Ep 19: Tome CEO Keith Peiris on Disrupting Powerpoint and Scaling To Millions of Users

Jacob and Pat sit down with Tome Co-Founder and CEO Keith Peiris to discuss Tome’s go-to market strategy, deciphering through the “AI tourists” to identify their ideal customer profile (ICP), and the different hardware cost considerations when reaching enterprise scale.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Speaker 00

  Welcome to Unsupervised Learning, Red Points AI Podcast. I'm Patrick Chase. I'm joined with my co-host Jacob Effron. And today, we had a really interesting conversation with Keith Paris, the CEO and co-founder of Tome. He spent seven years at Facebook and Instagram before starting Tome, an AI-native multimodal storytelling company that uses generative image and text models to allow users to create anything from presentations to story outlines and much more. As of March, the company had 3 million users and they had raised over $80 million with their last round of funding, valuing the company at a $300 million valuation. We talked about how they took a first principles approach to AI-powered storytelling.

• 0:36Speaker 01

  It's going to feel really different if we think through all of that from first principles. And I think that's going to feel way better than bolting a chatbot onto the side of Google Slides or PowerPoint. And we decided that was a big idea, maybe worth the next 10 to 20 years of our lives.

• 0:51Speaker 00

  We discuss how Tome designed their conversational UI to best help users.

• 0:56Speaker 01

  We often talk about LLMs aren't intelligent enough, or they're not coming from the right data. But often, the biggest problem is that the customer hasn't given the LLM enough input for it to know what to do, right? So I think it is figuring out this sort of conversational UI that's going to make all of this stuff useful.

• 1:13Speaker 00

  And Keith discusses his vision for the future of Tome.

• 1:16Speaker 01

  I think Tome in 2035, we would certainly be working on like brain computing rather than a web app. But I think even before that, we just want to be the tool that you pick up to talk about complex ideas.

• 1:29Speaker 00

  Thanks so much for joining us. I hope you enjoyed the episode. Keith, thanks so much for coming on.

• 1:33Speaker 01

  Thanks, Jacob. Excited to be here.

• 1:35Speaker 00

  Awesome. Well, we always like to start with, would love to hear just kind of the founding story of Tome. Like, how did you guys come to start the company? And can you talk a bit about like the evolution? Obviously, you've lived through a few different hype cycles of generative AI, even in just the two, three years that you've been doing the company.

• 1:49Speaker 01

  Yeah, for sure. I guess to start off, you know, Henry, my co-founder and I, we've been working on communication products our entire adult lives. You know, he helped design the latest version of Facebook Messenger. And I worked on a little bit of, you know, humans talking to machines or some NLP research. But I think the most relevant Experience I had to tell was honestly at Instagram where I was running the AR camera team, you know, the stories camera. When we left Meta, we were talking about how much work we've all done as a society, as technologists and in helping people express themselves through the camera. Right. And when you think about what it was like to take a photo in 2005,

• 2:33Speaker 01

  you know, using like an early digital camera versus what you can do now with like the Snap, TikTok or Instagram camera. We've taken this thing that used to be sort of laborious and hard. You used to need like a team of cinematographers and like an SGI machine to do anything cool with video. And now you can take hundreds of photos and videos a day with your phone. You can express yourself using AR. You can, you know, drag text on with your fingers and produce something really sort of powerful and expressive. We were thinking about how, you know, social media drove that innovation. Because, um, you know, all these social media companies are run, you know, using display ads.

• 3:18Speaker 01

  So the whole company is incentivized to get people to share as much as possible so that they can put ads in between the feed and make money. But on the other hand, if, uh, if you want to share something in your head, you know, like an idea or a concept, you don't have the, this AR multimodal video camera that you can use your fingers. You have to use like Word and PowerPoint and, uh, and like maybe Twitter. The real problem isn't slide assembly. The real problem is that you're, you're sort of stuck at your desk for hours being like, how do I take all of this information in my head and linearize it so that, you know, my audience will be persuaded.

• 3:59Speaker 01

  So, you know, we were incubating this and we were starting to look at, you know, GPT-2. And we're like, this, this thing is not bad. You know, a long way to go, um, a lot of work to be done. It's a little slow. It's got, you know, all kinds of problems, but the essence of it is, is actually not bad at helping you, you know, figure out what to say and how to say it. And, um, I think Henry and I were both like, well, It seems like if we were going to build a product using this technology, we'd have to think about all of it. We'd have to think about like the editor, the interaction,

• 4:40Speaker 01

  how the page is constructed, how you would get something out of this like non-deterministic system. And I think it's going to feel really different if we think through all of that from first principles. And I think that's going to feel way better than, you know, bolting a chatbot onto the side of Google Slides or PowerPoint. And we decided that was a big idea and maybe worth, you know, the next 10 to 20 years of our lives.

• 5:05Jacob Effron?

