AI Agents, VC Insights on AI, and Building in Public, with Yohei Nakajima, Creator of BabyAGI

Full Transcript

Transcript

Yohei Nakajima: 0:00 People are just doing what ChatGPT is telling them to do. Could you just remove the person from that equation? That was the question I asked myself, and the challenge I posed myself was, alright, let's prototype an autonomous startup founder this weekend.

Nathan Labenz: 0:12 Which is an insane sentence, by the way, but yes, carry on.

Yohei Nakajima: 0:15 And then I was like, oh, make the world a better place. And then just started building a list and started thinking about ways to make the world a better place. And I was like, oh, okay. I guess this could be more than a startup founder. I think the next one I did was build as many paper clips as possible.

Nathan Labenz: 0:28 Canonical.

Yohei Nakajima: 0:28 And it started by saying this is actually kind of dangerous. Let's first create safety protocols was the first thing it did. And so if you give companies the scaled ability to control what a chatbot might nudge a person to think, there's a lot of responsibility on that to nudge humans in the right direction or not take advantage of it.

Nathan Labenz: 0:49 Hello, and welcome to the Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week, we'll explore their revolutionary ideas, and together, we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my cohost, Erik Torenberg. Hello, and welcome back to the Cognitive Revolution. My guest today is Yohei Nakajima. Yohei is a venture capitalist, founder of Untapped Capital, who's become best known over the last year for what he has personally built in public, originally mostly for himself, but ultimately inspiring tens of thousands with the potential of AI automation and agents. Yohei, unreasonably modestly, describes himself as lazy. But after this conversation and after watching all of his project releases, I think his defining qualities are really curiosity, playfulness, and a love of exploration. Over the course of 70 public releases, he's helped the entire AI community get a glimpse of the sort of highly personalized AI powered conveniences that seem destined to raise living standards for all. And as much and perhaps more than any other creator or builder that I follow, it's always been very clear that he is having a ton of fun along the way. We cover a lot of ground in this conversation starting with his recent TED AI talk, then running through the many apps that he's built and how he goes about it, before finally digging into how he thinks about the AI space as an investor. The first half of this conversation is a great study for any aspiring AI operations specialists, especially considering what a key role no code proficiency has played in Yohei's journey. And I personally found a lot of value in his answers to my rapid fire investment framework questions in the second half. Overall, though, if there's one thing that everyone should take away from Yohei, it's the way that he's delighted in the process of incorporating AI into his life in a way that serves his own idiosyncratic and ever evolving purposes. If you find value in the show and want to help it grow, I'd suggest sending this episode to a person you know who tried ChatGPT and still says they couldn't figure out what they'd actually use it for. An Apple or a Spotify review would be helpful too. And, of course, you can always reach out to us for any reason at tcr@turpentine.co or by DMing me on your favorite social media platform. Now here's my conversation with Yohei Nakajima, venture capitalist and prolific in public AI builder. Yohei Nakajima, welcome to the Cognitive Revolution.

Yohei Nakajima: 3:33 Thank you.

Nathan Labenz: 3:34 Very excited to have you here. You have been one of my favorite Twitter follows for the last year or more. And I think you probably really don't need much of an introduction to our AI obsessed audience, but it has been fun to follow your progress as you've really gone down the rabbit hole of building all sorts of AI projects while also continuing to run your VC firm. And increasingly, I understand those are kind of bleeding together as you're automating more and more of the processes that the firm does. So I want to kind of get into all of that and understand your perspective on building with AI, investing in the AI arena right now. But for starters, I understand that you just are pretty fresh off of a TED Talk at the TED AI event in San Francisco.

Yohei Nakajima: 4:29 Yeah, that was quite an experience.

Nathan Labenz: 4:32 So tell me about it. And I don't think we'll have access to the video for a while, but perhaps you can kind of share your message.

Yohei Nakajima: 4:39 It should be up soon, so I wouldn't be surprised if it's in the next few days. But yeah, it was a pretty cool, wild experience. I mean, I remember when the TED AI organizers reached out asking if I was interested in talking, I'd been such a big TED fan forever. I was so excited.

Nathan Labenz: 4:54 It's an easy yes.

Yohei Nakajima: 4:55 Yeah, I was immediately, yes. And I'm pretty sure I sat down a day later, and just drafted out a version 1 of my speech as just a thought exercise. And actually really liked that. So I just kind of, I used 70% of that directly in my talk, but what was wild for me, it was really more the speaking experience. I don't know when the last time I memorized a 9 minute monologue was. So there was a lot of practice compared to other public speaking. But the experience was fun. I think all the speakers too. I mean, we don't get that opportunity to prep and practice a short talk. All the speakers were kind of all in their zone practicing and it was really cool to kind of all be in the back speaker room, pacing around, trying to remember a talk, all a little bit nervous and kind of bonding through that experience. And I was completely fanboying over, I mean, it was Andrew Ng and Stephen Wolfram. And, Jim from the Voyager paper, Joon from Stanford Smallville, all my favorite AI follows were all kind of there. Harrison Chase, Amjad from Replit, Sarah Guo from Conviction. So it was a really, really fun, just geeky nerdy few days.

Nathan Labenz: 6:09 Yeah, it sounds awesome. Was the audience an AI obsessed audience? I mean, typically at TED, right, it's super diverse, and then people kind of get exposed to all these, I understand it's often pretty new ideas, right, where part of the value is that people are hearing from people they might not always hear from. And obviously for the speakers, it's a high value audience. Was this still that kind of audience or was it a more intensive AI audience?

