Unlock AI Engineering, Venture Capital, & Open Source Innovation | Alessio Fanelli | Glasp Talk #28
This is the twenty-eighth session of Glasp Talk!
Glasp Talk delves deep into intimate interviews with luminaries from various fields, unraveling their genuine emotions, experiences, and the stories behind them.
Today's guest is Alessio Fanelli, partner and CTO at Decibel Partners and co-host of the leading AI engineering podcast and newsletter, Latent Space. Alessio has a rich background in software engineering, venture capital, and open-source development, playing an influential role in shaping the AI and engineering landscape. His contributions to both venture capital, with a focus on investing in technical founders, and the broader AI community through open-source projects, highlight his commitment to advancing AI-driven innovation.
In this interview, Alessio discusses his journey from playing with PlayStation in Italy to co-founding Latent Space and becoming a key player in venture capital. He shares insights into how AI and engineering transform the tech industry and advises emerging founders. Topics include Alessio's experiences building venture capital firms, the challenges of scaling AI startups, his take on the evolving role of AI in product development, and the future of AI engineering. Join us as Alessio shares his unique perspective on AI, venture capital, and the importance of open-source contributions in today's rapidly advancing tech ecosystem.
Read the summary
Glasp
Transcripts
Glasp: Hi everyone, welcome back to another episode of Glasp Talk. Today we are excited to have Alessio Farnelli with us, Alessio is a partner and CTO at Decibel Partners and the co-host of Latent Space, the number one AI engineering podcast and newsletter. With a background in software engineering, venture capital, and open-source development, he has been actively shaping the world of AI and engineering.
Alessio: Thank you.
Glasp: Alessio is also a key player in the venture capital space, focusing on investing in seed and series A companies. And besides his work in venture capital, he contributes to the community through open-source projects and by sharing his insight on AI engineering. And today we'll explore his journey in venture capital, his passion for AI and engineering, and his advice for emerging founders in the tech industry. Thank you for joining us, Alessio, today.
Alessio: Yeah, thank you for having me. I was wondering if you were going to mention that overlay Magic The Gathering based on my personal website. I was wondering if that was going to come up, but yeah, that was a very good, very good background. I'll reuse it for future bios.
Glasp: Oh, thank you, thank you so much. So yeah, I know you've been working at Decibel Partners and also you've been a co-host of Latent Space, but we wonder how did you become a partner at Decibel Partners and also started co-hosting Latent Space? Could you share a story behind it and also your background a little bit with us?
Alessio: Yeah, I'll try and do a speed run through it. I was born and raised in Rome, Italy. I started coding and kind of learning about computers, trying to hack PlayStations. So my family couldn't afford to buy minigames. So I was like, oh, you can jailbreak them and then download as many games as you want. So that sounds pretty interesting. And then after that, I started an open source company in university, which was a Google Nest alternative. Basically, we built our own hardware, our own software, and then we open sourced all of it. And then you could either build it yourself or kind of buy the finished product from us. So I got to learn a lot about open source and some of the incentives in that business, but also got my ass kicked in hardware. It's really hard to actually get things manufactured, as everybody would know if they weren't so naive to start a hardware company. And then I led engineering at a bunch of companies kind of like in YC, early stage. And I joined a venture firm called Six for Five Ventures about seven years ago. And I built one of the first data-driven sourcing platforms for VC. Basically, the history of venture capital is kind of network driven. So what funders do you know? Who do your friends know that are going to start a company and can you invest in them? That was really what venture was in the first 30, 40 years of the industry. And then about five, 10 years ago, there was, you know, the cost of building software kind of went down by a lot. So a lot of companies actually have products before they even raise a C round, which was not true in the past. You know, before the cloud, you had to like build, you know, a server rack and you had to like host it and you had to buy all the licenses. So like you had to raise money before you got started. Now you have a lot of companies that are public, that already have a product that had never raised money. So my idea, well, the firm's idea was like, Hey, why don't we look at things like Product Con, GitHub, Twitter, and find interesting products that are kind of like rising in popularity before they become venture-backed companies and invest in the C round of those companies. So our first one, the 645 was only 7 million. Then we raised a $40 million fund. We raised $160 million fund, and then we raised about 400 million for Fund4. And then at that time I moved to SF, actually in South Park, right where we met, to open the Bay Area office for the firm. And then eventually I moved to Decibel two years ago now, a little over two years ago. At Decibel, all we do is technical founders, building technical products. That's all we invest in. And what I kind of saw is that the firms got bigger and bigger. So the large firms now have like tens of billions of dollars under management. The middle ground is kind of like hundreds and hundreds of millions that try to do everything. But where we really saw the opportunity was taking a smallish fund. So like our current fund at the firm is 275 million, and invest that only in specific areas and only do a handful of investments every year so that you can put a lot of work behind them. So we have a community, for example, of about 2000 early adopters. These are like people in the enterprise that buy software. That kind of helps all of our companies with getting early customer feedback and kind of getting from zero to about five, 10 customers before they go on and raise a growth round. We have a full-time talent partner that can help our companies with hiring. If you're trying to, as you know, if you're trying to hire for any role across any type of companies, it's impossible to build a good network. So being focused kind of helps with that. And then we have four partners on the investment team. We do not have any associates. We do not have kind of any junior personnel. We do everything and then we use AI for the rest. And all of us were either founders before or were technical operators. So we can kind of relate to the founders in a different way compared to the average VC. So that's how I got to Decibel. And then the Latent Space story, it's a little random. So I knew Swix for a while before. So he has started this discord for developer tools investing for, you know, writing angel checks and whatnot. And we kind of got to know each other through that. And then I kind of wanted to start a podcast two years ago just to talk about, I just feel like a lot of people talk about startups, but they don't do startups. So I was like, maybe I should do a podcast and interview some of the people that actually know about it. And then Sean and I, we went to the OpenAI office in October of 2022. So this was pre-ChatGPT. And we were, there were made like 40 people. It was kind of like a hack day that Lenny Bogdanov organized. And Sam Allman was there and kind of like, you know, they were showing off the new DALL-E API. And they were basically kind of like, you know, there's a lot of cool stuff coming down the pipe. So we were like, well, Sean was going on all these podcasts. And I was like, do you want to do a podcast instead of just going on podcasts? And you know, I wanted to do a podcast anyway. So we're like, well, let's just do an AI podcast because nobody's really talking about this technology. So our first episode was with OpenAI. And we were like, hey, we should kind of have people know more about this company, you know, and what they're doing. Because I think people knew it as like, you know, Elon Musk's AI thing, you know, that was maybe like in the broader market, what people knew, but they didn't really understand the products. So yeah, that started, we released the first episode February of 2023. So that's been a year and a half. I don't know what the latest number are, but let's just check. We had about 1.75 million downloads or something like that. So obviously, a lot of people care about AI, it turns out. And it's been a lot of fun to do the pod and just get to talk to all these people before they become popular, because AI is so quick, you know, like we had the founder of Suno on the podcast, you know, many months ago. And now I even got my friends in Italy are like generating songs with Suno. And I'm like, how do you guys know about Suno? You know, like this thing is so new, but it just shows you how much some of these like multi-model AI things are like very international, you know, like text can be sometimes hard in different languages to really be good. But when you're making music, when you're making images, it really transcends some of the cultural barriers. So yeah, that's kind of the story and the background, but happy to dive into any of it.