  I love that. And I'm super interested, like as you were getting started, I'm sure, you know, you were probably thinking about all sorts of new ways these workflows could work. And I imagine there was some sort of tension between, let's just do something that looks totally different and is nothing like what a user has been exposed to versus, hey, most of our users are going to be coming from the Adobe Creative Suite, PowerPoint, Google Slides. And like, let's at least help them feel a little familiar or comfortable. Like, how did you think about that in the beginning in terms of how you designed the product?

• 5:31Speaker 01

  Yeah, it was a struggle, to be honest, which is to say that,

• 5:37Speaker 01

  you know, I think if you make it look nothing like a tool that anyone has ever used, I think it's just going to be hard to know what it's good for. But at the same time, like so much has changed since PowerPoint 1.0, right? Like PowerPoint 1.0 came from forethought in the 80s before Microsoft picked them up. And this was a format that was designed for 35 millimeter overhead projector slides, right? And because of that, it was, you know, designed for this four by three aspect ratio. It was static, you know, the cloud didn't exist. And it wasn't designed to be driven by.

• 6:19Speaker 01

  Any sort of models. So as a result, these sort of infinite X, Y, Z positioning. And it was it was sort of all designed for direct manipulation. So we're like, well, there's some things about that we have to challenge, you know, which is to say that most of us don't do work anymore just on a desktop screen. We're like looking at it on on mobile. We're looking at it on on different

• 6:45Speaker 01

  screen properties, like ones that you can touch. All of us were, at least the the folks that we we were dog fooding with, were working in the cloud. So we wanted to build something that was, you know, live and connected to to live data. We knew that it was going to be driven by models. And because of that, I think we wanted to build a sort of predictability into the page, right? Like, if you're going to be driving layouts, if you're going to be driving content being put in, it can't feel like I'm feeling lucky, like it's a different layout every time. So we did design a page system for it. And and we knew that NLP was going to be a sort of core

• 7:26Speaker 01

  or natural language was going to be a core way to people sort of interact with tomes. So we wanted to build with with with language commands in mind. So I'd say that's the technology piece. But then, you know, I think we had a lot of go to market considerations, right? Which is to say that it's a it's hard to create a slide go to market, you know, where we I think through some hard work and some amount of luck, you know, we were able to get a little further than than, you know, many of the the slide companies of the past. But it's it's a really hard go to market because if you try to embrace a consumer go to market, right? Most individuals and most small businesses,

• 8:11Speaker 01

  they don't create enough slides at a high enough frequency for you to spread and grow. So so you you you sort of hit the wall instantly if you're only doing like a founder's pitch deck that they make once a year, right? Or if you are trying to take You know, a top down enterprise go to market. You'll probably have to build for four to five years, you know, to build charts, diagrams, compliance, security, control. You know, before your first client will like even use your tool, right, which is inherently risky in itself. Right. So we were like, well, you know, let's pick our poison. Let's, let's take the consumer path.

• 8:55Speaker 01

  But in order to take the consumer path, we'll, we sort of need to embrace non slide prosumer use cases. You know, so we decided, okay, let's make a page system that we think can stretch to doing mini websites, to do visual docs, to do guides, to do homework, um, you know, and we'll sort of like, like grow top of funnel using this like a long tail of consumer use cases and then make sure that once someone likes Tome and they take it into their company, but it's a, it's a great slide product too. Um, so that was sort of the, the go to market side. And then maybe the last thing I'll say is, um,

• 9:35Speaker 01

  from the competitive perspective, you know, I think we, we believed really early on that there's, there's no startup to be built in AI slides. Right. If, if all we were doing was applying AI to slides, uh, then I mean, the incumbents will just get around to that faster. So I think because of that, uh, you know, we were always embracing, hey, we have to build our own format.

• 9:57Speaker 00

  I personally, I do think I'm feeling lucky for slides would be very cool, but, um, yeah, maybe that's coming later. The, uh, the, the question I had for you is, um, I, I think one thing that's top of mind, kind of for any founder that's building, um, maybe like an AI native SaaS app, for lack of a better, um, better term is, you know, how do users want to interface with AI? How does that show up in the product? Like, any learnings that you can share from how you've designed Tome? I'm sure a ton of thought went into designing it, you know, kind of to work with human and machine. Um, but I guess what have you learned about the, the workflows that, that people want and how they want to interface with AI?

• 10:36Speaker 01

  It's a good question. I, I think as a society, we've evolved. A lot, you know, and I think internally, we've, we've talked about the, the first phase of generative AI. Productivity was a little bit like a dancing bear, which is to say that, you know, we were all enamored by this idea that like, with one prompt, you could create an Oscar winning movie. You could create a pitch deck that would, you know, send directly to Redpoint and then your wire would show up the next day. And it was going to do your job so that you could go hang out in the park and, and, and live your best life. And I, I think because of that, we were like all enamored by demo wear.