Yohei Nakajima: 6:34 It was, I'd say, somewhere in between, right? Compared to a TED audience, there was a lot more AI enthusiasts, but compared to an AI event, there was a lot more TED enthusiasts. There are some people who just love TED who've been going to TED. Not surprisingly, San Francisco has a pretty large group of those folks, and this was the first TED in San Francisco. So it definitely did attract a pretty diverse crowd. And I mean, right now, if you're in SF and you're interested in TED, the likelihood that you're interested in AI is, I think, pretty high. So yeah, it was different from a lot of my other AI events to some extent, for sure.

Nathan Labenz: 7:07 So what did you think? It's an interesting thought experiment for me because I basically only speak to an AI obsessed audience. How did you think about the challenge of trying to say something substantive enough to be of value to those that are deep in it, but also accessible enough to those that are maybe newer.

Yohei Nakajima: 7:31 Yeah. TED is interesting in that their main audience is almost more online than in person. Obviously, the event is huge, but they get so many viewerships online, so they really optimize for production quality. And then when I was thinking about what to talk about, I was thinking, okay, it's probably going to be one of my more watched videos. I want to appeal to a wide audience. So I ended up going for a very broad theme of identity, and then kind of tied my work from BabyAGI into identity into kind of collective conscious and the opportunity for AI to increase kind of understanding amongst people. So that was kind of the direction I went.

Nathan Labenz: 8:12 Yeah, that's interesting. I definitely see a lot of potential for that, actually, going back to my experience as a GPT-4 red teamer. One of the more interesting experiments that I look back on now, and of course, now we can just continue building this sort of stuff, but was mediation between neighbors. I played the role of 2 hostile neighbors, and GPT-4 played the role of mediator between us. And they tried a bunch of these little simulations. One was a group workout thread where GPT-4 was basically supposed to facilitate and kind of encourage us. And another one was tech support for my 90 year old grandmother who has an iPhone and calls me up when she can't find the email she wants to find or whatever. And I do find it's just so striking how empathetic it can be and kind of understanding, and it can pick up on

Yohei Nakajima: 9:10 And it's great at translating ideas into a context that someone else might understand from a different background. I mean, translating language, but not just language, but it can almost translate culture to some extent too. So if used correctly, I think it can definitely help people better understand each other, both in the usage of AI, or you can use AI to help 2 people communicate with each other or just as an individual using ChatGPT, you can just decide to try to better understand somebody and use ChatGPT to help you do that. But also in the development too, tying it back to BabyAGI, a lot of BabyAGI, I'm trying to get it to work better. And my general ideation is looking at BabyAGI do something and then if it's not doing something the way I would do it, think about what do I need to adjust that would act more like me. So it's very much a lot of self reflection as I'm trying to build BabyAGI. It's just trying to close the gap between what I'm watching and what I would do. So I think both in the usage and development, we better understand ourselves and our cognitive process.

Nathan Labenz: 10:10 I do think that's becoming more and more striking all the time. I definitely want to unpack and get into all the good BabyAGI stuff in just a second. But it's been a striking theme in recent research that a lot of the patterns that the latest projects are demonstrating are looking more and more human like. And I'm someone who's pretty resistant, borderline hostile to AI human analogies because I think it's very, it's too easy to mislead ourselves and confuse ourselves with them. But some of these recent things, I'm like, boy, that's starting to look pretty similar to kind of what I understand our own architecture to be.

Yohei Nakajima: 10:50 If you look at the word AI, artificial intelligence, and assuming when we talk about intelligence, we're talking about human intelligence, then it is natural for advancement of AI to start looking more like human. We are essentially, not trying to map the, I think a large part of AI is trying to map our understanding of our cognitive processes onto a software stack to some extent.

Nathan Labenz: 11:08 So do we have to wait for the video? Or can you kind of give us a little bit more about the vision for the improved understanding future that you outlined at TED?

Yohei Nakajima: 11:18 I went more for a narrative that weaved together ideas. So I talked about the development process of BabyAGI being self reflective, talked about kind of our identities being less internal, but more our many roles amongst the many different groups of people that we're part of, whether a country, family, all that kind of stuff. So all the different roles make up who we are, which kind of led into the idea of kind of how our identities are actually, to some extent, overlapping because we have shared identities that we both feel like we're part of. And you have that. So we are complicated is kind of where I think we're at the middle point. And then I touched on language and how it's linear compared to our parallel world and how it's limiting and just understanding that kind of setting that context. And then jumping to anthills and how they act like neural networks and how the anthill has a better understanding of the colony's environment and intent than a single ant. And perhaps suggesting that that might imply that I too am part of something that's just really complicated that I would never fully understand, which led to talking about the journey of self discovery as a continuous process, as opposed to a destination, because you're never going to fully understand yourself, because we're just too complicated. It's what I was trying to get to. Then, and because our identities are overlapping, the journey to self discovery is a shared journey. And then I talked about the opportunity for AI to help kind of strengthen understanding, collaboration, and then end with a hurrah hurrah, let's, let's, let's collaborate, let's build cool stuff. Let's embrace the fusion of AI and our shared spirit for a better future. But it was a fun talk. It was very much more narrative, kind of slightly confusing on purpose and inspirational at the end. So that was what I was going for.

Nathan Labenz: 13:09 Well, there's nothing like the modern AI moment to create some confusion. So I think you're, in some sense, you'd be doing the audience a disservice to suggest that you or anyone else has it all figured out.

Yohei Nakajima: 13:21 I think accepting confusion is an important part.

Nathan Labenz: 13:25 Yeah. Certainly. So maybe I'll circle back to some of these kind of super big picture questions a little bit later. But for the moment, I'd love to get a little bit more into some of your projects. Obviously, the project that you're best known for is this BabyAGI project. And I've been following you for some time when that dropped and certainly know that you'd probably put out 20 different things before.

