Nathan Baschez: The New Age of AI Writing Tools

0:05Hey everyone and welcome to Generative Now. I am Michael Magnano. I am a partner Lightspeed And this week, I'm talking to Nathan Bashes, the founder and CEO of Lex. Lex is an AI-powered word processor designed for writers ranging from book authors to bloggers. Nathan and I talked about how Lex burst onto the scene as one of the first AI products to hit during the AI boom about 18 months ago and the misguided way in which many people talk about GPT wrappers and what separates great AI products from the rest of the pack. Take a listen. Hey, Nathan. How's it going? It's going well. Good to see you.

0:41Thank you for doing this. Pleasure to be here. So I've really been looking forward to this for a bunch of reasons. First of all, you're an amazing entrepreneur. You've got brilliant stories. But I think like in particular, the story of Lex is probably one of the more unique ones because I feel like you've been at sort of AI products like almost since the beginning. And granted, like it's been like only two years or maybe not even, but that's kind of the entirety of like AI, right? Or this current wave of AI. So in many ways, like you're an AI OG. I guess I'd love like the quick overview of sort of how you got started with Lex and kind of what's happened since then.

1:29Because I imagine it's been a lot. If I remember correctly, when you started, like GPT-4 wasn't even out yet, right? GPT-4 was definitely not out yet. GPT-3 was out, but it was before the whole kind of, it was before chat GPT. And it was before the sort of like wave of people building on GPT. At the time, all of the hype in AI was actually about image generators. So Dali was out, Stable Diffusion had just come out. And that's all what anybody was talking about when it came to AI. GPT-3 was out, but people just didn't realize what it could do, I think. Um, although I will say the big sort of success story at the time, um, was Jasper.

2:11There was, you know, they were, they were growing real fast based off of, uh, GPT-3 kind of powering, powering their product. But, um, anyway, yeah, no, it's, it's kind of funny because, um, yeah, it is sort of weird to be like, oh yeah, OG of like AI of like the current wave, I guess, of AI. Um, I don't have a machine learning background or anything like that. Actually, um, uh, my background is really in, uh, like building, building products kind of at the intersection of media and technology. And most of the time I've been on the technology side of that intersection. So, you know, started a company that had like this, uh, tappable visual kind of storytelling format for mobile called hardbound.

2:53Um, went to work at Gimlet media, the podcasting company that, that, you know, very well, um, kind of, did y 'all get acquired like at the same time by Spotify basically, or? It was kind of all part of one big cycle. Same day, actually. Same day. So literally the same time. Yeah. And then after I was at Gimlet, I was at Substack. I was the first employee there. And while I was at Substack, I got really jealous of the writers that were using the platform. I was like, oh my God, it's so boring coding the settings page. I want to write cool things for people on the internet. So I was there for a little while and then left to start Every with my friend Dan Shipper.

3:29and that was based on the insight that it's really hard to build a sub stack solo, but it's, it could be easier if you kind of work within a team and you have multiple writers and you're pooling an audience kind of like a magazine, you know, it's not like a completely original idea, but, you know, built around the sort of subscription unbundled era and sharing a lot more of the economics with the writers that wanted to join the bundle. Um, so anyway, did that for a little while. And that was the first time in my career where I was really writing and editing on a regular basis. And I went from using like Figma and VS code as my main tool set to like Google docs basically.

4:05And, um, it's a real love hate relationship I've discovered that writers have with Google docs, um, or word, you know, it's the other, the other one, if you're writing a book, it's kind of like word or like, I guess a legal contract, but other than that, pretty much everybody uses Google docs. And, um, there's just so many things that I was kind of like, yeah, this isn't great. It kind of reminds me of when I was a designer and before Figma, before Sketch even, Photoshop, right? Because it's like this general purpose. And in Photoshop, you're manipulating pixels and Google Docs, you're manipulating characters, but there's not a lot else built into the workflow.

4:39It's just a very general purpose tool. And so over time, I got more and more sort of frustrated with it and also more and more excited about what else could exist. Right. And so I started thinking about, you know, building, building my own Google Docs. How hard could that be? Right. Like, were you thinking about AI when you were like, well, a little bit. OK, it wasn't the first thought. The first thought was literally it was very unambitious. It was just could I make the minimum thing that would make us not have to use Google Docs that would solve these couple problems we have on a regular basis?

