker 00

  Yeah, I really feel like every single company is going to have, like, you know, every company has like, here's our logo, here's our font. And then I think every company is going to be like, here's our voice and it's distilled down into this bucket of information that we can apply to different LLMs. So if you are using anything generative, you have to use this data bucket. Otherwise you're going to sound like Exxon Mobil instead of Snap, you know, which are very different corporate images, I would say. Similar amounts of CEO control. Sorry. No business jokes on the podcast. That's a rule. So I'm curious about the market you've picked because as a writer, clearly you are making something that I am going to be very interested in. And I'm torn between thinking that you are going after a market that is too large and too fragmented and too competitive because there's everything from, you know,

• 21:04Speaker 00

  old school Microsoft Word all the way through to things like Notion. There's just so much that you can use to write. On the other hand, it's so big because everyone needs to write. So can you tell me a little bit about how you decided that this was a market that was going to be worth tackling from a business perspective? And again, I'm trying to think about like other founders picking where to deploy AI powered software.

• 21:23Speaker 01

  Totally. It's a tricky one. I don't think that we have all the answers to this one right now. I think that a big part of what we're doing right now is seeing where the product is the most useful and kind of tailoring to those use cases and leaning into it. And right now we are probably a little too broad. We'll see how it goes though. I mean, it could be the case that we've picked the exact right level of focus, you know, but with any early stage startup, it's always there's, there's adjustments that happen along the way. But I will say, you know, the reason why we chose to sort of focus where we are right now, which I should sort of define. So our current focus is if you're writing something that you're serious about, there's maybe multiple rounds of revision. It's probably longer than, you know, a text message or a short email. It's probably a couple paragraphs at least, you know, maybe more like in the couple of thousands of words, probably not book length. There are

• 22:07Speaker 01

  people writing books in Lex, but the AI becomes less useful for really long stuff just because of some of the current tech limitations. Then that's kind of our target market. And the reason why we chose that rather than just like, hey, this is AI for, you know, grant proposals, or this is AI for whatever, pick, pick your niche is just historically, we haven't seen a lot of very vertical focused text editors. And you can think of all sorts of reasons why you might want to build one of those even absent AI, right? Like you could, you could imagine, yeah, why isn't there a version of Microsoft Word that's like specifically for lawyers? Why do they just still use Microsoft Word, right? And the answer, my theory is that these tools have massive network effects that are really commonly, I think, underappreciated.

• 22:50Speaker 01

  And if you want to view an example of like a more modern company that was able to kind of overturn the old network effect and build their own, Figma is a great example. It feels like there was a period of like a year where it went from nobody's using Figma, like everybody's using Sketch, or maybe they're still using InDesign or Photoshop or whatever, to like everyone's using Figma. And that doesn't happen absent a network effect, right? And I think the network effects of creative tools are really based on just you want to use a tool that other people know how to use if you're working with other people, A. But then B, things like people create tutorials, people create extensions and plugins, you know, you learn from the best. So if you're like a new designer and you're asking someone who's more experienced, like how do you do what you do? They'll talk about their process and embedded in that is the tool that they use, right?

• 23:32Speaker 01

  And so there's sort of intergenerational network effects, like what I just described. There's ecosystem network effects. There's direct network effects, like just within a company, similar to Slack, you know, you set up your team on Figma and your team's going to use Figma, you know?

• 23:44Speaker 00

  Right.

• 23:44Speaker 01

  Anyway, I think that obviously that kind of business, it makes it hard in the early days, but it's what makes it more of a potential venture bet in the later stage if it works.

• 23:52Speaker 00

  Yeah, because if it does hit that kind of mushrooming moment, then it will grow to be massive and then you'll be able to theoretically make lots of money and hustle for everybody. But on the other side is if you don't have monetization in place, when your usage explodes,

• 24:05Speaker 01

  Right.

• 24:05Speaker 00

  you could rack up quite a bill. And so I was thinking about this show and I talked to a lot of people who are building software companies and it feels like a slightly different cost structure because

• 24:16Speaker 01

  Yeah.

• 24:16Speaker 00

  given that anything that's using a AI API has more discrete usage costs for those API calls.