Yohei Nakajima: 13:49 I think I'm at 80. Well, I was at, I was at 7 AI project. Number 70 was BabyAGI.

Nathan Labenz: 13:54 Wow. That's amazing. So maybe even start earlier than that. I mean, I guess you kind of just catch the bug, and you're just, this is what you do. Right? Nathan Labenz: 13:54 Wow. That's amazing. So maybe even start earlier than that. I mean, I guess you just catch the bug, and this is what you do. Right?

Yohei Nakajima: 14:02 It'll be helpful to go back before AI. I've always been a tinkerer in terms of, my dad was an engineer, and he bought me books. So I did start playing with code in high school. I never became a developer, but I always enjoyed it. And so I always found myself just hearing about something and just spending a really short, tight amount of time just digging in. I remember when Ruby on Rails picked up, whenever that was, it was really hot. I spent a Christmas break learning Ruby on Rails, and that was just my hobby project. And then at a couple of other places I've worked at, like Techstars and Scrum Ventures, I kept building internal tools almost for myself. So I'd have a Google Sheet script or a Google Sheet with a ton of scripts in there, just automating data entry and connecting the Google Sheet to a Crunchbase API, an email search API, all that kind of stuff. I was already building at Techstars and I kind of rebuilt it at Scrum. So I've always been just building stuff to be a more efficient guy. And then when I started our fund, I knew from day one that I wanted a custom CRM because I had all these workflows in mind that I've built at other firms and I wanted a CRM that was just all that was baked in. Interestingly, I actually initially tried to code it myself, and I was using PHP and MySQL because that's kind of what I knew. And I quickly realized that is not where general partners should be spending time. This was in 2020 when Webflow and Airtable, they were all kind of hitting that growth level. So I decided I want to learn about no code and just scrap my code and just decided to build what I was trying to build with code on Airtable. And so that was my first kind of foray into building in public. It was pandemic hit. I was starting to find was the first time I wasn't working for anybody. So as I started building, I just started tweeting about it. It was mostly just refreshing that I could, because I always thought I was building cool stuff, but I wasn't allowed to share it. And also pandemic, right? I wasn't seeing anyone. So just lonely building at home. And the result of that was I got a lot of no code deal flow. A whole bunch of no code founders reached out, a bunch of VCs that knew me saw me as the no code guy and sent me all their no code deals. This is an awesome hack. And then kind of organically, same thing happened in Web 3 where I just found myself interested in NFTs, launched an NFT project and got a whole bunch of Web 3 deal flow and then a whole bunch of Web 3 VCs. And I guess to some extent, AI was the same. It was more organic. I came across the OpenAI API in August. And then as soon as I started playing with it, I just was obsessed. And that's when I started building one or two little hacks a week. Initially, it was no code. And then eventually, when I realized that I could get ChatGPT to write code, switched to kind of coding code. And then one of the coding projects ended up being BabyAGI.

Nathan Labenz: 16:37 Hey. We'll continue our interview in a moment after a word from our sponsors. You've got your own little recursive self improvement, exponential takeoff phenomenon here.

Yohei Nakajima: 16:48 It's fun to think about starting with code, realizing it wasn't worth my time, switching to no code, and then now it's code is faster.

Nathan Labenz: 16:56 Yeah, man, that has changed dramatically. This is an area actually where I don't know if I struggle or maybe I just don't even have the right mindset to even begin to struggle. But I'm pretty savvy with all these technology things, certainly all about the AI these days. But even before that too, I can make things work in a no code environment. But I just don't find myself applying it to my own life very much. I don't build, I don't automate workflows for myself. I'll do that for projects or if the customer success team needs something, I can step up and make it happen. But do you have a tip or a perspective or something along these lines that would help me or others that kind of struggle in the same way to identify and develop this as a personal practice? Because I've never managed to catch that wave for myself.

Yohei Nakajima: 17:56 I think, and I'll caveat that to some extent, part of me building automations for myself is fun and exploratory. And so the return on investment isn't just efficiency gains, but also enjoyment and learning. If I was just looking at efficiency gain versus the amount of time I'm spending on building, especially right now, I don't think it would be a great ROI. But then when you layer on enjoyment, learning, deal flow, other stuff, then it does end up.

Nathan Labenz: 18:27 AI invites.

Yohei Nakajima: 18:28 Yeah, exactly. So the return on investment comes from other parts. But I think that the tip where I started was really just being lazy. I hate data entry. I hate doing the same thing twice. If I caught myself doing something, and I thought, I feel I shouldn't be doing this, can I automate this is kind of where my first thought went. So even using the Google sheet, a lot of people might, there's a lot of email hunting tools. I use Norbert. So you have a list of names, you need to get all their email addresses. You could sit there and copy paste names from the Google sheet into a web form and then copy paste output back into a Google Sheet. But as soon as I get to number five, I'm, oh, I don't want to do this anymore. And even if it takes me a minute, it's depending on how long the list is, I'm how long would it take me to build a script that would automate this? And then maybe I'll spend five minutes building a script and then just pressing go and then just watching the script do all the work for me. And I remember the first time I did something like that, I was just, oh my god, I'm gonna do everything. And then once you start building those, then you have these little sheets, little code that you can reuse. So it becomes more and more efficient the more you do it because I'm, oh, okay. I need to scrape 1000 websites. I already did that before. Let me just go find that project, pull that script and adjust it so I can use it for this project. And then so they do stack on top of each other. And at this point, I feel most of my ideas, a lot of my ideas do end up basically pulling small bits from past projects. So it just becomes more and more efficient, I guess.

