A very human vision for going all-in on AI

0:00Support for this show comes from MongoDB. You're a developer who wants to innovate. Instead, you're stuck fixing bottlenecks and fighting legacy code. MongoDB can help. It's a flexible, unified platform that's built for developers by developers. MongoDB is ACID compliant, enterprise ready, with the capabilities you need to ship AI apps fast. That's why so many of the Fortune 500 trust MongoDB with their most critical workloads. Ready to think outside rows and columns? Start building at mongodb.com slash build.

1:00of our two-part miniseries about how developers are using and building with AI in their products. Basically, I reached out to a couple of people who make apps that I really like and just said, can you come explain to me what AI does, what it doesn't do, and how you think it fits into this product. We had a lot of fun with Thomas Paul Mann from Raycast last week, and this week I'm talking to Sari Azut, who builds a tool called Sublime. And the way I would describe Sublime, you'll hear Sari describe it too, but the way I would describe it is as a bookmarking tool and then some, right? The idea is you collect stuff, you find a link that you like, or you highlight something in a Kindle book, or you journal into Sublime.

1:42And not only does it sort of put everything in one place and let you make collections and let you search stuff, it tries to use that to start a sort of discovery journey. So if you save a quote from an article that you really like, you open up that quote on Sublime's app, and it'll also show you this sort of endless feed of other interesting things that you might enjoy based on that quote. Sometimes it's tonally similar. Sometimes it's by a similar author. But the idea is it's sort of an explorer system built around the stuff that you save. I think that's very cool and really interesting. And I have actually really enjoyed using Sublime to just like dump things that I like.

2:21I read a lot for my job and just like putting in articles that I like and quotes that I find interesting and all that stuff in there and just sort of using it as a database to find even more cool stuff has been really powerful. At Sublime, they're using AI to make a lot of that stuff possible. And I think the thing that's so interesting to me about Sublime is that Sari, as you'll hear, is a very human-oriented person, right? Sublime is not some like whiz-bang new AI app. It is very like deliberately sort of quiet and relaxed, and it's not a high-tech feeling thing. If anything, it feels like sort of old internet in a way, but there's a lot of AI underneath all of that.

3:04And I wanted to talk to Sari about how to make those two things match. How do I have a place that feels private and mine and human and sort of interesting and curated and cared for when it's just, you know, chat GPT all the way down underneath? Is that disingenuous? Is it possible? Do those two things eventually run into each other and cause conflict. This is all the stuff that I've been thinking about with, frankly, every product I use. And I think lots of developers I talk to are thinking about this. How do you use AI in useful ways and not just sort of shout loudly about AI to make venture capitalists want to give you more money?

3:41There's a lot of that. That's fine. More power to the venture capitalists. But I think the best apps are the ones that are going to figure out how to strike this balance. And I think Sublime is actually doing it really well. So that's what we're going to talk about. Sari's always a good time. She and I have met a few times and I'm very excited about this. All of that is coming up in just a second. But first, I have to go and make sure that nothing in my Sublime library is actually going to get me in trouble in case anybody finds it after we do this episode. This is the Vergecast. We'll be right back.

4:11Support for this show comes from MongoDB. Are you tired of database limitations and architectures that break when you scale? Maybe it's time to think outside rows and columns. If you're a developer who wants to innovate, then it's time to check out MongoDB. MongoDB is a flexible, unified platform that's built for developers by developers. It is asset compliant, enterprise ready, and with the capabilities you need to ship AI apps fast. That's why so many of Fortune 500 trust MongoDB with their most critical workloads. Ready to think outside rows and columns? Start building at mongodb.com slash build.

4:54Sarah Azut, welcome to the Vergecast. Thank you, David, for having me. Excited to be here. Siri just went off when I said that, which is just a very funny thing to have happened. Happens all the time. I believe this. I think the place I want to start is I just want you to describe a little bit of the story of Sublime, because it's a thing I find really fascinating. You and I have talked about it a few times in the past, but for people who don't know, what is Sublime and kind of where did it come from? Yeah, I mean, Sublime is in a very concrete, utilitarian way. It's a personal knowledge management tool.