• 24:24Speaker 01

  We've got COGS, basically, unlike most software businesses which have very minimal COGS, like we have real COGS.

• 24:29Speaker 00

  And COGS is cost of goods sold, essentially deduct that from revenue and you get gross margin. Software companies have high gross margins and I'm sure that Lex will-

• 24:36Speaker 01

  Unless you're Spotify.

• 24:36Speaker 00

  Sorry, that's a sore, sore spot. I didn't even know that was a button you could push on my body, but yeah. So because you're using modern AI tools

• 24:47Speaker 01

  Yeah.

• 24:47Speaker 00

  in your product, does that shorten the window of time you can build a startup that has a lack of paid features essentially? Does it increase your burn? So you have to start charging faster?

• 24:57Speaker 01

  Yeah, it does.

• 24:58Speaker 00

  So how like, I mean, if you didn't have that cost, how much more time might you have to experiment and futz around before you have to charge versus in your current roadmap?

• 25:07Speaker 01

  I actually think charging is helpful for experimentation because if people are just using your thing for free, you don't really know what they think of it. And of course, we want to have a free tier that people can use. And I do think that if you look at companies, for example, like we're just talking about Figma, there's no accident that they have a pretty good free tier, right? And they also have a pretty good network effect. So definitely want to have a generous free tier. That being said, you know, yeah, important to charge early and something that we'll do soon, just because I think it helps the experimentation, you know, and it's not, it's less about like keeping our costs under control or anything like that. And it's more about like actually sorting out who's really using this thing. Like, what's the real market for this?

• 25:46Speaker 00

  And then in terms of pricing, because you do have COGS, I presume that, you know, down the road, I'm not going to pay $1 a month for Lex. But given the way costs are currently baked into the use of AI tooling, can you sell this to a power user for tens of dollars per month? Or is that not work out financially?

• 26:04Speaker 01

  Yeah, it would.

• 26:05Speaker 00

  Awesome. Because I've been very curious about that, because I presume that in the next 12 to 18 months, I'm going to be personally paying for a number of tools that have an AI component to them. And I'm just not sure quite yet what my price expectation should be. So I'm not really anchored, if you will, to anything apart from like GitHub Copilot, which I don't use, and then Office 365 Copilot and stuff. So there's a couple of early points, but they're from Microsoft. So I don't really want to like, sure, take that as gospel for everybody out there. But it sounds like it's going to be consumer friendly when it does come.

• 26:32Speaker 01

  Exactly. And it depends on the type of power user, right? So like, there's always going to be, for instance, some people who it'll be really hard to make a profit on if they're like really going, going wild with it. But like, you know, someone who even like a professional who uses it pretty much every day as a part of their work and uses the AI features. Yeah, you can charge tens of dollars a month, like not, not getting into the hundreds and still be profitable on that user.

• 26:53Speaker 00

  Awesome. No, I'm really glad to hear that because that means that stuff that I want to use AI in should be accessible to me at a price point that I'm okay with. And if what you said comes true, I mean, do you see a decline in price as compute power goes up? Then frankly, I don't see why any AI power company can't have SaaS margins in a couple of years, once we get all that sorted out. So it sounds actually like the future of that isn't too far away, which is exciting.

• 27:15Speaker 01

  It is exciting. Yeah. And I think especially as open source is getting better, it's this huge, I mean, the big difference between AWS and large language models is AWS is providing a service, whereas large language models really under their hood are its information. It's a big collection of ones and zeros that has extraordinary intelligence encoded in it. But you can download it. And then like, it's up to you where you want to host it. Maybe then you have to go to AWS. But the product AWS provides is like, they have buildings that like require a lot of electricity and like skilled people to keep it running and humming. And large language models aren't that. They're just information. And there's the whole idea of like, information wants to be free, which is a very old like, kind of idea on the internet. And, you know, I don't know how much it's like, I always believe that. I think there's a huge point of like, especially creators, like charging for their work.

• 28:00Speaker 01

  You know, if you, if you're creating writing or videos or whatever that people really value, like, yeah, you can and should charge for it. But for something like a large language model, where it's not like a personal relationship between me and the craftsperson who made it, you know, like, it does affect the economics of that business, for sure.