Nathan Labenz: 19:58 Yeah, that's interesting. I guess I have to develop that practice. Maybe I, in some ways, have a little too high of a tolerance for pain and just power through more of that stuff than I should.

Yohei Nakajima: 20:11 And then especially now, I think now it's a little bit different. Now I have a whole bunch of curiosity based projects. When I ask myself, oh, I wonder if I could get Dolly to do this. Last night, I was wondering if Dolly could do 360 images. I asked ChatGPT, I just asked it if it could do it. If I have a curiosity, when I have a question about whether or not GPT-4 or Dolly can do something, till I try it and post about it is usually 12 hours. If I think about it in the morning, I'll do it during my lunch break. If I think about it in the afternoon, I'll do it at night after I get my kids down. And so I just have a really fast curiosity to when I think about it, I'll just open up ChatGPT and ask the question. And then I'll get to it and fix it when I have the time. But I'll start the project immediately.

Nathan Labenz: 20:57 Yeah. The modern language models are an unbelievable boon for that.

Yohei Nakajima: 21:02 They're so fun. Yeah.

Nathan Labenz: 21:04 So you had 70 projects leading up to and perhaps including BabyAGI. How many of those are they all just kind of retired at this point? How many of them would you say continue to play a role in your life? I think that's also kind of an important calibration for people who may not understand just how much of this stuff is kind of spun up, but also eventually left behind. I don't know how it is for you.

Yohei Nakajima: 21:28 Yeah. I mean, I'd say it's hard to say because some of the projects do stack. When I say 70 projects, some of them might be improvements on an existing project to some extent. Right? So my investment memo to my AI analysts that are doing a lot of research for me, there are probably 15 projects that are stacked on top of each other where the first version was no code, but it was slower. And then I rebuilt it on code. And then I added kind of API integrations. Right? But they stack on top. I mean, I think probably 20 to 30% of the projects I actively use. So probably, I think 70 to 80% get to some extent retired. But of the retired percent, of the 80%, maybe 20 to 40% are kind of just adaptive. So people can still use it. I still sometimes come back and play with it. And maybe I'll come back to improve on it later. But I see 50%, yeah 50 to 60% get retired.

Nathan Labenz: 22:27 Yeah, we talked to, you mentioned Abjad from the Today Eye event. He was an early guest on our podcast. And I also think about Flo from Lindy in this context. Just the notion of single use or single use would be even more extreme, but kind of disposable software or single use software is definitely something I'm increasingly playing with these days. It sounds like you build a little bit more to last than that.

Yohei Nakajima: 22:55 I definitely do single use though. I remember even, I mean, going back to just trying to automate everything, when we saw that our home was assessed really high one year, and I was, okay, I want to contest this. I just went to ChatGPT and said, okay, I want to contest my home assessment price, walk me through it. And it just laid out a step by step. And I was, okay, go find some comps. I was, I'll go find some comps. And then eventually it even wrote software, wrote code to calculate adjusted price of all my comps. And then eventually just wrote my whole assessment or the report that I ended up sending in, and it actually totally worked. But that was a one time use, right?

Nathan Labenz: 23:34 Take us then to the BabyAGI moment. I mean, there's a lot of interesting angles to this. It was one of the first, if not the first of kind of a small Cambrian explosion of its own of projects. I guess I would describe them as putting the language model in the loop. Yeah, basically. So yeah, I mean, tell us kind of how you started that project, and obviously it's continued to evolve.

Yohei Nakajima: 24:00 Yeah. It was actually in March. It wasn't technically the first, but it was the first popular one. It definitely kicked off the whole trend. But I was looking at Hustle GPT. I don't know if you remember that, but people were using ChatGPT as their co-founder. And there was a whole community and Discord of people starting with a 100 and growing it. And I was just fascinated by that idea. And I wanted to do it myself, but I thought I'm too busy. I should not do that. But when I was, but as I watched people do it on Twitter, they were just doing what ChatGPT told them to do. So I was, I feel if that's all, people are just doing what ChatGPT is telling them to do, could you just remove the person from that equation? That was the question I asked myself, and the challenge I posed myself was, all right, let's prototype an autonomous startup founder this weekend.

Nathan Labenz: 24:44 Which is an insane sentence, by the way, but yes, carry on.

Yohei Nakajima: 24:48 Well, then I thought, okay, how do I work? I wake up in the morning, I look at my task list, I start at the top and I go through them one by one. And then at the end of the day, as I finish tasks, as I get emails, I add new tasks to my task list. And then every once in a while, I reprioritize my tasks. And that was pretty much, okay, that's basically what a founder probably does too. So let's just start there. So it was start with an objective, it would create a task list, and then it would just start executing the task list. Anytime a task was finished, send it to OpenAI to see if it wants to create any more tasks. Anytime there's a new task, send it to a prior reprioritization agent that would just look at all the tasks, reorder them, and then send it back into the execution queue. So it's really just three prompts looped around a while true loop. And when I asked it to build a mobile health company, it would just start kind of start thinking through. And then I think somewhere in there, I also stored at least the task names in Pinecone embedded. And then when it was creating new tasks, it would look at previous tasks that, the most similar tasks that it created. So it kind of created that kind of randomness and kept it at least initially from just repeating the same tasks over and over again. But yeah, when I asked it to start a startup, it just started thinking through business model, fundraising strategy, hiring strategy, let's research some competitors. And it was mostly language. I mean, it was all language because it was just OpenAI, but it was thinking through all the different aspects of building a startup. And then I posted that video online kind of casually, woah, if I press go, it just keeps going. And that's when that video went wild. And I think it quickly went to 1,000,000 impressions, people asking, woah, can this do more than be a startup founder? And then I was, oh, make the world a better place. And it just started building a list and started thinking about ways to make the world a better place. And I was, oh, okay, I guess this could be more than a startup founder. I think the next one I did was build as many paper clips as possible.