• 11:17Speaker 01

  You know, we were all enamored by, look at this cool image, look at the scene I created, look at this eight page presentation. Uh, that, that I made and um, I think in reality, when, when we talk to our, our happiest customers that are in it, that are like using Tome for work, that they're not using AI to try to obviate their entire job, you know, they are using AI to help them shape something they already have, to help them sort of riff when they've got a little bit of writer's block or to help them start something. Um, you know, when they're like not quite sure how to, how to fit a story into a structure.

• 11:57Speaker 01

  So, so I think, um, you know, we've gone from AI is going to like do your job better than you to, to words. You know, AI is actually just like a reasonable person who's available 24/7, who can like, you know, iteratively riff with you, um, to help you get your work and your thoughts to a, to a really good place.

• 12:17Jacob Effron?

  So the iterative riffing, I feel like people that really know how to use these tools, it's unbelievable to watch them like go, right? And obviously, you know, a lot of people, you have so many people that come in and they just get exposed to Tome for the first time. You know, how have you thought about like teaching users or even making the product intuitive for them to get into that like riffing, you know, back and forth with the models that you guys have versus like just staring at the blank page or, or expecting things to kind of just work on a first prompt?

• 12:41Speaker 01

  Yeah, it's a, it's a good question. You know, we just released, um, Our new version of page generation a couple of weeks ago. And I think this sort of is moving in the direction of being useful at work. So now, if you open up to home, we'll default you to creating a single page instead of a whole presentation. And you know, we invite you to type anything. You could type dogs if you wanted, or you could type out, you know, you're trying to come up with your, your pricing strategy for your bottoms up SaaS tool. And now, instead of giving you one answer, we're going to give you four.

• 13:15Jacob Effron?

  Seen that approach before.

• 13:18Speaker 01

  We're like, here are four different approaches to think about this page. You know, maybe one is you're talking about sales assist. One is you're thinking about pricing and packaging. One is you're looking at like viral growth. And you know, from there, you could be like more like this or less like this. And we're, we're, we sort of like make it really obvious that these aren't done. You know, you should take one of these and keep going rather than, you know, here is sort of nine pages. And hit share when you're done. And you know, as we're getting into our new version of presentation generation, we're just sort of coming out in the next few months.

• 14:00Speaker 01

  It has a very similar feel like we're probably never going to show you like one prompt 10 pages that happens instantly. We're going to show you the outline, make sure that you sort of agree with the outline. You can edit it if you, if you don't, then we show you a page, see if you agree with the page, then ask you to regen. And, and sort of get into that flow. Actually, I talked to someone who was a management consultant for years. And, and she told me that's, you know, even if you work at one of these big management consulting firms, you, you write out your outline first, and then you go and show, you know, the, the partner or the, or the case manager. And you get them to approve the outline before you go build out more stuff.

• 14:44Speaker 01

  We often talk about L L M's aren't intelligent enough or they're not coming from the right data. But, but, but often the biggest problem is that the customer hasn't given the L L M enough input for it to know what to do. Right? So, so I think it is figuring out this sort of conversational UI that's going to make all of this stuff useful.

• 15:04Speaker 00

  That makes a lot of sense and Jacob is the consultant, so he would have.

• 15:07Jacob Effron?

  Yeah, I was getting some PTSD there.

• 15:09Speaker 00

  Yeah. It's kind of interesting because I, it makes sense, the iterative creation. But when you're talking about it at the beginning, it's like, hey, Tome can help you come up with this narrative. Or you're going to figure that out together and then go figure out the visual side of it. I honestly would have thought it would maybe be less involved at the beginning and more involved at the end. Like, hey, you have, once I have this narrative, then Tome goes and kind of makes it beautiful. But it seems like it's kind of showing up at the beginning all the way through the end. Is that kind of the right way to think about it? I guess it depends on how users are using the product.

• 15:42Speaker 01

  Yeah, I completely agree. And we try to be flexible, but we've found that we can help people the most sort of earlier in the process. And I think our sort of design team's goal has always been, you know, you should be able to show up to Tome with less than what you'd show up to a traditional productivity tool with. You should be able to be like, you know, I think I want to write this sort of product pitch about this idea. I've done the research already, but I'm not quite sure how to land it. And if we can get you to show up to Tome with that, that feels like a completely different plane of interaction than, you know, sort of the, what I would call a last mile productivity tool.

• 16:23Jacob Effron?

  You guys have had just this flood of like inbound, it sounds like people just coming to the website. Everyone, I mean, the, the presentation people share are super viral, you know, people are really excited about it. And I think a lot of AI founders right now are trying to think about how to channel that like initial excitement and people first getting exposed to the product into like long-term users that stick around. And I'm curious like what you guys have learned about that. And if there's like things that you're like, oh yeah, once someone does that in our product, like we know we've got them. And it's not just like, you know, hey, they tried that and tomorrow they try the next thing that's viral on Twitter.