Glasp: Thank you. That's really amazing and background. And so and yeah, I write, by the way, I really like, you know, the Suno generated AI music in the beginning of the podcast and also AI.
Alessio: It's very polarizing. So some people say they love it. Some people are like, you got to stop with this thing. So but anything, anything, anything that makes people talk is good, you know, so we keep we keep doing it. And most people like it, I would say.
Glasp: So yeah, I like it. Yeah. Great. And thank you. And so I'm curious, you know, you've, you know, interviewed a bunch of people like in from AI and listen to the ultimate, you know, guide of prompting, right? And so how do you find these people, reach out to people or, you know, how do you decide people to interview with? You know, that's something that maybe Swix and I don't even fully agree on, you know. Our biggest point of contention is like, should we try and interview Mark Zuckerberg or not? And I say we shouldn't, because I feel like everybody has already interviewed Mark and like he said, everything he needs to say and Sean is like, well, maybe we can ask better questions. So usually when you think about guests, there's like two things, right? One is how many times has this person already talked about what they're doing or what they want to do? If they've done a lot of interviews, it's kind of hard to get them to say something that new, you know, like. They're just not that many tokens that a person can generate. So that's one. The other one is, can you get somebody that has not been interviewed that much, but is still interesting, you know, because there's, there's also a reason why the same people get interviewed all the time, right? Definitely the most interesting things to say. So it's always a balance of who is kind of up and coming, you know, and what are like things that signal that. That's why we got Mikey from Suno on the podcast, for example, it's like a car potty and like tweeted one of the songs he generated through one of the early models that Suno had. And so we're like, Oh, this is the music was not great per se was funny, but we're like, you know, this works. It's going to become definitely like something that people really like. So we reached out to Mikey and then on the podcast, he did like a demo of like the V2 model and now I think they're on like V3, but for us, it was basically, can you see this product being interesting in the future? And can you see the story mattering? You know, like nobody had interviewed Mikey and like, they were previously working on like financial services data. So it's like, why were you working on financial services and now you're making AI generated music? Like there has to be something there that like, there's some interesting story to how you got there. So that's, that's one way we think about it. So that's more of the up, up and coming thing. And then the flip side of it is like, what are the companies that are maybe already popular in our circles, but that are maybe not approachable for many other interviewers, you know, like Bush on and I are technical. So for example, we interview, you know, George Hotz who interviewed Chris Lanner from Modular. Like the work that they do. It's so technical in nature that like, if you're not really technical, it's hard to get a good interview out of it. And that's why they don't like, you know, you'll never, not never, but like, you might not see one of them on like CNBC or like a big kind of like publication, because it's like the really interesting stuff, nobody asked them, you know, like nobody's asking Chris Lanner about garbage collection in Python, like everybody's asking them about, Oh, you raise a hundred million, what are you going to do with it? And it's like, that's not that interesting, you know? So, uh, that's kind of like interviewer arbitrage, you know, what are like, who are like people that don't get the interviews that they deserve, so to speak, and we can like put together better questions and content for them. Um, so that, that's kind of what goes into it. And sometimes we're also like, you know, in AI, there's a lot of ads. So just like, Oh, this tool is like the hottest thing. And then it disappears. Um, that's one thing that so far we've been pretty good at avoiding, you know? So we try not to like rush and interview some startup or like some founder, like if we're not sure that they're like for real, you know, because then you also look bad, it's like, can't believe you had an interview with like this guy. And it was like a fraud, uh, take like the reflection 70 B thing, the drama about like that, the model, like there's a lot of drama. We don't want to be in the drama business. Like our goal is to like make a podcast that you listen to it. And then the next day at work, you're going to like use something from it to help you be better at your job, you know? So, um, it's a bit of a different format, but so far so good. We've done 89 episodes, I think so far. So we're approaching number 100. Um, and, uh, yeah, we, we still got a lot to learn. So more, more to come.
Glasp: Yeah, that's exciting. Yeah, that's awesome. By the way, so in that sense, you know, like, do you take, do you accept the paid guest? I mean, sometimes I think, you know, you know, a lot of people in an interview and I think some people want to get featured, you know, by you guys.
Alessio: Yeah, we get a lot of emails, um, of people that want to come on the pod. We'd never charge anybody to be on the pod and we would never get money from anybody to put them on the pod just because like, it's not, it's really not worth it. Um, and same with ads, like we don't have a single ad, uh, there, there's no advertising of any kind just because you cannot do, Hey, welcome to the episode sponsored by, you know, NVIDIA. And then in the, in the episode, you're like, Oh, I wish AMD would be better. You can't say that. So, uh, we tried, we tried to avoid any commercial interests. Also, if you pay, if somebody pays you to be on the podcast, you cannot ask them hard questions. You know, you need, you need to make them look good. So that doesn't, that doesn't work.