5:13Like getting an image out of a Google Doc. Like, my God, you have to like go to the inspect network tab and like filter to like image requests to like find the actual source image. You can't just right click on it and hit like copy or copy URL like you can normal images. Or like if you want to use an emoji, you know, on the Mac, the keyboard shortcut is I think I just muscle memory. So I don't know. Command option space or something like that. It brings up an emoji picker. It doesn't work on Google. There's just a lot of little things like that, that like, and also like just pagination or if you turn off pagination, the width gets all weird and there's just lots of little stuff with it.

5:44that was honestly the main thing. And then the other big thing is copying and pasting. Like in Google docs, if you hit enter, it's like, there's no paragraph break. So if you paste into a CMS, a lot of times you have to manually go in and like add a lot of extra paragraphs or remove extra paragraphs that you made in the Google doc. Cause like the underlying sort of data model doesn't match. Anyway, I just want to fix those things kind of to start. It wasn't like supposed to be a startup, you know? Um, it was just like, maybe we could build our own thing. And it's kind of my fun software project because like nights and weekends, because it's, you know, I wanted to write code again and solve this problem I had.

6:17And then, but the more I started working on it, I did get curious about AI and it was kind of like, yeah, what could it do? We're not using it at all in our writing process or our editing process. It feels like someday we will. So it'd be fun to have like a playground to experiment with different ways that AI could maybe help the writing process. And so basically just built the kind of really simple thing, launched it with like the only AI feature and the main AI feature was like, if you type plus, plus, plus, then the AI will fill in the next paragraph for you. And the idea is like, when you're writing, oftentimes you get stuck, or at least I mean, I get stuck and I beat my head against the wall and I don't know what I want to say next.

6:57Very painful experience. You know, usually I would just like go to Twitter or something. I don't know. It's like a, not, not a fun place to be, but if you have this other thing to do, then maybe it helps you kind of streamline the writing process where you're not like using a lot of stuff that the AI actually says, but it just kind of gets the wheels turning and it gives you something to do to help you get unstuck. So that was what I really launched Lex as, like within every, it's kind of like this side project that was just like a fun AI experiment. And the response to it was like dramatically different than I ever would have imagined.

7:27It was like, people were like, oh my God, I can never imagine like writing the same way again or whatever. Or like this, literally one person said it feels like an iPhone moment. And I knew that the thing that was the iPhone moment about it was not my little demo. It was GPT-3, right? And so no delusions that it was like, cool, I've created some iPhone on my own nights and weekends, like little hack project. But the thing that I did do was like, kind of connect GPT-3's amazing capabilities to like a problem that people could relate to and experience and made for a really good, like, you know, 30 second YouTube video, or like within the first 30 seconds of the video, like you kind of understood, wow, that's, that's pretty cool.

8:12And you, in a simple way to just sign up and try it. Right. Um, and so we had like 25 ,000 signups just based on like a tweet, you know, um, and a YouTube video that the tweet links to within 24 hours. That was like just the first day. Uh, and then it kept going from there and it was just kind of like, wow, that's incredible. I accidentally have this like incredible wedge on my hands. I know it's not the thing, right? But just getting to 100K signups in a matter of a month or two is wild. And I'm just having more fun than I've ever had working on this thing. So I just want to keep doing it. And it really deserves to be talked about it with my co-founder, Dan.

8:52It felt obvious that this is a different company. It's not necessarily just a side project within this other company. and so we decided to spin it out um which is like you know a whole process went through that uh worked with our investors and dan to like make sure everybody was was as happy as they can be with like the deal and then um you know once the spin out was closed raised the seed round from true ventures and up until that point it was pretty much just me working on it and so it was great to use the seed round to kind of like hire a hire a small team and really the team formed kind of like last fall so in some ways i feel like wow i've been working on this almost two years.

9:29In other ways, I feel like the starting line was almost like the end of last year, because that's when the real nucleus of the team formed, you know? And the stuff that we've been doing in the last like six months has dramatically outpaced anything I ever did on my own. And we're just kind of, I think you'll see another wave of Lex, you know, later this summer and this fall, as we have some really big things that we're excited to launch that we're sort of developing and testing right now. But it's definitely kind of, our thinking has evolved a lot and we're really, really excited to, I don't know, just keep going with it because it's so much fun to build these things.