• 16:54Speaker 01

  Yeah, it's a good question. I, you know, I would say we brought in a lot of AI tourists, you know, and I think, you know, it was awesome. You know, it sort of helped create explosive top of funnel for us. But I think the challenge with that was that we had to sort through all of the noise, right? Because before we launched Tome, when we were sort of in early access, we were only bringing on, you know, ICPs that we knew. Would be like great team idea shares. So we were bringing in startups, tech companies, PMs, you know, sales execs. And they were giving us really detailed feedback of I tried to hire Tome for this job. It let me down. Here's what I need.

• 17:38Speaker 01

  And now we've moved to, you know, we've like tipped 10,000 people in a high school in the Philippines. Right. And, you know, we're suddenly getting feedback of Tome is terrible for, you know, me trying to do my homework. Because it doesn't like cite the sources that the LLM uses. Where huh, is that the most important problem for us to go through right now?

• 18:01Jacob Effron?

  So I hope that's actually true, that there's 10,000 people in a high school in the Philippines using this.

• 18:06Speaker 01

  That is true, that they hit our workspace limit of 10,000. And one of the enterprise ready things we need to fix in the coming months. But to be honest, the first thing was that we decided to be a lot more discerning in, you know, whose feedback we're prioritizing. So we went sort of back to the personas who we thought were going to get the most out of Tome. And being really permissive of everyone using Tome, you know, we think anyone has a compelling story or anyone could have a compelling story to tell. But we are really focusing on a couple of personas whom we think can get the most out of Tome. So that's been really helpful. The metric that we use, and it's imperfect like all metrics,

• 18:49Speaker 01

  but we sort of look for someone wanting to share a Tome with someone other than themselves. You know, and we're like, that's sort of a sign that you've created something good enough for work. And, you know, at the start, there was a little bit of noise in this metric too, because it would be like, look, cool, I created an AI presentation. But I think once that stuff's decided, we were starting to be like, well, you just shared this Tome with somebody else. What were you trying to do? What were you hoping it would do? And sort of work backwards from there. And that's been really, really good.

• 19:29Speaker 01

  I think the other thing we found was to be honest, we weren't that excited about generating and self-serve revenue for the sake of it. But we thought it was really good for feedback qualification. That we were like. You know, we're going to set a really low price for now because the product isn't done. You know, we're not trying to be a giant self-serve company alone. We're going to charge eight bucks. If you care enough about this product, you could probably pay eight dollars for us to prioritize your feedback. And that's like really, really helps in figuring out who to focus on and what to do next.

• 20:05Speaker 00

  One question I had on the model front. I'd imagine there's a ton of that goes into the AI that you guys are using behind the scenes. And there's always this kind of trade-off, like build versus partner. Like, how have you guys thought about kind of curating your own models or partnering with the, you know, the model providers? And then also, like, any shortcomings or learnings from, from building there?

• 20:29Speaker 01

  Yeah, I would say we, we were very opinionated with wanting to own the customer experience, really understand the problems our customers want to try and solve, and then like get there in a V1 with off-the-shelf models. And I think to sort of be even more opinionated, I think as a technology company with a, with sort of a consumer go-to-market, you either need to be big or technology, or you need to have distribution. I think sort of everyone else gets squeezed out. You know, we, we decided, hey, we want to be like a, a distribution company that owns the relationship with customers.

• 21:07Speaker 01

  So to be honest, we were thinking about it like we're, we're going to build the race car with an engine that we purchased. And then when we're really happy with the aerodynamics and the chassis and we, you know, we've got like more data, we'll, we'll go build our own engine in the future. So, so we, we decided to build sort of V1 of all, using a bunch of off-the-shelf stuff. And then we started to, to build our, our own data pipeline. And, and this was maybe one of the most illuminating things we've ever done. You know, we, we built our AI team and we, we started to get this pipeline of, you know, a couple billion edits a month.

• 21:46Speaker 01

  But it's sort of wild to see, you know, we have all of this data of people taking generation, editing it, throwing it away. Some of it gets shared. But then we also have manual content that's being authored into them. And, you know, we, we started to realize Actually, this the stuff is pretty helpful. And it's all anonymized, I should, you know, state that. But we found that with this data pipeline, we can actually fine tune much smaller open source models and get way better output than some of the sort of large foundation models that are getting a lot of notoriety.

• 22:30Speaker 01

  And I think this has been awesome for a couple of reasons. One, smaller models work faster. And it turns out when you're building an experience that's supposed to be an extension of someone's minds, you know, the milliseconds matter. And then the other thing is, you know, we're starting to get a little bit sharper with the output. And, you know, we're able to sort of do things in a first class way that felt like a hack on some of the traditional models. So, for example, we, you know, we're trying to build a multimodal storytelling tool. And what that means is that sometimes, you know, you want a table, sometimes you want a diagram, sometimes you want a chart.