Glasp: And you don't give money to guests or podcasters, you know, interviewees.
Alessio: No, we don't have it. We don't have any money. We, uh, we have some folks that we have busy, like a, uh, Patreon like set up for like, you can support the show if you want, but, uh, we lose a lot of money on the podcast every year because, you know, between like the, all the events that we do and like the editing and like all these things, it's like, it's definitely not a cashflow positive thing. Um, and usually if somebody wants to get paid to be on a podcast, it's, it's not a good signal, you know? So, uh, yeah. If anybody asks for money to be on Glasp Talk, do not pay for it. Just go interview somebody else.
Glasp: Yeah, we don't want to, yeah. Also like, you know, good idea should be, you know, you know, we're spreading, so, and we want to do that and share it with the world. So it's part of our mission as well. Yeah. But so just curious, like, you know, how many hours or how do you prepare for the podcast interview? Um, yeah, in general. So how do you prepare and how long does it take?
Alessio: Um, it really depends. So some episodes we interview somebody that maybe we're familiar with, but we don't know very well. In that case, it takes us a good amount of time. I would say, I don't know, 10, 20 hours or so. A lot of times also like its new projects. So there's not that much to research, you know? So it's kind of like, you're trying to understand about the market broadly, but like, you're more trying to create the content that people will research later more than studying the research. Um, the ones that take, then there's like the episodes where we know the person really well. So like, uh, when we did, uh, Lang chain with Harrison, or we did, uh, Jerry from Lama Index, it's like, we know those guys and like, we know their products and like, we're pretty well versed on it. So for those episodes, there's like less prep on research and there's more prep on like, again, how do you make them say something interesting that they haven't? Because both of them speak at a lot of conferences. Both of them do a lot of tweeting. They do a lot of blogging. So you're trying to get them to say something that is kind of like unique. So for example, when Harrison came out of pod, I don't know if you remember, there was like this hacker news viral post, which is like, it's a Lang chain pointless. You know? So we asked them, it's like, it's Lang chain pointless. You know, it's like, nobody's asking that. Uh, and so we had a nice, nice discussion on it. So it's more about being, uh, up to speed on like the Zeitgeist, so to speak. Um, and then some of them like, uh, George Hotz or like Chris Lander, those take a lot of research on the technical side to really, uh, understand, you know, so those might take more like 30 hours, you know, 40 hours of, of prep. Um, but yeah, it's, uh, it varies a lot and I use a, you know, I've like Obsidian for taking notes and I use a Reader from Readwise to do highlights and then they all get sunk, sink back into, into Obsidian. So, um, a lot of times I really have notes from the months of reading content before, and then I just try and put them together.
Glasp: I see. Oh, by the way, I was working on something like Obsidian integration or something, and I saw your commit. And so, oh, oh, this person shares a really great insight and, and later find out, oh, this was you when you have, yeah, yeah, I try all of this stuff. I open source all of it. I, first I, I was using Pocket and so I built like a Pocket to Obsidian, uh, like importer, uh, but then I switched from Pocket to, to Readwise. So then I turned that into a more open source, kind of like knowledge management sinker thing. Uh, but now Obsidian is a lot of good plugins. So I don't have to write a lot of code for it myself anymore. So I write code for other stuff instead.
Glasp: I see. Yeah, that's amazing. And also I saw your small podcast and, you know, to automate, transcribe and, you know, showing people. That's amazing. You know, I love to see you're contributing to, I love the open source, uh, culture and yeah, people sharing, contributing, amazing.
Alessio: Yeah. Small podcaster is open source. Um, thank you to my only contributor, Nathan, uh, Nathan Lambert from the Allen Institute. Some of you all might know him as the hourly chef, uh, God. Uh, he also writes, uh, interconnects AI, but, he's been super helpful in adding some, some features there. Uh, I'm actually doing a hosted version of it. Um, so more people will be able to try, even if they cannot run their own, their own code. I'll probably need to do like, you can do one episode a week, max, because otherwise I'll run out of money. But, um, yeah, that's almost, almost got 300 stars on, on GitHub. So people, people like the podcast.
Glasp: Yes. That's, that's awesome. And also you, you, you, you know, since you mentioned about Obsidian and tools you use, and I'm curious, you know, since, you know, we are also, you know, uh, part of productivity tools and what. no like productivity tools, not taking a writing tool, knowledge management tools do you use daily basis for work or personal project OSS?
Alessio: So I use a lot of things. I would say Obsidian is definitely my kind of like daily note-taking thing. I use for writing it's funny I mostly just try and write wherever I can write you know so I don't have a very good it's like sometimes I write in Obsidian just because it's easy if it's gonna have a lot of images I just write in Google Docs because then it's easier to get the images to Substack or whatever else. Sometimes I just write in Substack directly so I'm not very opinionated about the writing itself. Yeah I use Readwise for kind of like annotating things and things like that. I think like I should definitely like try Glasp again because I feel like the first time we talked I talked to you about I forget there was like something about the mediums I feel like at the time you were like more focused on like YouTube highlights and kind of like note-taking from like video content which I never do. That's one thing that's funny I use I don't really use YouTube that often for like research things you know. What I do I just take the video and then I use I have a bunch of tools that I built to just like extract the whole thing and then I put it in Obsidian and kind of go through it in Obsidian you know so it's like a different a different workflow and then I have a lot of tools that are built for myself so I have like an AI email client that I built that is basically like a it you can only respond with AI like you cannot use it to write emails you can only use it to like triage it with AI so it's all like pre-built buttons so it's kind of like a schedule you know so if somebody wants to schedule a meeting I just click schedule and then it writes the schedule email looking at my calendar if it's like you know a founder you know like a there's kind of like a not interested button you know and then there's like a schedule meeting button. I found the you know the easy triage things are usually like the most ROI on time that you get because as you know it's like if you get 50 emails like five of them you have to really think about and write something very specific like 45 of them are like either you just want to get rid of them or like there's like some very basic response that you want to write in it so that that's been that's been super helpful small podcaster helps a lot on like the podcast side what else we have a lot of tools internally for like finding companies so and also build this kind of like research agent so whenever I'm looking at a new startup I can ask the agent to do just like a basic research on similar companies and like summarize what they do and just look at all the website and find interesting things out of it yeah I think today with like cursor like some of these tools it's just so easy to like just create something that is like pretty simple that only works for you but it's like it only needs it doesn't need to turn into a startup you know like it's just something I feel like the email client it's pretty good like you wouldn't be able you cannot compete with superhuman you know but it's better than superhuman for me specifically so I think that's like my big wish just for like AI encoding especially you know to just like allow everybody to build these things like even obsidian plugins or like whatever else it's like now you can like pretty easily build your own plugin even if you're not somebody that codes so I think that kind of changes a lot on how things work so it feels it's more about data extraction you know like I would guess that in the future like the most important thing that you'll build even in Glasp it's like the ability for people to have like a central repository of highlights you know like it's not about it's not even about acting on the highlights it's more about just this is how I get them all in one place and then I build my custom way to sort through them and like learn from it and like whatnot so I'm just curious about AI UX overall so yeah I'm curious to hear what you guys think like you know not that people can like customize UIs and like kind of things can be regenerated in time like are you thinking differently about how you want to build like a product?