• 28:16Speaker 00

  Yeah. No, I'm just, what I love about this particular recording is I presume in six months, half of this is going to be at least slightly wrong. And that's why this product technology is so interesting to me, because it feels a little bit less iterative than what we've seen in the last five, 10 years.

• 28:31Speaker 01

  Yeah.

• 28:31Speaker 00

  And I was talking to a colleague of mine about the kind of like 2010 through 2015 internet, and how tech felt fun.

• 28:39Speaker 01

  Yeah.

• 28:39Speaker 00

  Like, there were more like, there was more mobile OSs competing for the crown. You know, things weren't quite as like, steady state. And so it felt like because there was more stuff going in different directions, there was just more to chew on total.

• 28:51Speaker 01

  Yeah.

• 28:51Speaker 00

  And I just, I kind of missed that era of tech. And so I'm hoping that LLMs, transformers, whatever you want to call it, this new generative AI moment, is going to make tech a little bit more sclerotic, and therefore more fun.

• 29:02Speaker 01

  Yeah.

• 29:03Speaker 00

  And so far, so good. I just hope it kind of persists in the next couple of years, because if I have to cover more incremental SaaS companies, even I will eventually get tired of that.

• 29:12Speaker 01

  Yeah, totally. No, I mean, I'm right there with you. I think that it's fun to live through a big new shift in what's possible and see the ways that it sort of plays itself out, some ways faster than anyone would have expected, and in some ways, much slower, right? You know?

• 29:28Speaker 00

  Yeah.

• 29:28Speaker 01

  And it's always interesting, this sort of juxtaposition between, I feel like there's like, the societal equivalent of thinking fast and slow, you know, of like, things like ChatGPT come out, and it's like fastest growing product of all time, or whatever, for a period of time. And then, but at the same time, like, the market we're targeting, like writers, like, most people don't use AI, really.

• 29:46Speaker 00

  Well, I mean, that's the funny thing about your TAM. Your TAM is everyone who's alive. The question is just getting them to use Lex and then also getting them to not buy in per se, but kind of learn to use intelligently these tools that you're trying to build. Because when I was playing with Lex, it took me a little bit of time to like, Yeah. Okay, so I'm here. Now is when I can use. It's like having a new screwdriver to reach for. Yeah, exactly. When you have a particular issue. Yeah, yeah. We don't have time to get into privacy. We'll do that the next time we have you on the show to talk about AI and startups, but there's a lot going on there. But Nathan, thank you so much for the chat and the context. You know, this moment you have some revenue numbers you want to tell to the press. We'd love to have you back on to talk about how Lex is monetizing when it's time. Amazing. Well, thank you so much for having me. It was a lot of fun. And if folks want to find you on the great wide internet, where can they do that? Twitter's a good one. N B A S H A W.

• 30:31Speaker 00

  And what is the Lex URL? Because I can't not give you a shot, shouting that out. Yeah. Lex dot page. L E X dot P A G E. What is that T L D? Is that a country T L D? Or is that just one of the random ones they added later on? It's one of the random ones. It's like Lex dot limo or whatever, you know? But it's dot page, which I thought was very nice because it's like, yeah, that's what it is. It's a page that you can write on. So if I didn't like Lex, could I get like Lex dot crap or something? Just like a spoof site, like just? Hopefully, yeah, you should. Could you like spite URLs? Someone email us about that if you can. All right, anyways, we got to go. That's Nathan from Lex. I'm Alex. Equity comes out three times a week, Monday, Wednesday, Friday. And of course, you can follow us on Twitter and threads where we tweet and thread under the handle EquityPod. All right, everybody. Bye.

• 31:14Speaker 00

  Equity is hosted by myself, editor in chief of TechCrunch Plus, Alex Wilhelm, and TechCrunch senior reporter, Mary Ann Azevedo. We are produced by Teresa Locan Solo with editing by Kel. Bryce Durbin is our illustrator. And a big thank you to the audience development team and Henry Picabette, who manages TechCrunch Audio Products. Thank you so much for listening and we'll talk to you next time.