• 23:13Speaker 01

  This is kind of the stuff that we're working on right now. And we can create multimodality by getting the LLM to produce JavaScript to go and fetch that object for us from Tome. You know, and we found that, you know, you can get GPT to do that. But, you know, given that this is sort of a core problem to Tome, we decided to build that part ourselves. And now, because we built, you know, our front end in React, we can get the model to choose a layout. And then we can get the, then create JavaScript and sort of create a layout using classification that we've built into LLM. So I would say we're slowly seeping our way

• 23:58Speaker 01

  down into eventually building our own models. Super cool.

• 24:02Jacob Effron?

  It makes a ton of sense. And obviously, you sit on such an interesting set of data. And I think, obviously, you're able to kind of lower the inference costs, you know, when you build it yourself. And but I think the sequencing of doing that. After you have the distribution also makes a ton of sense because you have those like billions of data points on how people are actually using it to to build something differentiated. And I'm curious, like you really are, I think, just the scale of folks that are using Tome. You know, you're you're one of probably one of the most prominent examples of really deploying these, you know, large language models and image generation models at scale. You know, I think there's there's tons of things we hear all the time. Hey, there's tons of things that are cool demos. And then when you try to get them in production at scale, God knows a million things break. I'm curious on your end, like, what were some of those things that as you've tried to do this at just a really large scale

• 24:44Jacob Effron?

  that you've had to overcome or things you've had to build in-house to to really be able to do this?

• 24:48Speaker 01

  Yeah, for sure. I would say, you know, with with a lot of these these model companies and we're we're close with with many of the great ones. You know, they're so early in building, you know, an enterprise-grade product, but they haven't built out capacity monitoring. They certainly aren't working at 5.9 reliability. And there's a bunch of stuff that is like is supported almost by happenstance, but but not supported in a first-class way. But for example, if your LLM read a large corpus of material and it happened to have read enough Spanish, then, you know, it might be fine at generating some Spanish.

• 25:34Speaker 01

  You know, but but then we would find we move from model to model and we actually like make our generation quality worse. You know, in Spanish, even though none of us had even thought about it and the LLM had, the LLM team certainly haven't thought about it. So I think we've had to build a lot of monitoring and a lot of sort of eval work in in order to get things going. We actually rebuilt Tome's architecture with models so we could hot swap a model within a day. Just knowing that, you know, there's so much movement that you don't want to ultimately build to in too customized of a way for a single model.

• 26:16Speaker 01

  So that's been pretty good. We sort of built our own capacity management.

• 26:22Speaker 01

  Which is to say if we're like moving towards peak, we can slow down the free tier a little bit and prioritize, you know, uh, pro users.

• 26:32Speaker 01

  And I would say just all kinds of little things like that around really productionizing, building a, an app with, you know, tens of millions of users. That, that, that's sort of run on, on the foundation models.

• 26:44Speaker 01

  And then, you know, I, I think there's, there's all kinds of shortcomings that we've had to, to, to learn to overcome. You know, which is to say, you know, I'm sure everyone's talked about uh, LLM hallucination. And, uh, and, and we've, we've actually found the, the best way to combat hallucinations is to, um, actually just build a claim classifier. And, uh, so we built a claim classifier and then we, um, we, we send it through a web crawler. And, um, you know, soon we'll be able to replace those claims with things you can find on the web. And, uh, you know, one of the first, um, features we rolled out with the claim classifier is this references feature, which is in our text editor.

• 27:26Speaker 01

  So you can highlight a claim, hit reference, and it'll find a couple of links that, that you could cite. But we're going to use that to replace, um, you know, the text generation in the future. You know, I think with every image gen model, um, you, it sometimes gives you like broken text and, uh, broken diagrams and broken charts. And really you don't want any of those things. Uh, you'd, you want a real chart or a real diagram. So, um, you know, we're, we're, we're getting creative with negative prompting and, you know, uh, creating JSON to go and fetch charts and diagrams coming soon to tome. Uh, in the, in the next few months, I would say that the last thing is that we're, um, I think with some of the image gen stuff.

• 28:10Speaker 01

  Uh, I, I, I think our customers certainly complain about being in the uncanny valley with faces and hands and, and, and all of that. And, uh, we, we, now we're, we're sort of building a way to detect, uh, you know, is there a face, uh, in, in this generation? Does it sort of pass the threshold? Uh, if not, should we just use like, uh, image, like web image? API instead. So I think we're just getting more and more intelligent in figuring out, you know, what to use generative AI for versus, you know, what to use the internet for.