5:24I didn't start wanting to build a personal knowledge management tool. I wanted to start by building a different way of being on the internet. And what we've built really is a way to curate and build a personal library. The catch for Sublime is it's connected to other people's libraries. And I think that unlocks just a different experience on the internet where you are connected to other people. there is serendipity, but fundamentally, the way you're interacting with the tool is as a personal knowledge management tool and not some performance circus. One of the things I always liked about Sublime is it seems like, and I'm curious if this is your experience, there are sort of people who get that basic activity of like, just save things that mean something to you in sort of the broadest possible way, like things that make you feel something, put them in a place and trust that sort of something will come back about that.

6:18There are people who like instinctively get that and find that activity even sort of on its own really cathartic, just like the act of bookmarking does something for some people. And then there are other people who are like, what in the hell is the point of this? I'm just creating this massive, you know, text file of nonsense that I'm never going to look at again. And it really feels like you can sort the world into those two groups of people and there is nothing in between. I mean, 100 percent. That said, I would argue that it's been a niche thing because the tools that serve this need don't have a big ROI.

6:50Like, what's the point of storing this? And I think if you experience sublime, you save one thing and you go down a rabbit hole that is just like, you don't regret the minutes spent. It's just unlocking a different experience. I do think a big part of it is what's the point because the ROI of that in the past has been, I've got a massive archive of things that I will do nothing with. The timing that you've been building Sublime is like right next to this crazy AI boom. Did you think about this as sort of an AI adjacent product from the beginning? Because I think spiritually, it both is and isn't in ways that I really want to talk about.

7:30But like, were you thinking about what some of these language models might do in sort of the early days of building out this product? I think the question of what is your AI strategy or starting even with tech first is very foreign to me. Like even just companies or founders that tell me I'm building a Web3 startup. I'm building a B2B startup. That tells me nothing about whether, you know, like that's like a trend that is very easy to like mistake for a real market and a real product opportunity. So the way I think about it is, what is a problem that I want to solve? And in what ways are the current technologies available going to allow us to solve this in a better way?

8:11If you think about a tool like Sublime, there's, broadly speaking, three jobs to be done. We help people collect ideas. We help people connect ideas. And then we help people create with those ideas. So looking at those three functions. So on the collect front, I am very opinionated about this is taste-driven. This is judgment-driven. Only the user knows what deserves to be in their personal library. We have connections. You can plug in your Instagram bookmarks, Twitter bookmarks, Kindle highlights, all the things. But the act of just the judgment of what deserves to be here, we don't use AI. You could theoretically have some argument for somebody that's just monitoring what you're doing in your browser and deciding what's worth adding.

9:00We decided not to do that. Now, there are some things in the collect experience that we use AI for where there is utility and not judgment. So, for example, Podcast Magic. With just a screenshot, you can send the transcript of a part of a conversation that you like straight into your library. There is no, like you made the judgment call, but we use advanced AI to go from screenshot to a moment. Nobody needs to know that AI is a part of that process. It just is invisible, which is I think a big part of the way we approach AI is it's invisible. I remember I got feedback in the early days by like sophisticated tech investors saying we should say like this is AI driven and our landing page should have like, you know, AI powered connections.

9:44And, you know, there is an argument that the window of there is a window of opportunity that opened because people were more open to trying new things. But I think for the most part, once the shock and awe fades, people go back to the baseline and they just want their problem solved. So I think a big part of it is just play the long game. It doesn't matter what is powering this thing. It matters that it's solving something interesting for people. So that's the connect piece. So collect, connect. Then there's the create piece. The create piece is interesting because this is where users need a lot of control.

10:18So you might have a collection of really cool, maybe I'm collecting ideas for Sublime merch. I have got a collection of cool things that I've curated. One potential interesting thing is how do I combine all of these things into something new? These are my sources and references. If AI is a generalization machine, it's going to retreat to average. What can AI do with my sources? So ideally, it can recombine them into something new. Now, I think the problem is not a model problem. It's an interface problem where at some point, the sort of like AI-assisted iteration stops being great, And you need that like manual tweaking.