10:05And what's so amazing about this, again, like this is, like you said, this is all sort of, I think the story you just told, although correct me if I'm wrong, was kind of pre-ChatGPT, around the time of ChatGPT. Like what did that moment do for Lex? Like in what ways did that help Lex? and what ways did GBD4 help Lex? And obviously there have been so many things that have happened in AI since. Like I have to imagine that you've just had tailwind after tailwind. Yeah, definitely. I mean, there's tailwinds and headwinds, right? The tailwind is like the core products that we offer is still at its essence, a really simple yet nice version of Google Docs that has AI built in to help you with what you're writing.

10:50And so because we've got new versions of GPT coming out from OpenAI, we have Cloud coming out, new versions of Cloud, right? It's like the product just keeps getting better and better without us having to do anything. Of course, the downside of that is everybody else can access those same models too. So if we're a pretty thin layer on top of those things, then the benefit doesn't accrue to us so much as it does just to users and to the world because they have so many different options of things they could use. And then the other headwind is because I've never in my entire, like, you know, over 10 years in tech seen a hype wave this big, you know, there's like a million, everything from startups to side projects and everything in between, like building and kind of writing is like the most obvious first thing.

11:38now um you know i think that over time uh the stuff that we're building in lex uh is moving further and further away from kind of like whatever you might call a rapper or whatever but there's no doubt that the first version of lex was like you know it it was it became like the hello world demo kind of it's almost like hey we have this new thing called ajax and like maybe the hello world demo is like a to-do list that updates without having to refresh the page or something like that. Like, um, we kind of like built that and then it becomes the, like the hello world for a lot of things. Like even, um, you know, tip tap, which is like the open source text editing framework that, that Lex uses, like built in a lot of the AI stuff after we built it, um, to let any of their users use it.

12:24Right. So it's kind of like, uh, it's an interesting, like you've got to figure out, um, a specific vision of who you're serving and what problems they really have and beyond just sort of like, let's give people access to AI and kind of like a simple way. Um, and so that was like, in some ways it was hard, but also it was the most fun because, um, and the thing I think that makes Lex the most different is because it comes out of this thing where like, you know, I've been writing my whole career and I did it full time for like three years as a writer and as an editor. And I got a chance to work with a lot of really amazing editors.

12:58So I think the perspective that Lex has on what are the kind of problems that that writers specifically face not just anyone who's typing into a text box but like someone creating a blog post or a book or an academic article um you know all that what what problems do they have it's um i think we have a different um perspective on it than a lot of other companies you you mentioned you mentioned the term uh rapper or i don't know if you said gbt rapper or gbt rapper like that's that's obviously a term that a lot of people yeah a lot of people a lot of mention that like and I think it often gets used in a negative way I'm actually like not as convinced it's it's necessarily a negative I mean I think like it implies that whatever is wrapping it is thin but I think the benefit of it is it enables you to focus right and enables you to really focus on the details that the thing that it's wrapping GBT is is not going to focus and And then going to the complete opposite end of the spectrum, maybe something that's not a wrapper, like an incumbent product, like Google Docs, as an example, they're not going to be able to move fast.

14:09They're not going to be able to move fast and focus on those little attentions to detail that a company like Lex can. So I don't know. I think it's kind of interesting. And I'm curious. I'm sure people have asked you about this in the past. And I'm curious how you sort of respond to it. Yeah, I think that the term is a hilarious one and points out there's a kernel of truth in it, but there's a lot. I think that people use it in a way that is misguided. Um, and, um, I'm kind of, uh, I'm kind of, I'm kind of pumped that we were like, I think Lex was maybe part of where that term came from because we were like, you know, this thing that got a pretty insane amount of traction within like a very short period of time on a very thin, you know, value proposition on top of open AI.

14:58Right. Um, I don't think we're the only thing obviously that it refers to, but I think like part that term kind of came out a little bit immediately post Lex's crazy initial run there. And like, the thing that I think is true about it is if you are literally a thin wrapper, then yeah, there's not a lot of value. There's not a lot of defensibility. The thing that's false about it is any meaningful product might seem really thin or really simple, but actually involves a lot of work, A lot of work in understanding who your user is, what their problems are, how to speak to them, how to design something in a simple way.