Nathan Labenz: 26:42 Canonical.

Yohei Nakajima: 26:43 And it started by saying this is actually kind of dangerous. Let's first create safety protocols was the first thing it did. And then somebody tagged Yud who pointed out that, oh, it started with safety protocols. This is better than most big AI companies. And he retweeted that project, of course, so all the doomers saw what I was working on, which turned into a whole bunch of doomers and EACC type people DMing me opposite messages.

Nathan Labenz: 27:07 So this was just before or just after GPT-4 when you first did this?

Yohei Nakajima: 27:12 This was before GPT-4, I think.

Yohei Nakajima: 27:12 This was before GPT-4, I think.

Nathan Labenz: 27:14 Yeah. I mean, if it was March, it was certainly right around that timeframe. So obviously, this set off or helped set off a moment of, oh my God, AI agents, these things actually. There had been sort of a sense of a big separation between agents and language models, at least for many people, where agents are the kind of thing that DeepMind does, and they're very intensively trained and kind of, I think of them as hard intelligence where they're extremely good, often superhuman at what they do, but they have these narrow domains and outside of them, they don't do anything. And then obviously the language models are I kind of think of that as soft, sort of flabby intelligence for the most part, where it's getting really amazingly good at a lot of things, but it's not really a specialist in anything, but the powers in the generality. And I think what you and others kind of showed is that you can get this agentic behavior out of a language model very easily by just prompting it. And with, as you said, just 3 prompts, you can kind of put it into a cycle where it can kind of keep iterating on its own work and maintaining its own state and updating its own plans and increasing it with tools and affordances of all sorts, actually accomplishing to some degree things in the world. And then everybody was, oh my God, we're going to have AI agents, and they're going be all over the place before we know it. And then that kind of slowed down for a minute. And I'd love to get your take on kind of the post, obviously there's this huge surge in interest in GitHub Stars galore. Seems like you must have seen something early there where some others that had a viral hit raised money or kind of were, this is going to be a company. You did not do that as far as I understand and have kind of kept it as a hobby. So starters, I wonder how you decided in that moment where the world was beating a path to your door that this was not something you were gonna go all in on, even though you probably could have easily pivoted your whole life into that project.

Yohei Nakajima: 29:28 So I think it's pointing out that BabyAGI was 105 lines of code. So was really simple. And it was really more, hey. Here's a simple framework to start working on this idea. It was in my mind, was how I open sourced it. It wasn't like there was clearly a lot of work to do, because it was just a large language model. But obviously, I definitely fueled the fire and making sure it went wide and spread, because I noticed that it was going viral. But yeah, as an investor, well, yeah, that's a good question. I did explore turning it into a company. I talked to a couple of VCs. I talked to a couple of studios. The one thing I knew was that I didn't want to be the CEO of a company while trying to run a venture fund. And to some extent, that could be from PTSD I had when I launched an NFT collection while I was running a venture fund. And it was just running, I was running an NFT collection called pixel bees to learn about web 3. And I was trying to be the CEO while, basically run a project while doing that. And it was just too much. I was just I can't, I don't want to do 2 things at once. So with BabyAGI, I kind of poked around to see if somebody would want to run it. Really I quickly saw a whole bunch of different people built quickly building things were just much more powerful than what I'd shared. Because I shared 105 lines of code and then I went back to work running my VC fund. By the end of the next week, there was multiple dozen projects where teams of 2 or 3 people who are much better developer than I was building on top of this framework to build really cool stuff. So for me, I actually saw it almost as just as an investor, that's the ideal opportunity of all these really cool companies starting. And they're all kind of starting to go in different directions, some were in gaming, were in research, some was in financial areas. I thought, I'll keep working on BabyAGI because it's a good way to connect with the community. But ultimately, I want to invest in these different companies because I think it's going to be a very, I guess I didn't have a conviction in this space. And I and I like the idea of being able to work with different multiple companies tackling different niches.

Nathan Labenz: 31:27 That's if nothing else, it's a powerful object lesson that you can take forward with you as a VC that certainly in this AI wave, the tumble and the kind of things crashing over one another is extremely, extremely rapid. So now and I do I definitely wanna get more into your outlook as an investor in a minute as well. But the project has continued to evolve. I've noticed a couple interesting things about it. One is that instead of kind of having a canonical version, you sort of seem to be just forking off these kind of different mods. And the latest is BabyFox AGI.

Yohei Nakajima: 32:06 It UI for the first time.

Nathan Labenz: 32:09 So, yeah. And it's distributed remarkably.

Yohei Nakajima: 32:12 People are very confused by the GitHub. So what happened was after I released BabyAGI, people wanted to contribute. It was great. And I pulled 2 in and I realized, okay, reviewing pull requests is not where I should be spending my time, I don't even understand the code that's being submitted. And I had community members offering help, so I said, okay, go for it, let's pull in stuff that sounds good, and we started pulling stuff. And then when I sat down to wanna work on it again, I quickly realized I didn't like reading other people's code. And so I just took my original version and modded it as my next playtime. And then when I showed it to some of the community members, they're okay. Why don't we just create a classic folder and you can just stick it in there? And then that's how it started. And that one I called BabyB AGI. And then I just, at that point, just kept building on top of. BabyB turned into BabyCat, so I just stuck it back in the classic folder. And for me, I think, one, just having the animal names was the fun way to keep track of different projects because I don't remember numbers. So if I call them BabyAGI 1, 1.2, I wouldn't remember which one I had to go back to to find the idea. So having animal names is a way to just make them easier for me to remember. And I guess to some extent, so far, the animals have also been animals that are represented in the Pixel Beast collection, so there's a weird tie in there that was just for fun.