11:00And that's where I think chat sort of like ends. Like I've just tried so many times to tell ChatGPT, like change this in this image and it all just kind of breaks. So I think we've been very cautious on the create front because we've seen a lot of people just like panic to add AI features everywhere. And people don't like it if it doesn't work. So that's sort of where we're at with it is I think it is interesting to think about source grounded creation. But I think the quality is so subjective and users need a lot of control. And I think a lot of the disappointment with tools is the lack of control.

11:34Yeah. Yeah. I mean, I think I like that framing of it. And I think a thing you and I agree on is that that connect phase is actually something these models are already really good at. that like one example I give all the time is if you want to just do a useful, valuable thing with ChatGPT or any of these other tools, get movie recommendations. Tell it a bunch of movies you like and let it feed you other movies that you might like. It's like, it's low stakes in the sense that, you know, if it's wrong or it hallucinates a movie or whatever, like that's not the end of the world. But it is able to just sort of like open up a bunch of doors for you and be like, are any of these anything?

12:17and sometimes that's not what you want and sometimes that's exactly what you want and I think in a case like this one of the things that I immediately gravitated to about Sublime is that idea that okay not only am I going to save a bunch of stuff and I like build this library full of things that are meaningful to me but also it's going to help me extend that library because every time I save something I save a quote from an article I like as soon as I open that thing up it just gives me like an endless feed of other similar i guess is the word i would use things um but this goes to one of the things i'm thinking about is like as you're as you're tuning these systems and building them how do you even decide what success looks like right because again in the in the sort of old way of thinking about like movie recommendations it's like okay if i like this movie i probably like other comedies i like people i like movies by the same director i like movies with the same actors like we had we had parameters, right?

13:10And a generation ago Sublime would have been the same thing, right? Like, oh, I saved a GQ article. You probably like GQ. You probably like articles by this person or about this person or about this subject. You're trying to do something much fuzzier. And it seems like both a sort of opportunity and a challenge to figure out, like, what is that fuzzy thing that you're trying to do? It's a huge challenge. It's a huge challenge because it's not deterministic. Like, the software in the past is rules-based. This is, you know, this is every piece of content that you add to Sublime has an embedding.

13:42And it's just going to show you things that are like semantically in proximity to that embedding. And, you know, we try to tweak certain things here and there. I mean, there's this whole field of AI evals that is emerging. But I think your point is very clear. What I like about this in particular in the Sublime use case is that I think people think of hallucinations as a bug. And I think in this context, it's a feature. It's the fact that it's going to take you down a path that is completely unexpected that is interesting. But I think your point is clear. And in fact, it just changes software development because in the past, I've got my project management tool linear, and here's the ticket and do this.

14:31And it's like approved, done. And you could be in endless cycles of iterating. And in fact, when I test a lot of this stuff myself, we've got a database of millions and millions and millions of things. I can run tests on 10 things today. And by the way, the output for you is going to be different for me with the same card. So I think we're still learning, but it's a completely different era. Yeah, I mean, it doesn't seem like you have to develop a philosophy around what's good on that kind of thing. Because I assume one thing you have going for you is that everybody has saved all of this stuff into the database, right?

15:05Like by definition, your database is full of things that people have gone through the effort to save in one way or another. And that is like a everything, every single one of them has some signal of importance, right? Figuring out what that is and what that looks like much harder. But it is like you've sort of deleted all the crap from the internet just by virtue of what the tool is. But then like deciding what is good to put next to my stuff is so hard. Yeah, I mean, it's interesting. I learned a lot about recommendation systems from Ken Stanley, who wrote an incredible book. And something he was saying about social media and the way that algorithms work on social media that really resonated with me is if you are driving down a highway and you turn around and look at a car crash, the algorithm interprets that as, you know, you love looking at car crashes.

16:02And then it starts to show that to everyone. And so it's very based on, you know, if anyone has experienced going viral online, once you're past like a certain number of likes, it just keeps like cascading. The way that we think about recommendations on Sublime is we do have a human page rank of sorts, meaning you added this card or idea to Sublime and then I added it to my library. So there's some signal there that two people like this. What we are saying is past a certain threshold, which is very low, it could be three people. It doesn't matter if three people have added it or 100. It's just a basic threshold.