Glasp: Yeah as you mentioned AI UX is interesting space and also you know we want to have a like a customizability I mean custom like a capability the users can customize the UIs and because people have different tastes and also preference right and so just you know we can be as you mentioned like a centralized place for highlights or something and notes so that users can use the data as they need as they want some people want to use it or show somehow in obsidian or notion somewhere and yeah user UI can be really flexible I think in the future should be yeah totally agree with that.
Alessio: Yeah yeah also it would be nice if we can access more like you know open source like in a community-based like flagging in the future for example so but we don't have any basic or fundamentals about it but so in the future yes you mentioned that so you are using obsidian and thank you for mentioning about class but so where do you where is that like you know source of information so you don't want YouTube for research but so you know what kind of podcast or newsletter Twitter do you follow and you know?
Alessio: Um that's a good question who do I follow I feel like I don't really follow anybody but I read a lot of things often you know like I guess I'm like very um kind of like referral driven you know like we have the latent space discord I would say most of the things that I read it's like something that somebody has posted there and like likely swicks is like the number one content aggregator in the world so um I get to uh I get to get all the best links from him every day on every different uh thing um but I'm sometimes slow to lead them to read them so I'll put them like in my reading queue um and I think that's also just like often it's like if two three days pass by then you realize you didn't really care about something you know so I just try to not get over um influence but like what people are talking about and make that what I should think about because I'm also not trying to then tweet about the thing at the same time you know I think it's just a different um I don't know I don't have a lot of FOMO usually I'm like oh I need to know about the thing that everybody's talking about I just try and focus on like you know what's the thing that like I care about that would actually be useful to me and usually that thing if I read it today or like in a day or two it's still gonna be interesting so it's a good kind of like a test sometimes when you're unsure if you should care about something especially when it's like you know like a paper or like a lecture something like that it's like takes like an hour or something to get through it um podcasts yeah I would say I just have to spend so much time preparing my own podcast and like sometimes it's hard to listen to them I definitely listen to the acquired podcasts it's funny because like the fact that it's so story driven even though they're really long it's really easy to like cut them up you know so you might listen to like 20 minutes this morning when I'm going on a walk and then listen to 30 minutes later versus when it's like an interview you know you're gonna it's hard sometimes to stop it um yeah but but I would say social is kind of like the number one source of it and then decide what matters to read later because there's definitely not enough time to read everything that's that's for sure so you gotta you gotta choose somehow.
Glasp: Yeah I see yeah just curious you mentioned that reading lists but so where is your reading list like it's at Obsidian or like?
Alessio: Um yeah I would say I try and usually put things in the Readwise queue uh but also use like the Safari um kind of like reading list thing just because I usually read on my iPad so like it's easy to like sync between the between the two um yeah sometimes I'll just email it to myself um sometimes I will put I use like the tasks um extension in obsidian so I'll just put that as a task to do when I really want to read something you know I'll just put it there and then I need to get through it um yeah I don't really I don't really have a super super structured way to do it I feel like the most like you really only need the structure to like retain the learnings like figuring out what to read it does shouldn't be a structure you know it's almost like you know what you need to read at any point in time almost um so.
Glasp: I see, interesting. Also, you are the CTO at Decibel Right Partners, the VC, and you are the first person who I met, you know, the CTO at VC. And I'm very curious about the role, and you mentioned you internally, the VC, your firm has kind of automating and research agent and so on. But what do you do as a CTO at Decibel? Do you write code? Do you hire engineers, or do you also help in hiring?