Nathan Labenz: 33:25 So what would you say you've learned over these kind of 6 or so iterations that you've done personally? There are obviously a lot of development efforts going on in a lot of different places to try to make agents work. You've implemented some new strategies, but I wonder kind of what your upshot is right now in terms of the prospect for agents and what are the things that are that have come online over the last 6 months that are really working?

Yohei Nakajima: 33:54 And I think there are a lot of different approaches to take them. And I think they to some extent, there is a little bit of the specialized agent versus full generalized agent as slightly different things to focus on. The fully generalization, I think will just take longer to do. So as a startup, it's to some extent slightly riskier because it might be harder to find the ideal customer. It might be harder to get it to a point where people are happy with it every time they use it. And then more niche ones that start niche probably are easier to monetize, easier to find an ideal customer, easier to build for them. So the startup, what I'm seeing more, I think, especially the ones that have revenue are more specialized starting with a specific customer type and a problem type and working there. I think that makes sense. And I do believe they will over time generalize more and more, and that's how it seems like they're building. Having the luxury of not needing to monetize BabyAGI, I've decided to just slowly chip away at the generalized Asian route even if it takes longer. And I benefit from seeing what everyone else is doing all the time because I'm watching everybody else. That's kind of my design process. And then I see BabyAGI, what it has been for me and most of my coding has been just a way for me to share ideas with the greater AI and innovation community. So when I think about what to add next, I don't get excited about the idea of adding something that I've seen 3 other agents do. So I'm usually purposefully looking for a novel idea to test is how you think about it. So with Baby Fox AGI, the novel idea on the UI was being able to run multiple tasks at once, separating out tasks from chat to make it seem faster. It does feel faster than ChatGPT when I use it. So yeah, that's kind of my thought process. And it's a fun way to collaborate with the greater AI agent development community.

Nathan Labenz: 35:48 Would you say that the generalist agent model has hit a point where it provides actual practical utility for you personally yet, or is it still more for fun?

Yohei Nakajima: 35:59 I think it is. I use it all the time, mostly for research. But now that it's connected to not just the web, but I have a Crunchbase API, so it's connected to Crunchbase. It's connected to my CRM, which is Airtable. So now when I ask it to do research, can have it look for people who live in San Diego and then research what they're working on now and put together a report for me.

Nathan Labenz: 36:22 Give me that in real detail. So you sit down, and I noticed this is distributed on Replit, which I love, and we've done a couple of I mentioned Amjad already, we did a whole series on Replit, and I think it really has a chance to be.

Yohei Nakajima: 36:34 Oh, wait. I have to say something about Replit. My tech stack is starting on ChatGPT mobile, because that's usually I'm out and about, is is where I start coding. I copy paste it into Replit to get it running.

Nathan Labenz: 36:47 Mobile app?

Yohei Nakajima: 36:49 Yeah, mobile app, usually.

Nathan Labenz: 36:50 The best developer mobile app I've ever seen.

Yohei Nakajima: 36:53 Yeah. And then when it's working, I will go on the computer and copy paste my Replit code into the GitHub UI because I still don't know how to use Git and then open source it.

Nathan Labenz: 37:03 That's incredible. I think that is also a really good There's a guy on Twitter also that I'm sure you see plenty of his stuff. Levels.io is his Yeah.

Yohei Nakajima: 37:11 I love his style. Yeah. The one file PHP companies. Yeah, that's definitely my type of

Nathan Labenz: 37:16 It's a good reminder that you don't have to master every professional tool to build useful stuff. And I do think that is something that occasionally people need to be reminded of. But so in the actual practical use, okay, so you've coded something up and then you kind of give a short version of the have it do research for me. I'd love to hear that in a little bit more detail. To be totally honest, outside of ChatGPT where I have had some pretty agentic experiences with CodeInterpreter, where it will explore a dataset for me and kind of iterate on code to get the dataset into some workable form. That's the one agent experience I've had where I'm, wow, that thing actually kind of did what I would do where it ran into a problem and tried a couple things to poke at it and eventually got to some sort of breakthrough. But I really haven't achieved that with anything else outside of CodeInterpreter. So how do you set it up or kind of talk to it just the right way to get that to actually happen for you on your own little projects?

Yohei Nakajima: 38:24 I mean, I think the one I've used most, because it's this one's easy to kind of explain too. Also, my versions aren't that stable. So actually sometimes we'll use a different tool called BabyAGI UI that someone else separately created on, I think, TypeJS, but it's actually more stable. So I use that one a lot for research, for web research. But I had a lot of companies reach out to me after BabyAGI wanting to know about how they should be using large language models. And one of my very anytime I'm jumping on a company, a call with a big company that's well known public, I will always ask BabyAGI to research the company, find research their business units, research their revenue drivers, research their cost drivers, and for each business unit, come up with 3 strategies that leverage large language models that would be impactful to the business. And then I just step away, and it takes a few minutes to do because it's a lot of web search, a lot of calls. But when I sit down, if they have 7 business units, it's this company, 7 business units, 3 ideas each, that's 21 ideas that's prepared for me. And as I jump on a call, quickly glance through the 21 ideas, and I'm fully prepared to blow them away, because I can just pick 3 to 5 of them and then just expand on it from my own knowledge.