• 28:44Speaker 00

  Yeah, I feel like the faces and hands, they always, you know, like it's like such a perfect generation and then their hand is in a really strange spot or their their right and left are mixed up or something. But yeah, there's I can imagine some and it's really cool what you guys are doing with the JSON and like actually being able to fetch the item. Yeah, that's the super clever. My question, my follow up question was on the the model eval. And it's really interesting. You can kind of hot swap them and are running multiple in parallel. Can you say a little bit more about what the model eval looks like? And then are you guys kind of like, do you have a team internally that evaluates those or externally or that's doing kind of labeling and eval? But I think that's something that's really top of mind for founders.

• 29:27Speaker 01

  Yeah, so our, you know, our, our team in general has a principle of we have to like the stuff that we produce before we ask our users or before we ask an external eval team to to rate it. You know, I think one of the few things a startup can export that a big company can't is good taste. So, you know, I think we always have this bar. Well, do we like this? Then if we don't like this, then we shouldn't even bother. So we started with internal eval and we were looking at sort of concision, like how concise are is the output. Turns out, most people want their slides to be concise and terse. They don't

• 30:08Speaker 01

  want it to sound like sort of flowery AI output. You know, the other thing that we've that we sort of look for is, you know, precision and and sort of fact correctness. And then we look at, does the story add up? You know, does does page five look like it's connected to page six? And does the whole thing sort of add up? And and we found creative ways to evaluate that. I think where we're going from here is what we're going to be hiring a small team to do some labeling. But eventually, the place we want to get to is where we we just have our users tell us if they they like, you know, generation one or generation

• 30:52Speaker 01

  two or generation N. And Eventually, the place that we can get to is that we're all sort of consumer product engagement nerds. So we were like, we should just monitor time spends and we should monitor feature retention. And we should be able to sort of run A B tests at our scale of, you know, model one versus model two. And like which one has higher feature retention over seven days or 30 days and be able to sort of like move through with that.

• 31:23Jacob Effron?

  You were talking earlier about like, you know, about kind of, you know, alluding to this kind of like enterprise product. I mean, I believe it was in the context of a high school in the Philippines, which maybe isn't going to be your first enterprise. You know, that you go deep on that with, but I'm curious like how you've thought about, obviously, I think the sequencing of the business makes a ton of sense. That you wanted to start, kind of build a product consumers love and then eventually bring themselves into the enterprise. You know, how do you think about like what that enterprise product needs to look like? And what does that mean for kind of like the Tome roadmap going forward?

• 31:49Speaker 01

  For sure. I think one of the things that excites me the most about our enterprise product is that I think we'll be able to sell something really differentiated. And I think the sort of last gen of productivity tools, they, all they could sell was access control, right? So if all you could sell is access control, but then you almost like by necessity need an adversarial go to market, right? Where you're trying to get every single person at the company to use the tool. And then you could be like, look, Mr. CIO or Mrs. CIO, like 30% of your company uses blank, so you have to pay me. And, you know, I think that that's like a enterprise go to market. But I think we all wish that you could go through the front door and offer something valuable.

• 32:34Speaker 01

  I think the nice thing with Tome is that we, you know, our enterprise at Tome email is full of people asking for very simple things. You know, one, can I connect the generation to more data sources? You know, can I feed the generation my Notion, my Google Drive? Can I feed it like a list of terms that my business uses? And then, you know, I'm the generation pull from that. That's sort of second. Could we connect the generation to a single source of truth or facts like the company dashboard? You know, and if so, like, could you pull those fact in and cite it?

• 33:19Speaker 01

  And then ultimately, the place that we want to get to over the next couple of years is you should be able to sort of pull objects semantically into Tome from a wide variety of cloud services. Right? You should be able to say, like, make a page about our 30-day retention. And then Tome should recognize that, you know, we keep our data in Redshift and we look at it through Mode. It should be able to use its like Mode Sync to go and fetch the data and then render it in a native Tome way and then sort of have it sync, you know, whenever the author wants to refresh it.

• 33:59Speaker 01

  But you could imagine going one step above if you're connected to the analytics, if you're connected to like the expense reports, if you're connected to, you know, where the roadmap lies, you should be able to, you know, automatically create reports. You should be able to sort of build like through like multi-platform cloud reporting that'll all sort of happen through these stories that people need to tell at work anyways. So that's sort of where the enterprise product is going and, you know, I think we feel really lucky that there's a there's a whole lot to do there.

• 34:36Speaker 00

  Yeah, the integrations with the other cloud services, I think, is fascinating. Just because I think that's kind of always been something that was challenging with PowerPoint, I guess, is like, hey, you can screenshot this dashboard and put it in there, but like if you want it to update in real time or tell more of a story around it, then it's not always the best for that. And it was interesting around all this generative AI. There's been this wave of companies doing kind of generative video. And I was thinking about like, if you could couple that with a generative presentation that would be dynamically, you know, made on the fly, then you could have, you know, the CEO of a company like walking you through the presentation

• 35:15Speaker 00

  that was created in Tome and it would all be personalized. I don't know. It's like kind of a crazy future to think about. Yeah.

• 35:24Jacob Effron?