Alessio: Yeah, 100% of the tools that we use, I wrote. So there's no external, there's no contractors, engineers, there's no, you know, there's no junior engineers and whatnot. That's usually a good forcing function. So in venture, there are not that many things that need to be built. You know, if you think about a venture firm, it's kind of like a repetitive process in a way. And you're not trying to reinvent the wheel every time. So usually what you're trying to do is build tools to help you be a little more productive at like certain tasks of the job. And it's really hard to know what those tools are if you're not an investor, you know. So if you're a software engineer, like when I started working in venture, the things that I thought were most helpful were not helpful to an investor, you know. So it's hard sometimes to delegate and it's hard to explain why a certain product should look a certain way, you know. Like for example, in our software, there are two different ways to ignore a company, you know. One way is like, I ignore this company forever, I never want to hear about it anymore. And then the other one is like, I don't care about it right now, but if something interesting happens, maybe I'll care again, you know, which is like a more open-ended way to ignore. And this is all for like the top of the funnel of companies that we have not met yet, you know. So we're just trying to get through as many as possible. So yeah, in our software, you can like surface companies. So we like to scan the whole web. Then we use a bunch of different LLMs to do a kind of pre-screening and kind of categorization. So we use 4.0 Mini for a lot of it, just because it's a lot of companies. Like every day you're looking at like, you know, 15, 20,000 projects between GitHub or like Twitter and like things like that. So you know, you cannot use a big model. And then we put all of those things in front of our partners. So the four of us kind of like spend time going through them. And it's a really very friendly UI to just quickly make decision. So again, it's almost like, what's the least amount of information that you're going to show? You know, like if you were starting in the venture, you would say, okay, how can I build the most comprehensive thing to look at a company, to make a decision, but like, actually the most important thing is like, what's the minimum amount of information that I need to consume to make an initial decision? You know, because you can actually make it pretty quickly, especially if you're focused. And then once you say, okay, this company actually looks like interesting. So you send it to like the shortlist, then we use bigger models to do kind of like more analysis and reasoning and kind of like some of the research work that I was talking about. So that's kind of on the company side. Then there's the other side of like the early adopter side. So we have this like big kind of database of like buyers and early adopters, and they change jobs often, you know, they kind of like move around, they do different things. So we use language models to like keep their records up to date, not on, you know, obviously you can automate the change of title with like the people data labs API or something like that. But for example, we also want to keep track of like what stack they use, for example. So using some of these tools to like extract that information, that's something that you can do from a, you know, job descriptions and things like that. So yeah, there's a lot. And I would say seven years ago when I started, there were maybe four, five firms like in the whole world that we were actually doing something interesting. I would say today it's not that many more. It's maybe like a couple, a dozen or a couple dozen. But I would say every firm, if you talk to any venture firm, they'll tell you they have some sort of software internally. Most of the time it's like Salesforce and they built a dashboard on top of it. But some people are also building software. But as you guys know, like productivity tools just have to have like the right ergonomics, you know. It's not really about the information. It's like, does it feel like, does it really fit with the type of work that you do? And like, does it understand the limitations of your work, you know, like it's really hard to buy an off the shelf product that works for every firm, you know, because everybody does really different things and thinks really differently. So yeah, that's a, that's kind of some of it, but yeah, I write all the code. So thank God for Cursor because if it wasn't for these AI models, I would be shipping like a tenth of what I ship today. But yeah, it's a, it's, it's fun. And it's hard to invest if you don't actually use the things, you know, so it's definitely an advantage in itself to do it.
Glasp: I see. That's very impressive. You write all the code. And by the way, so since you are technical, right, then you understand when you see startups, you understand technology, I think. So, in that sense, what do you look for in startups? What makes you think, oh, this is really interesting technology? I mean, because your bar should be high, right? Right. You are so technical. So the first bar is like, can I just build it myself in a week? If that's true, I'm not going to invest in the company. There are a lot of those companies.
Glasp: Even if they have a bunch of users and growing fast?
Alessio: Yeah. It doesn't matter because like, especially today, the bar, it's so like how, if you're building, say, I mean, now people don't do it anymore, but before it was like maybe 18 months ago, oh, we'll try and help you like do all of your marketing copy and do your sales emails. Those companies are growing extremely fast, you know? But if you understand the technology, you're like, well, everybody will be able to generate a marketing copy basically anywhere on the web in like 12 months, you know, because the models are just really good at it. So usually what I'm looking at is like, can the model do some, like, can the company do something that the model can't? That's kind of like the first thing like you need to build something that allows the LLM to do something that they weren't able to do before. That's either through doing kind of like more, you know, chain of thought and kind of end-to-end workflows. Now as you're seeing with O1, it's like, if all your company does is do a chain of thought, that's not enough to be a company. But usually, they're working in domains where the data is very weird, for example. So like we invested in a company called Brightwave, which does financial research agents, which was started by Matt Conover, who previously built Dotly at Databricks. And then Brandon, his co-founder, ran a derivatives training. And they were, you know, their idea was like, look, asking questions to PDFs is fun, but nobody makes money doing that, you know? So what they did is like, you know, what we do is like help hedge fund and investment banks and investment advisors do net new research on markets. So for example, what their product did is like, hey, we don't just tell you what happened. We tell you like what could happen what are like the risks and what are like the opportunities. So they just put together all these models that are fine-tuned to have different both skills and personalities. Like if you think about, if you read an S1, which is like a public filing for a company, 99% is the same for every company. It's all kind of like boilerplate things that they have to put in there. It's like, oh, in the future, we might make less money than today, blah, blah, blah. So if you're doing a semantic search, kind of like without just really, you know, thinking about what kind of content you're doing it on, you're going to get all these answers that are like not really relevant, you know? So you need to do a lot of the data processing to kind of handle that. We have another company, Dropzone, that does an AI agent for security analysis. So if you work in a big enterprise, there are like, there are these people there that are like stock analysts that review like thousands and thousands of thousands of alerts. It's basically like this email was flagged as phishing and they got to go in there and check if it was really phishing and like mark it. So what Dropzone built is like an agent that kind of automates a lot of these reviews. So now 90% of