15:37And even companies that seem like the most literal possible wrappers where they're just providing literally like a chat experience that's like an alternative to chat.openai or cloud.ai or whatever. Even those, I believe that there will be large businesses built that look pretty much like very similar to those things. And the reason why is, like, you can imagine there's companies who do it like, oh, but it's for a team context in a way that's like, in the long run, are OpenAI and Anthropic really going to focus on, like, all the different sub -variants of, like, just milking the chat interface?

16:15No, they're going to have to pick a lane. And it seems like they're picking the lane of consumer where they want to be, like, the top-level entry point, kind of like Google, you know, competing with perplexity on this front. But that kind of, that's a focus. And that means that you're not doing the like enterprise version of it, or you're not doing it as well, right? They'll take some steps in that direction, but it's really hard. Most companies don't end up being all things to all people. And the kind of the different contexts of value are like fractal. Like you think it's simple from far away.

16:45You're like, that's all basically this thing. But the secret of every great product person that I've ever met is they're extremely sensitive to small differences because they understand that in the market, those differences actually matter a lot. It matters a lot more than it might seem from a distance. And it's almost like the more you learn about a problem, the more detail you see, the bigger the differences appear to you, the more you understand that those differences matter. So when I think about something as different from a chat interface as like a writing experience where you're dealing with versions and you're dealing with comparing different versions and suggesting changes and all this kind of stuff, it's like, that's really, really different from chat.

17:18And all the stuff we're working on now is like, it's mostly just about how to collaborate on a piece of text that, you know, merits revision, that goes through a revision process and how to make those revision processes that happen within companies. Like when you're putting out a press release, you're putting out content marketing or an academic paper or an article on a news website or a newsletter, like you want to go through a sort of standard process, right? Developers have this with GitHub, with pull requests, with linters, with tests that run. Writers will have all this too, is just my really firm belief.

17:53they'll also have better interfaces to compare different versions of things. Cause right now the state of the art is kind of like suggest changes mode, you know? And I, that was one of the biggest things that like, if you ask any writer, they're like, it's a nightmare when you have a document with hundreds of suggested changes and like multiple people suggesting different things. It's just, everyone's fighting over the same space. It's horrible. So these are the kinds of problems that we're focused on. And like the more you zoom in, it's like, yeah, we're barely, it's like not even anything to do almost with, There are some chat parts of the interface, but basically anyway, to bring it back to the term GPT wrapper, just the closer you zoom into actual users and problems and building interfaces that solve those problems in the most elegant way, things start to become way more different than they might seem at first glance.

18:39And the differences matter a lot more. I actually wonder if we're now going to start to see, I know we saw like a wave in the beginning when, you know, like the beginning of sort of the AI hype. now that people have had these tools and these SDKs and these libraries for a year and a half, two years, like, are we actually going to start seeing a bunch more? Like, are we going to start to see many, many more applications, products at the application layer where the teams have sort of like figured out the best ways to leverage this stuff. And now they're starting to invent new experiences. So, um, I think it's like a really interesting time.

19:12And I think the benefit that Lex has is like, you've got this huge headstart, um, and you've been able to focus on, on these details and figure out what actually matters. And it sounds like collaboration is a big thing. Totally. And the thing you say about timing, I think is just so important. It takes time. It's so funny because people like, you know, even now, I think it's still, it's still early, but like a year ago, it was so, it was laughable that it's like, why aren't there any AI startups that have a moat? Like none of them have moats. It's like moats take like five years to build in a lot of cases.

19:44Exactly. And it's like, even companies, there are some cases of companies that start to develop very real modes much earlier in life. And typically what that looks like is a network. So if you look at the trajectory of an Instagram or something like that, then clearly they had some sort of very real mode much faster. And that is the interesting thing with AI is it's not very, it doesn't seem like there's new networks really being built on it. But for the category of companies like creative tools, right? Those things especially take time to build because it's a complex piece of software, you know?

20:21Like Dylan Field from Figma has recently said that like you shouldn't take as long as they did to launch. But like, I don't know, like could they have really launched sooner? Because like I think they were showing it to designers from a very early, I think they probably took a little while to get to the point where they decided that they were gonna focus on product designers. But once they made that decision, takes a long time to build all the things that product designers need just as table stakes features. And for a lot of other categories of creative tool, you have a similar problem. We certainly have a similar problem.