16:44And then there's just a different level of openness. So, you know, we have some philosophy around let's not just show the obvious stuff, you know, because that's everywhere on the Internet. And part of the philosophy with Sublime is how can we help more people find the non-obvious stuff on the Internet? So, you know, algorithms are, you know, like all of this stuff is like opinions embedded in pixels. So there are, of course, underlying opinions and decisions that we're making that are philosophical. But I think your point of how open-ended and subjective this is remains true. Yeah. And it does seem like going back to what you were saying about the collect idea, it makes giving people the tools to make that particular corpus of content in the way that you're describing becomes super, super important.

17:38And I think this is where I get into what is sort of the human-AI relationship in all of this. And I think a thing that you and I agree on is that the best thing you can do for your AI tool is just feed it good data that you care about. And it feels like we're in this place now of like, you know, I think you talk about the sort of collect, connect, create loop. And I think what a lot of AI companies want to do is just collapse that into like all sort of one thing where like you do a prompt and then the AI does the rest. And it's doing the collecting and the connecting and the creating for you.

18:10and that all of it is just sort of happening and collection and creation are the same thing. And that to me is like a bleak future that I kind of hate. But I love this idea of like, I am going to build a library and database of stuff that I care about. And then the tools job is to help me make meaning and something out of it. But it does start with like, it's a hell of a lot of work to make people make that corpus of stuff. Yeah, I mean, so a couple of things. I think I 100 % agree and believe that the value of dumping all sorts of context about any given thing in one place has gone up with AI.

18:53Whether it's coming up with a marketing plan, all of your research for a book, just centralizing all of that in one place that you can dump and then export easily, that has immense value. because AI grounded in your sources is just going to unlock immense value. I feel like Notebook LM was the thing that like turned a lot of people's brains onto this, that it's like, oh, just stick a bunch of PDFs in a place and then you can sort of talk to this tiny corpus of data about whatever you're working on. It was like one small version of this thing that made people go like, oh, when I start, when I tell you what matters and then we do something about it, like cool, cool things start to happen.

19:30A hundred percent. So this context question becomes more and more important. And I think we, you know, notebook alum is probably more about like sales and just the positioning is a little bit more enterprise-y. But I think similarly with creatives, I mean, some of my favorite use cases for Sublime could be something like I have a collection of phrases I love. And when I'm creatively stuck, I will just, we have a one-click button, I'll open ChachiPT with those sources and I'll be like, help me get unstucked, use like for this other thing I'm writing, using these references. And it's just fun. You know, I also think people take AI so seriously and a lot of it is about speed and volume, but some of my favorite use cases are just where it's not about speed.

20:23Like I would argue a lot of the places where it's helping me is, it's not necessarily faster, it's just better because I'm constantly iterating and spending more time with ideas because the cost of doing so is lower. Yeah. It does seem to me, though, that there is something, I don't know, sort of essentially pro-human about Sublime that like, and you and I have talked about this before and you write this news that everybody talk about it all the time that is like, this is not supposed to feel like mining a database of, you know, AI-generated highlights from magazine articles. Is it, is it, Is there a risk the further you go into, like, we're building these sort of cool AI tools that are going to help make sense of all of this, where, like, it stops feeling like this sort of human discovery process?

21:12Can the AI get too good at this over time? I think that in a world of AI, you can think of companies as automate the things people don't want to do or help people do more of the things that they choose to do, that they get to do, that they have more time to do. And I think that we are designing Sublime to be a joyful experience. And I think if you just keep your eye on the kinds of, the feeling that you want to orchestrate, you don't get, like, distracted. So I think part of, I think, building in this day and age is putting blinders on, you know, the way, like, racehorses are. You could look around and everyone's like, you know, all tools look the same.