them, like the model can handle, and then the 10% that are like unbiased, you know, like the person can kind of step in for. And again, this is data that is like not really in the pre-training corpus, you know, because like the type of attacks change very often and like you basically have to look at data from different systems. A lot of times it's like the nomenclature is very obscure, you know, so you kind of have to build a lot of guardrails around it. You cannot really do it in ChatGPT, you know, way that it's like secure. You can pay it, paste an email on ChatGPT, and ask, is this phishing? But like, usually what you want to look at is for example, look back at like the domain that the email comes from and then go check when was it registered. Like, is it like a known domain? Like, has it been associated with other breaches? There's kind of like all these things that you need to do. So that that's like one of the core things I think about when investing in more like agent companies is like, what are like really, what's like the tooling that you've built around this LLM that the model themselves just don't have. And then the other side is more on the infrastructure layer. So we're investors in a company called E2B, which is like a code interpreter, API. I think in the last six months, they have like 6 million sandboxes being spun up. So they have like a ton of usage. And that goes back to like, what are things that seem easy that nobody wants to deal with? So, you know, cloud runtime management, like if I'm building a data analysis tool, I don't really want to think about, is it safe to run the code that the LLM generated, you know, or like, where should I run it? And like, if my customer is in Europe, where is it being run? Where's the data going? There's kind of like all these things. So that's like another type of investment. What are the things that are kind of like in the LLMOS, we call it like, what are the tools that the LLM needs, but that is not core to what the agent actually needs to do as business logic, you know? So yeah, again, it's really hard to answer this question because there's not, if there was an easy way to structure it, then you could automate it. You know, but I think like, it's really hard to like, like I always tell people if you were to automate a venture capitalist job, they would never make any investments because there's always a reason to not invest, you know, like, especially at the early stages. So a lot of it is just like the muscle memory, you know, I think by now in the last seven years, I don't even know how many founders and companies I met. It's like thousands and thousands and thousands of them. Like you just get a feel for like, there are like small things, you know, that are like tells of like, somebody's just BSing you or like, why did you say that? Exactly. Like Mike Mabels once told this story on a podcast, he was like meeting a founder and the founder was like, oh, by the way, I own the world's record for like the fastest Rubik's cube or something like that. And then they were like, okay. And then they went and looked it up and it was like, not true. And then they were like, why did you say that? You know, it's like, well, and so like, you know, that's like a small, thing, but like, maybe if you haven't thought about it, you wouldn't even go look it up, you're like, oh, okay, cool. But like, you're like, whenever founders try and say these things, you're always like, why did you say that? And why have people in the past told me that that's like an extreme example, you know, but sometimes you see it with like how founders represent how much technology they built, you know, and then you're like, but did you actually build that or are you just using, you know, uh, some API that does most of the work that you, that was like the early stages of AI. It's like, oh, we built like a, uh, email copy of writer. And it's like, is it just a GPT-3 API call? And they're like, yeah, but then we'll fine-tune our models, blah, blah, blah. It's like, will you though, how are you going to do that exactly? And then kind of like the castle starts to fall, you know? So, um, yeah, it's kind of like in the, you know, you have to be a detective a little bit, just try and figure out what's real and what's not, um, and things are not always perfect. Like every time you invest in a seed company, more likely than not, it's going to fail, um, but you always want to pick your battles, like for me, I never want to invest in a company where the risk is like the market doesn't care, you know? Like most of the things I invest in, it's like, if anything, we're not able to build it, you know, like it's too hard to actually do, but if you do build it, then everybody should use it because it's just like either so awful, like so much better than the existing alternative versus sometimes you're just hoping that the go-to-market works and like, that's just not, that's just not what, what I do. So yeah, that, that's a long story short to just say, I don't know how to answer that, but hopefully that gives you enough kind of like, uh, nuggets and in pieces of the thinking process.
Glasp: Yeah. And, by the way, in that sense, have you missed, is there any deal that you missed or you passed, but went great?
Alessio: Um, there's always a lot, right? There's always what I'm just trying to think what could be, well, there's actually, so sometimes you might think a company is good and then you're like, well, maybe that's too expensive or like that. That's something that I don't do anymore. It's like, if I like a company, I'm just going to invest unless it's like, you know, like some insane price, but it's like, if the difference is small, it's like never miss it. But yeah, there was like one company where, um, it was like YC company and I was like on WhatsApp with the founder. I'm like, Hey, um, by the way, can you send over like, uh, the pitch stack and whatnot, because like my team wants to see it. And he was like, I don't need to build a pitch deck because people are going to invest in me anyway. And I was like, okay. And I was like, you're probably right. You know, but, uh, and this was like in my previous firm, but it would just kind of like when you're building, when you're like in a bigger firm, there's like a structure to like investing, you have to go to a partner's meeting, blah, blah. Um, and so I was like, okay, like, we're just not going to be able to like invest because they want to raise it like 50 posts or something. Um, so I was like, sorry, you know, we just cannot do a round of that size at that price and like not even have a pitch deck. I think the company announced for like $4 billion or something like that. Um, so obviously it would have made a lot of money on it. Um, but yeah, the, the worst misses are the ones where like, you kind of like, kind of think that it's interesting, but like, don't do it, you know, because then you're like, man, I kind of had it, but sometimes, you know, there are like other companies where I was like, man, this is just like useless, you know, but then the, maybe it pivots and it does something else and then it becomes really valuable, you know? So, that's a thing. But again, if you think about the venture, like I only do every year, maybe like two investments. So, you know, I'm meeting a lot of companies and it's just not possible to invest in all of them. They're just not, it's just not how the business works. So what you just gotta be, you know, okay with, it's like missing a lot of these companies and just saying, Hey, you know, I just couldn't, couldn't do it. It still sucks, you know, but it's like, at least you're seeing them, right? Like the worst thing would be, you'd never missed any successful company. All that means is like, you'd never had the opportunity to even invest in a successful company. So you should go do something else with your life, you know? But, um, yeah, that's kind of how, how the job feels sometimes.
Glasp: I see. And also, you know, you know, I'm, I'm, I have a question regarding like a founder quality. You mentioned technical, uh, capability and also like ideas, right? But, you know, since AI is getting better and I do think idea matters or, or, you know, what else, you know, matters like when you see founders, you know, besides technical capability, but, you know, do they have, I don't know, something exceptional and yeah.