20:49Um, and you know, it's some things you just build it the way that it already works. Some things you, you want to innovate on and it takes, you know, some art to kind of choose those things carefully, but either way it takes, it takes time to build that and you can't just throw more software engineers at it and expect to come to a better solution faster, that you still have to have like a tight team with a high standard of craft to build something that writers actually want or any type of creative field wants to use. So I'm really excited for kind of, I think we're getting to the point now where some of these products are starting to mature a little bit.

21:24They've been able to go through enough feedback cycles and spend enough time to become as sort of feature rich as you need them to be to switch your primary tool. And now's a really interesting time. uh from also the big existing players are obviously they're they're building in more stuff so i i think over the next couple years is when we'll really move beyond like okay here's like cool prompts for chat gpt as like the main way that people use ai in their work or like you know okay engineers have co-pilot kind of experiences um to like a much broader class of creative adoption um so we'll see but it's i think you're right that the timing thing is so So people overestimate how much you can get done in a year or two and underestimate how much you can get done in like five years or 10 years.

22:12There's this narrative out right now about how, you know, AI is overvalued, like the technology and the features. There's so much hype. People love the idea of it, the concept of it. People are trialing every product in the world. But like, are they actually getting value out of the AI? What have you learned in the year, 18 months or whatever it's been since Lex has been out about the things that people actually like about the AI or maybe don't? And as you mentioned, it sounds like AI is actually becoming less of a focus. Yeah. Does that suggest that the AI is not as valuable to the users as you initially anticipated?

22:52um well it's funny because i had no real if we're talking about really initial anticipation like i did not anticipate how how many people would find it yeah you know interesting or whatever when we first that's true yeah like you said it was a surprise but i will say um a little bit as anticipated a lot of people they try it they feel like wow this is really cool and then there's some subset that they actually settle in with in the long run. And the most important thing is like being very specific about the problems you can solve for people and picking and making sure that that works really, really well, which is just as much an AI problem as it is an interface problem and a marketing problem and a design problem, all that kind of stuff.

23:37Um, and so for instance, like, um, you know, one of the features that I constantly use in Lex and wasn't a part of it at the beginning is you can select any line of text and leave a comment where you at mentioned AI and the AI knows the whole context of the document and it knows what line you've highlighted and it can give a response. So I use it for asking for feedback on specific parts. I use it for like brainstorming or research. Like if I'm writing, like, uh, I'm making some point, I'm like, is this, what's like some counter arguments to this point? Or, uh, what are some other examples of this phenomenon?

24:14It's just amazing for stuff like that, where it's just like this really specific in context kind of quick interaction that I want to have. Um, and then the other thing that we've learned is it's really good for in Lex for sort of like enforcing a guideline, a rubric, a style guide, you know? So if you, let's say, um, you know, you've got show notes for this podcast, right? Um, AI is not amazing yet at generating the show notes for you, right? If you just feed at the transcript and you say, here's some of my past show notes and here's the transcript, can you like figure out how I turn a transcript into show notes and like do it for me?

24:52AI is still not great at that. But what it is really great at is if you're like, hey, here's the like five key things I want my show notes to do. And here's some examples of it doing each of those things. Now, can you sort of like grade this draft against that and give feedback and give suggestions? It does a great job at that. It also, whether that's that sort of like a line level or like high level kind of feedback. And so we're building workflows for teams that's like, hey, here's a new edition of our newsletter. Here's a new edition of our changelog. Here's a new press release. Here's what our standards are for those things.

25:28And as a part of our workflow, can we have some of our editorial principles sort of like, let AI give them a first pass. Usually there's someone in a team who kind of plays the role of editor that like has the final say on what goes out. And often they're a bottleneck. And so if we can even help sort of make their job a little bit easier and help also for the writers on the team who, you know, you want to turn in something that's good, right, to your editor. So it's great to be able to like sort of self-edit with the help of AI. You learn a lot faster and it just feels very empowering. So it's kind of funny because it's totally opposite to this narrative of like, oh, like, you know, AI is just going to like put people out of work.