21:56It's like a big prompt box and, you know, VZero and Bolt and lovable. And everything's the same. And I think it's just very easy to be part of that mimetic cycle. But the antithesis to that is just what is the feeling that you want to generate for your customers? And keep your eye on that. And the big prompt box is actually exactly why I brought this up, because to me, there's a really easy version of Sublime to build if you wanted to that says, OK, we have this giant database full of interesting, meaningful stuff to lots of people. And what we're going to do is we're going to build Sublime Bot and you're going to use it to query all the greatest quotes in history, right?

22:39Like tease this out long enough. And it's like, okay, we presumably have the 10 most relevant meaningful quotes on any subject you can think of. And we're going to start to, we're going to start to collate that for you. We'll, we'll make a wall calendar for you with meaningful, like there's just sort of a million ways you could go. if you start the process with prompt this database that we've built, right? But, and you can like sort of do that in the search of Sublime now, but like Sublime really doesn't want you to do it that way. It really wants you to start by saving something or making something and then go from there.

23:11And I feel like that at some level has to be a deliberate choice, right? To say this is not fundamentally like a thing you can go look at. It's a thing you do first. Yeah, so that was very intentional. I don't think that what you're describing is antithetical to the vision, meaning I can imagine a world where we've grown to a very, very large database where people might want to interact with that database very differently. Like you can imagine thinking, what's on your mind today? And instead of posting on Twitter, how many people actually post on Twitter? The vast majority of people are just consuming.

23:47So you start with a thought, an intention, something that's on your mind. And then you go down this rabbit hole of related ideas. I do think that intention over attention is one of my guiding principles. So instead of a for you feed, it's what do you want? I think there's something very fundamentally different about that that feels important to me. But I think there's a lot of ways in which we can extend the functionality that we already have and enable new kinds of experiences. and still stay true to that sort of feeling. It's nuanced, right? But there is something that gets very flat and boring about a world where everything is served to you in summaries.

24:33And there is, I think, like as with everything, there's an opportunity to be at the extremes. For us, it's never going to be about just the summary. It's here's some very interesting, cool substack by this person that you didn't know that relates to this other thing. And I don't think it's mutually exclusive with the other reality. I think we'll just live in a world where both coexist and serve different needs. All right, we got to take a break. And then we're going to come back with more from Sari Izout about Sublime. Be right back. Support for this show comes from MongoDB. Innovation is the key to success.

25:10But innovation also comes with its own set of challenges. And if you're a developer who wants to find a new way in and think outside of the box and think beyond legacy code, it's time to check out MongoDB. MongoDB is a flexible, unified platform that's built for developers by developers. MongoDB is ACID compliant, enterprise ready, with the capabilities you need to ship AI apps fast. That's why so many of the Fortune 500 trust MongoDB with their most critical workloads. Ready to think outside rows and columns? start building at mongodb.com slash build.

25:52So you mentioned podcast magic, and I want to talk about podcast magic because I think it, to me, it's sort of a perfect example of how to build, as far as I can tell, like an extremely AI driven, essentially like AI dependent product that doesn't, that still sort of hits the like human spirit of the thing you're talking about. This is pure like AI as tool, not as lifestyle decision, right? And I am like very bullish on AI as a tool and very bearish on it as a lifestyle decision, just to put my own cards on the table. Just tell me a little bit about podcast magic and kind of where this thing came from.

26:28Yeah. So podcast magic essentially is the simplest way to capture insights from podcasts. It came from both a personal need of sometimes I listen to David on the Vergecast saying something really smart. And I am constantly, you know, constantly. that happens all the time. And I am walking, working out with my kids out for a walk, whatever it is. And the only way I get to kind of retain that piece of information is when I'm back on my laptop or there's some like gymnastics I could do with other tools. But the normal thing people do, which is like the existing behaviors, often take a screenshot.

27:05It's such a thing that people do. In fact, when I tell everyone this idea, they're like, yeah, I have already a lot lot of screenshots of this stuff. Of course, they never go back to them. So I remember one day seeing somebody add a Spotify link to Sublime for a podcast. And in the notes, they said, minute 1114, Lenny is talking about, you know, whatever. And I posted it in our team chat. And I said, this is so silly. People want the equivalent of the Kindle highlight for podcasts, and it just doesn't exist. And Alex from our team said, this is probably a silly idea, but it should be a screenshot.