Alessio: Um, yeah, that's a good question. So I was actually in a meeting a couple of days ago and then, it was like some person from like a sovereign fund and they were like, how, how do you see like enterprise software evolving? Um, and what I told them is like, well, why should enterprise software exist at all when you can theoretically just like regenerate any UI at any time, you know, if you had a good enough model, you know, so, um, that's obviously like the extreme of it. And I don't think most people will do that, but, um, I think it's less about the technical skills and building maybe the interface, because I think the interface will be more malleable, you know? Um, I think it goes back to like the ergonomics, you know, like what are like, how do you have an idea and an insight into a market that is like unique, you know, like if you were like, Hey, uh, I'm not doing Glasp anymore. I'm going to do an app for like, um, uh, I don't know, airplane pilots. And I'm like, okay, why? Like, what do you know about airplane pilots? You know, it's like, it's easy to write the code, you know, but like how, what do they need, you know? So you always got to look at that. And even like the idea, it's like not, that's not yet like the idea and the insight are like two different things, you know? So the insight is like, Hey, I understand these people like myself in venture, like I understand how venture capitalists work better than like most people that are not venture capitalists, you know? So that's kind of like, I think these are like the two things that most people don't understand about it. And then the idea is actually like, okay, how do you act on it? So I think in the age of AI, it's like the insights are much more helpful than the ideas, because then if you have the insight, you can like kind of work your way around it until you find the right product shape for it. If you only have the idea for a product, but you don't know why that's helpful, then it's kind of hard to like, you know, so like iterate around it, you know, and maybe you can still make there. That's why I think kind of like the long tail of products is going to be very, like, uh, say the AI-generated photos, you know, it's like, is it a good idea? Probably because people are making money, but like, is there any unique insight that people have that makes it like a durable business? Like not really, you know, it's like, there's not, there's not a lot of Delta between all of them, you know? So I think like, if you're like somebody that wants to be a founder and start a company, just like, what are like, what's some unique insight that I have into the way people work and operate, or even it could be an insight about the future too. You know, it's like, there are going to be no more customer support people because they're all going to be replaced by AI agents. So now all of a sudden you have all of these people that are really good at interfacing with the public and like problem-solving, how can I build a platform to reuse that talent somewhere else? Like, that's like a simple example of like something that is not like a product idea per se, but it's like, this is kind of like the insight or like the trend that I want to build around and that's going to evolve, you know, but like, yeah, it makes it easier to write code, but it doesn't make it easier to like come out with like what the code should do, you know? And like the idea itself can only get you so far, you know, you might do a product that has some success, but it's like, you know, yeah, that was like the thing, it had generated photos where the thing 12 months ago, like today, I don't see that many people doing that anymore either, for example, it's like, it's kind of like a fad that comes and goes. So, yeah, it's hard, it's hard to come up with a good company.
Glasp: I say, yeah, yes, yes. But by the way, quick question, do you, how about design? Do you believe in the power of design or UI?
Alessio: Um, I think design for me, it's more like the UX than the UI, you know? So like, are you like, it might look pretty, but you're asking me maybe like do too much when you could just infer it, you know, or like, you know, yeah, I think that's why the insight is really important because you need to understand what the person wants to do. Like if you're like when you're doing software for like a doctor, there's kind of like the UI, which is what it looks like. And then it's like the UX of like, do you understand that most of the time I'm standing and I'm like walking and I'm like, I don't have a keyboard, I'm typing on a touch thing, you know, like to me, that's like more important to understand. Um, and then yet, you know, today the bar is like so high because you're going to assign it to use tailwind CSS and do an ICUI. So, uh, you know, the minimum viable design is much better than it was before. Uh, but I think if anything, you know, sometimes less is more, you know, just have a, just have a simple UI that has a good UX and like, just get out of the way of the, of the person that's why Obsidian is so good. It literally got nothing, you know, I just, you just write in it and, and go.
Glasp: Yeah. That makes sense. And since time is running out, but you know, I'm curious, you know, you have interviews, you said you have, uh, you have done 89 interviews and not only interviews, I know, you know, speaks and you just, you know, had a conversation, you know, several times, but you know, how do you see, since you are the expert of AI and do you know, you know, what's happening in AI in this space? So how do you see AI evolving in the next, let's say a few years, five years, 10 years? And what kind of role do you want to play in the future? I mean, do you want to keep, you know, do you want to keep becoming a venture capitalist side and support ecosystem? Or do you want to keep contributing as an open-source contributor or want to start your own company at some point?
Alessio: Yeah. You know, I would say even five years nowadays, it's like too, too hard to predict. Um, but what, what I do know is like the models are turning more and more, or at least the model labs are turning the models more and more into platforms, you know? So if you look at like, even just basic example, like, Oh, one, Oh, one does more than just generating tokens, like, like the outcome is more than just asking to generate the next token. It's kind of like, okay, my goal is to reason, you know, I'm expecting that more and more will be kind of like swallowed by the model. So like people, if they use one, they don't need to know about the chain of thought, you know? So now all of that part of the industry has kind of been swallowed back into the model. Um, now OpenAI also has the evals and kind of like the model distillation feature. Um, so I would say less and less, it's less and less interesting if you're a founder to like, try and build these small things around the model, you know? Because like the model labs just need OpenAI, OpenAI, if they need to be worth $160 billion, they need to do a lot more than charging $2 per million tokens on four or many, you know, like they need, they need like a lot of things to be built around this. So that's like the first thing, kind of like the model labs becoming more platforms, kind of like in more like the clouds, you know? Like if you look at AWS, it's got like a runtime, it's got databases, it's got authentication, it's got all these different things. Um, I think that will be, that will be similar. Um, and then, yeah, our main theme is like services, software. So, um, you know, usually people used to buy software to make people more productive. I think today in the future, people are going to buy AI to do jobs for them. Um, and that is not, I know a lot of people are always worried about kind of like job displacement and whatnot, but even you guys, like, if you think about most people that run a small business, they just, they could do more with more people, you know? So they're usually not, it's not like, Oh, I have all the people that I would ever need. Most people could do a lot more if they had more capability and bandwidth. So actually think this is going to help the long tail of businesses, you know, kind of like do, do more and hopefully more of like niche things also do better. You know, I feel like maybe the last, especially here in the U S the last 10, 15 years have been a lot about kind of like a centralization of like brand power. It definitely likes in retail and like stores and whatnot, but like even like, you know, like sports is a good example of like, it used to be about teams and that's about players, you know? It just kind of like a few people kind of gather all the, all the attention, you know? So I'm excited to see more and more kind of like, even, even, you know, we don't do it as a business, but even our podcast is like, we're able to do a lot with very little because of AI, you know, like if we didn't have AI, we wouldn't be able to do AI-generated songs. We wouldn't be able to do nice thumbnails. We wouldn't be able to do all the transcripts and like all the chapters and all of those things for each episode. So having that helped us be a more popular and bigger podcast. And I think the same thing will help to like a lot of other folks and publications and like things like that in, in different areas that like we might not even imagine, you know? Um, but yeah, that's, um, that's kind of how, how we've been thinking about it. It's like, people don't want to buy software anymore and they kind of want to hire an agent to do something for them. I think that kind of changes the way people build companies overall, to begin with. So, um, yeah, I wish I had, you know, like, I wish I had a better answer, but it's really like, you know, yeah. Anything, anything I see today, like in three months is already outdated. So I'm more like trying to figure out how to help my companies be relevant in three months than in five years, if that makes sense, you know, you're more trying to see what's like the next thing coming up. But it's a great time to be in tech overall, you know, like I always tell people, like if the models do become AGI or whatever, then it didn't really matter anyway. Right? Like, it's going to be great for everybody and like, whatever. And then if the models are not going to be AGI, then you kind of wish you had just started building, you know, and doing things. So either way you should be building and experimenting with these things and try and do something instead of just sitting around, uh, and don't try to do five-year prediction, just try and do shorter term, you know, iteration.