  I don't know if you see like the world heading in that direction or like what you think about how Tome could interface with like videos or avatars or any of that, but I think it's pretty interesting.

• 35:34Speaker 01

  Yeah, I think it's incredibly interesting. Yeah, I think one of the things that I learned about video when I was at Instagram was that I think whenever you need to make an emotive point, there's just sort of no replacement for video. So it feels a little out of place if you're like, you know, talking through a graph. But I think in the concept of, sorry, in the context of pitching, which is what a lot of people use, like a slide format for, there's really no replacement for video. So I think there's a wild amount of innovation happening and I think we're just excited to see where it could go.

• 36:08Jacob Effron?

  And obviously, as you talk about the enterprise product, I feel like a lot is, you know, tying other data sources into, you know, to kind of augment what you can do in Tome. And I'm curious, like, I feel like you could take two, you know, schools of thought around that. Like, one is you would go in and do some of the dirty work of actually getting some of that data into Tome yourself. Or you could say, hey, that's not really what we do. Like, there'll be other sets of folks that eventually the plumbing will get better here. And like, you know, there's all sorts of companies that are working on feeding this data eventually into models. Like, we'll, we'll, we'll kind of stay, stay back from that. Like, how do you think about where, like, how, how in the, in the blood and guts of these enterprises you want to get?

• 36:42Speaker 01

  Yeah, it's a, it's a good question. I think one of the challenges of being so early in, you know, this, this journey with AI is that it's hard to know what we should build, what we should wait on, and you know, what we should ignore, because it'll sort of be obvious in, in the coming months and years. And I think to your points, I think someone is going to figure out there's a data syncing layer with, with LLMs and, you know, we'd, we'd love to, to buy the solution from them. So, so we're really focused on right now, the, the sort of data syncing into Tome through the enterprise and the, the sort of core product of charts, diagrams, multimodal output.

• 37:25Speaker 01

  Really building robustness there because that's ultimately the stuff that differentiates our company. And we, we know there are, you know, many a startup. Figuring out enterprise data injection into L L Ns. And I, I know that we've, we've all talked about like hacking the context window. We've talked about embedding. We've talked about small finely tuned models and I'm sort of excited for, for someone else to take the lead on that stuff.

• 37:48Jacob Effron?

  I mean, another topic I know folks will be eager to get your thoughts on is like pricing. I mean, I think everyone's trying to figure out, you know, given inference costs that are incurred with different user actions. Like, you know, how do you think about tying pricing to that? And I'm sure, I mean, it's, you know, obviously you guys are still figuring exactly that, that out on your end. But curious like how you've thought about that. And, you know, maybe it's different approaches you guys have thought about.

• 38:09Speaker 01

  Yeah, I've been told by, by many great C E Os that I'm going to be working on pricing and packaging until the day I die. So, so it's hard to say that we've figured it out.

• 38:20Speaker 00

  Jacob's a consultant, so he, he has to ask about pricing. It's like in, in his D N A.

• 38:25Jacob Effron?

  When you leave a consulting firm, it's contractually obligated. If you ever host a podcast, you have to, to ask about it.

• 38:31Speaker 00

  Yeah, that's my second consulting joke about, about Jacob working. I'm going to see if I can hit three by the end. We'll see.

• 38:37Jacob Effron?

  I believe in you.

• 38:38Speaker 00

  Yeah.

• 38:38Speaker 01

  It's amazing. You know, I, I think to be honest, we, we had a very market centric approach to start, which is the, hey, let's be honest. We, we love the product we're building. We know it's early and because of that, you know, we, we'd love to have as many paid subscribers as possible using the product. And, and we know if we, if they love it and they ask for more, we can build more SKUs later. And, and because of that, we decided to, to start, you know, I think as low as we possibly could. And, uh, and sort of be, be open and permissive.

• 39:15Speaker 01

  The other thing I've found about, um, you know, all of these, uh, compute costs is that, that, that there's a bunch of downward pressure on them from, from a couple of factors. One, by the time you get to tomscale, you, you no longer buy inferences per token or action. You just buy, uh, G P U's, right? And, and at that point, all you're interested in are, um, peak or peak capacity and, and sort of, um,

• 39:44Speaker 01

  Air traffic control at times of peak. So I think that actually puts a bunch of downward pressure on your costs. The other thing that's happening is that the models are becoming smaller and more efficient. Right? I can't imagine a world where we're running Tome in thousands of enterprises using off the shelf GPT four or five or whatever it is, right? It doesn't seem plausible, so I imagine we're going to be using smaller models over time. That'll be cheaper to run and more efficient. And then the last piece is it sort of seems like energy is going to get cheaper, GPUs are going to get cheaper.

• 40:27Speaker 01

  So I think over like a long enough timeline, I think that this is going to look a little bit more like, you know, your regular cloud storage costs, rather than it being this big expensive thing, which is why, you know, I think we've had this perspective of, you know, we should make sure that our gross margins look fine, you know, but this isn't sort of the time to sort of like maximize profit when we're this early into the company's life cycle.