Nathan Labenz: 39:36 That makes a lot of sense in the context. How reliable do you find it to be? Are they is it getting all the business units? If it how often are we kind of missing things? And maybe it doesn't matter in an intro call context where you just wanna be prepared to be impressive.

Yohei Nakajima: 39:51 Actually, think that your question brings up a good point. A lot of people ask me, where should we start with using AI? And what I tell them is start using AI and automations in things that you wish you were doing but don't have time to do, not the things that you are doing and are critical to you. Because at first, it's not gonna be doing it yourself is gonna be better. And replacing a task that's being done by a person with AI to get worse output is not how you should start using AI. And it's not where you should start experimenting and messing around. You have a good human workflow. So doing deep research into somebody right before you jump on a meeting is something I wish I had time to, but I usually don't. So for me, if it misses a business unit or if it comes up with ideas that aren't good, I can easily it's better than not having

Nathan Labenz: 40:36 Don't say the ones that aren't good. Yeah.

Yohei Nakajima: 40:37 Yeah. Yeah. Just don't say the ones that aren't good. And if it misses one and I notice it during the call, I can probably come up with any ideas on my own. But it's more just it takes me a 10 second prompt and a 30 second review to do what would have taken me 15 minutes of research ahead of time.

Yohei Nakajima: (40:37) Yeah. Just don't say the ones that aren't good. And if it misses one and I notice it during the call, I can probably come up with any ideas on my own. But it's more just it takes me a 10 second prompt and a 30 second review to do what would have taken me 15 minutes of research ahead of time.

Nathan Labenz: (40:53) So this is actually maybe a perfect intersection. The things that people wish they were doing that they're not doing or feel like they should be doing and just can't get around to, I think intersects very naturally with something that I understand was kind of a unique approach that you had originally with the VC fund, which was sourcing your deal flow with outbound. So that's fascinating because I think almost every company, almost every organization would be like, should I be doing more outbound? Yeah. Probably. Right? Everybody needs more sales. Everybody has at least experimented with like, yeah, I could go find a ton of people on LinkedIn and message them, but the hit rate's really low. And obviously, it can be time consuming. Same deal on recruiting. If you're trying to source candidates, man, would it be nice if I could reach out to 10,000 candidates? Yeah, but who has time for that? So take people through that a little bit on, from your expertise originally in outbound, and then now everybody's sitting there thinking, yeah, I should be doing more of this. How would you coach them on how to get started with that near universal problem?

Yohei Nakajima: (42:06) I mean, one, to have an outbound funnel. Well, even if it's a Google Sheet, just have a list. Have a list where you can put people's names if you want to reach out to them. Right? And that's the most important thing. Have a place you can put it. Once you have that place, depending on where you're stuck, if you're having trouble getting people's names on there, then you could connect a bookmarking tool like Pocket to it using Zapier so that if you come across somebody's LinkedIn, and you want to reach out to them, you can just press Pocket, and it would bookmark it in Pocket. And then you connect Pocket to your Google Sheet or Airtable, and it would just start adding people to it. So that works. That's the let's get stuff into my funnel really quickly is use a bookmarking tool of sorts that you can just, and I'll use Pocket for anytime I find a startup URL. Use Pocket. It's on my Chrome. It's on my mobile, so I can use it from anywhere. So that's kind of the step one is just get to a point where you have a list and your list is growing. Once you have that, then you need to figure out how you're going to reach out to them. If you are going to reach out to them, I use a tool called Voila Norbert, V O I L A N O R B E R T, but there's a couple of many other tools to get email addresses, whether it's Clearbit, Hunter. So you need to figure out how you're going to reach out to them. In my case, if I have their name and domain name, Norbert will find their email address for me. So then now you have a list of people and all their email addresses. And then I guess you don't want to automatically send an email, but you can build a template. You could use something like Zapier where you have a checkbox column, or if you, when you check the column, if there's an email address, it'll kick off a zap and it'll go to a template. And then it can send them an email. And if you want to personalize it more, you can just create a new column in your Google Sheet, where you would type in the personalized message, and then just pull that into the email message. And you can just do that directly in Zapier. And if you just do what I did, what I said, then now you have, now you can just go through, go look at people's LinkedIn, click Pocket, which will pull their name and maybe company name. If it's hooked up correctly, it'll automatically find their email. And you can just go through, type in a personalized message, click check, and it would automatically send an email from your email address for them.

Nathan Labenz: (44:17) Interesting that you're not automating, if I understand correctly there, the personalization step?

Yohei Nakajima: (44:24) Yeah. I guess if you wanted to use AI to personalize it, then you would take whatever data you were fed in and you can just add an OpenAI step in Zapier to send the personalized info to OpenAI and say, can you write the precise message? Actually, that would make more sense if you have that info.

Nathan Labenz: (44:38) That caught my ear because I was thinking maybe you didn't trust it to do that or...

Yohei Nakajima: (44:42) I actually write all my outbound emails myself. I did a lot of AB testing at Techstars on outbound emails. I found that 2 to 3 sentence emails just perform the best. And so when I'm reaching out to somebody, I just send them 2 to 3 sentences that's directly to the point that's personalized. My bottleneck is my time. I can send a lot of outbound emails, but I just don't have the time to take all the meetings. So that's not where my bottleneck is. So I guess that's probably one of the reasons I write my emails.

Nathan Labenz: (45:11) Interesting. Yeah. That'd be different for most sales organizations where you probably have a pretty high accept rate on the outbound to the meeting.