Glasp: Yeah, that probably makes sense. And, by the way, do you think, let's say OpenAI keep the same position in the future or other players, Google, let's say Google incumbents like Amazon take over?
Alessio: I would say at this point, it's funny, there's not really a lot of room left between where the models are today and the more kind of like, whatever, AGI-like things. I would say the models today are really good at email generation, document review, Q and A, and coding, all these like average enterprise tasks. And then I think the next step from here that is really meaningful is just doing all these tasks completely end-to-end without anybody being involved. And I think we're like a little further away from that, we're maybe like a couple of years. And it just starts to see who will win from there. Like if nobody's able to do it, then everybody's gonna catch back up, you know? But I don't know, Sam at the Dev Day, I was pretty adamant that like the models are gonna get extremely good in a very short amount of time. So maybe they stay ahead, but also he needs to raise another tens of billions of dollars, so he obviously needs to say that their models are gonna get very, like a lot better. So yeah, it's hard to say, but I definitely think it's more than one winner, you know? Like it's hard to see OpenAI being, you know, 10X better than Entropic, you know? Like they're just like too many, they're just like too many people are friends, you know? Like they're just like too much overlap amongst all these people. They keep switching between them, you know? And it's like, it's hard to really have like a big gap between them.
Glasp: Yeah, that makes sense, yeah. I saw executives or core members moving to another company or Google DeepMind and so on. Yeah, I see this transition every time.
Alessio: Right, exactly.
Glasp: Yeah, yeah. So, all right, so I read your one year anniversary blog, actually, then you listed some questions and I wanna ask this question to you before ending. So what's one message you want everyone to remember today?
Alessio: I would say my message would be to stop arguing on things that the outcome doesn't really make a difference for you one way or another, you know? Kind of like, especially in AI, people are like, oh, is AI real or not? It's like, are you asking because it makes an impact on your life or are you just asking to argue? So whenever there's somebody that wants to argue with you about, oh, AI is useless or like this is useless, it's like, if they don't have, if it doesn't really impact them in any way and they just wanna talk, I just try and avoid those conversations because it's like, it's not worth my time. Like there's nothing to gain from this. Like you're not incentivized to have a good opinion. And then whenever you do think that you have a different opinion than others, just try and capitalize on it, whatever way that is. Like going back to having an insight to build a company, if you do think you disagree with most people on something, just go do something that makes you better off by doing that. Don't try and convince the others that you're right. You know, like too many people are like, oh no, the opposite side is like the people saying, no, AI is not a fade. AI is really like it, it changes everything. AI is awesome. And then they don't do anything with it. It's like, that's also useless. It's like, don't argue on Twitter, just go build something that helps you, you know? So I would say that's the message I want everybody to listen either go build something or like don't waste your time talking about it. Like nobody has the time to read your messages.
Glasp: Great message. Be a builder, yeah. Cool. And then, sorry, this is the last question. Since Glasp is a platform where people can leave what they're leading, and learning as their digital legacy, we want to ask you this big question. And what legacy or impact do you want to leave behind for future generations?
Alessio: You know, there are like two, I would say maybe two things. One, for like the broader, you know, world, it's just like that it's okay to do more things in the open, you know? I know it's like, Sean is also a big proponent of kind of like learning in public. And I know you guys are too, because you're free tier. So you can use the product for free if you have like public highlights, but yeah, just try and be, you know, inspire people to do it. I think most of the time, especially in my business, like in venture capital, people are very secretive, you know? And like, they try to like, just make it look like they have some like big, unique thing that they do or whatever. So I think, you know, one thing would be just making people more open about sharing what they do and understanding that like, look, what you have is really not that unique and that you can still succeed, even if you share it. You know? I mean, to me, that's like the biggest thing. It's like, it's easy to share it. Having the example of like, I did it and I was still successful at what I did, like impacts it a lot more. And then the flip side, not the flip side, but like the other side of it is like, you know, I grew up in Italy, which is maybe one of them, we were like the first country to ban chat GPT in the whole world, you know? It's a country that is not very tech-forward. So trying to change that is one of the things that I hope, you know, that like we can try and make people say, hey, change is good for you and technology is good for you. It's hard. So I would say that would be much more of a legacy than learning in public because I think most people already kind of buy into that. But yeah, that was like the hardest question when you were like, hey, can we do this pod? So yeah, you can tell that's something that everybody should think more about, you know, what to do, what to do. Have you guys answered that question on the podcast before or like written about it online?
Glasp: Yeah, I answered, you know, several times because, you know, our project, I mean, Grass started kind of, the idea came from my kind of near-death experience. So when I was 20, I had a subdural hematoma. The left side of my body was paralyzed. At the time I was almost dead, and I was so scared, but before I died, before I passed away, I wanted to leave something that like for other people so that I could feel a sense of contribution to other people. So, and then I've been thinking about how to do this. And then, as you mentioned, learning in public is one approach we could take. And I was learning a lot of things, reading a lot of things, but why can't we share what I learned? Or, you know, things inspired me. And what if everyone shared the things they learned or their wisdom with others? What would the world look like? So that's, you know, I was always wondering. So I wish people do this. And yeah, so that's what I want to do.
Alessio: Yeah, that's a great mission.
Glasp: Thank you. And yeah, thank you so much for, you know, beautiful, you know, answers. And also, yeah, really, you know, insightful answers and sharing your experience.
Alessio: Thank you guys for the time. This was fun.
Glasp: Thank you so much.