• 40:56Speaker 00

  Yeah, I totally agree. And it's been interesting to see just how much that cost has come down over time for both training. I think the training was down maybe 10x in a year and then inference also down 10x in a year. And you can imagine as that curve continues and there's kind of increasing pressure on the model providers, that that trend will continue. Moving towards the last few questions here. Would love to know, you know, I think you guys have you have an incredible vision for kind of Tome and, you know, a massive undertaking that you have gone on. Like, where do you want Tome to be in maybe five to 10 years? Or where do you see the company and kind of that longer time horizon?

• 41:37Speaker 01

  Yeah, I mean, I would say that I would want Tome to be a multi-format company. Yeah, which is to say that, you know, our sort of grand vision of Tome is that we're we're reducing the number of steps between you conceiving an idea, manifesting it in digital canvas, and having your audience understand it and be persuaded. All right, and I think, you know, Tome in 2035, we would certainly be working on like brain computing rather than a web app. But I but I think, you know, even before that, we just want to be, you know, the tool that you pick up to talk about complex ideas.

• 42:17Speaker 01

  And, you know, I think that means that we sort of have clear usage inside of an enterprise for pitching your CEO on your next big idea, or like, you know, pitching your PM on the next sub feature that you want to build. But but I would hope that, you know, we sort of maintain cultural relevance. You know, we're just to say that I would love to see climate activists using Tome to talk about big ideas, which are about Twitter, or like, energy discussions happening in Tome. And I think that that's sort of been the the the grand vision and also the challenge in communicating Tome early on.

• 43:00Speaker 01

  Just to say that, you know, when you start a company, everyone tells you to to go vertical SaaS, right? Pick like one persona, pick one use case, and then land and expand. But but in reality, you know, everyone needs, a lot of people need to tell a compelling story, and all of those people need text, they need video, they need charts, they need diagrams. And I think it's certainly a feature, not a bug, that Tome has this sort of horizontal broad appeal. And I wouldn't want us to sort of lose sight of the the prosumer business as we we figure out enterprise.

• 43:36Jacob Effron?

  Well, we always like to end our interviews with a with a quickfire round where we ask you a few questions and get your quick thoughts. And so maybe I'll kind of kick it off. What's your favorite thing that someone's built with Tome?

• 43:45Speaker 01

  It's a good question. I would say I've seen a couple. Someone published a children's book after generating

• 43:53Speaker 01

  And then printing it out and it's now on Amazon and Barnes and Noble. And then they have two. So the other one was that we found this tiny agency in LA. We like pitched Angelina Jolie to work with them using Tome. And they sent her a Tome and she opened it on her phone and she decided to work with them. So I think those are maybe my two favorites. Instead of me talking about some startup raising a series A or B, which just happened a few times.

• 44:27Speaker 00

  That's amazing. I guess next fire round question. What's something that's overhyped in AI today? And what's something that's underhyped?

• 44:34Speaker 01

  I think one shot generation is overhyped. Whether you're working in video, multi-modal, text. You know, I think that we've just overhyped that you type something into a prompt and you get an end result. I think it was like an easy media narrative, but I don't think we're going to work that way in the end. You know, I think what is underhyped is, you know, AI is a tool for human amplification. It's just you, but better. It just helps you get to a better result faster. And I think we need to talk more about that and celebrate those kinds of companies a little more.

• 45:11Speaker 00

  What's your favorite AI product outside of Tome?

• 45:13Speaker 01

  It's a good question. I am actually a big fan of perplexity. They're starting to break my habits of going to Google to ask hard questions. And I think it sort of shows the power of what an LLM can do.

• 45:25Speaker 00

  Yeah, perplexity is super impressive.

• 45:26Jacob Effron?

  Yeah, we both enjoy using it too. Well, I feel like there's a ton of threads that people are going to want to pull on here. So I will leave it to you to say, where can people go to learn more about you, Tome, all that good stuff?

• 45:38Speaker 01

  Yeah, I'd say if you're interested at all, try and Tome, go to tome.app. You can try it out for free and I'd love to hear what you're using it for. So you can find me on Twitter, I guess X, sorry. And send me a DM, would love to chat.

• 45:57Speaker 00

  Last thing, you were talking about this vision where, you know, vision for AI kind of. Maybe in the beginning, before people had played with it, where you know, you click a button, it generates a tome or generates a presentation that gets sent to Redpoint and then they automatically fund it. We only wire the money if it is a tome. We do six hours of diligence, at least if it's not a tome. But if it's a tome, we just wire immediately. Yeah, if it's a PDF or a PowerPoint, then it's extensive diligence. Otherwise, then you know, it's AI native and then we're all in. Thank you guys. New firm official policy. The best advocates.