Yohei Nakajima: (45:20) Yeah. So, yeah, that's probably it. It's like if I send an outbound email, I'm probably gonna get the meeting. So if anything, my issue is if I send too many outbound emails at once, my calendar gets too full.

Nathan Labenz: (45:29) So where would people start today if they want to get into this kind of agents stuff? Obviously, a lot of different projects out there. I saw you tweeted the other day too that you were cited in a new paper that introduced a whole framework for agents. Yeah. I can imagine you might feel like that's the place to start, or you might feel like, yeah, that's kind of overbuilt. Maybe start with something simpler and then graduate to something more full featured. But what's your advice on just day one with agents for a, let's say, an AI engineer who's actually gonna get into it now?

Yohei Nakajima: (46:02) I'll caveat that I haven't gone through looking at where would I start since the agent space started because I was already in it. So to some extent, I'm not fully caught up to date. Obviously, AutoGPT, if you haven't heard of it, really popular project. Again, I haven't looked at it recently, but that one uses kind of the React style agent where it's kind of doing one task and thinking through. So that's, I think it's worth looking at. I think there's a lot of integrations. There's probably a lot to learn. I'm personally a big fan of LangChain, Harrison Chase. You can build an agent with LangChain. They also have a plan and execute agent, which Harrison said was inspired by BabyAGI to some extent where it does generate a task list ahead of time. Right? LangChain agents were initially React style where it would just do one task, think about it, do the next. But their plan and execute agent will plan tasks ahead of time. And the reason LangChain, I think, would be a good place to start is because there's so many things built into LangChain that if you start just learning how to play with LangChain, you can actually just learn all the things that you can do. And then if there's something you're really interested in, then you can just do it without LangChain. And that would be a good next step.

Nathan Labenz: (47:11) I guess maybe my last kind of big question on the agents before turning over to the investment side of your activity is, are they all gonna wake up when the GPT-4V comes online? It seems like so many of the problems to date have been just getting lost in kind of the friction of the web and all this nonsense. HTML is super bloated these days for all kinds of reasons, but the interfaces are meant to be interpreted visually. It seems like GPT-4V is really good at that. Do you think we're on the verge of sort of a threshold effect where all of a sudden they're gonna be able to do 10 times as much?

Yohei Nakajima: (47:52) I mean, to some extent, yes. I don't know. 10 times is whatever vague number. I feel like anytime a new model, especially something if GPT-4V becomes available as an API, there's gonna be so much you can do that you couldn't do. So it is gonna feel like, wow. These have 10x'd. But I would suspect that we're going to have many, many moments like that. So it's going to feel like, oh man, they can do so much more. But we just keep feeling like that every time there's a new model or whatnot. Yeah, it's going to be many of those moments where we're just going to continue to be impressed. But the reality is, we're incredibly complicated. So getting where the agents are now to being able to do generalized can do anything a human can do is going to be a really long ways. And to some extent, the further along we get, the further the bar is going to get. As soon as agents can do everything online for us, it's gonna be, oh, yeah, your agent still can't make coffee for me. And now we're gonna be getting into robotics, which is also moving fast right now. So it's all gonna merge together. But again, just because GPT-4V's out is, it just moves the bar up to the next one.

Nathan Labenz: (48:51) Is that a reasonable model? And my first gut instinct is like, yeah, it does kind of seem like a reasonable model. I've been working on this on the safety and control side as well because there's been some recent research that's like, wow, that seems like it gives you a whole order of magnitude improvement on ability to detect bad behavior in models or whatever. And then you're also kind of like, but, if it's really serious, there's still a long way to go. There's still multiple orders of magnitude needed. On the agent side, it's kind of coming the other direction where BabyAGI can maybe do one in a million things that a human could do in its original form, and maybe we're headed for 1% pretty soon. And those next two orders of magnitude, I wonder how quickly they come. It seems like you don't think they're gonna come that fast, but I wouldn't be shocked if we're sitting here...

Yohei Nakajima: (49:42) I know. I think there will be things that go fast. The question, I don't know, and again, this is beyond my technical capability, but I do feel like there's, I mean, I'm experimenting what you call the orchestration layer, right, that you're kind of stripping it out of the model and doing it in orchestration. But it does seem like a lot of the stuff, the learning from building in the orchestration layer can just basically be baked into the model. Again, that's a little bit beyond my technical capability, but I think if we start seeing stuff like that, models that are building that can, that can, multimodal, that can do a web search from within the model somehow. I don't know if that's even possible, but to decide to do a web search and then write new skills and store it, the stuff that we're orchestrating, if that can be done within the model, I do think that's going to move so fast. But again, I don't know if and when that's possible.

Nathan Labenz: (50:30) It does seem like it's at least somewhat possible, and somewhat coming soon. But yeah, we'll have to see just how fast. So let's turn then to the investment angle. And you could either take this from a sort of what you're looking for or maybe more of an advice to founders. But I'm really interested in just kind of, everybody has this big question, right? Where is the value going to accrue? Is anything defensible? Everything's moving so fast. It's, I think, become pretty fashionable to say the incumbents will probably keep all the value in this wave because they've already got the distribution, and it's pretty easy to implement the AI. So what's your take?

Yohei Nakajima: (51:09) I think at a high level, we can all agree that a lot of value is going to be created. Compared to previous tech innovations, incumbents will get a larger share of that value. And that comes from largely noticing that both the C levels at every large company know they need to do it. And many engineers are just curious and want to play with it themselves. And I think if you have those two things, it's enough to, it's a lot more than we've ever seen, right? It's usually the engineers are interested in technology or the C level wants to do it, the other side doesn't. But we have both in AI. So incumbents will move fast. They will capture value. There's no question about that. That being said, I don't think there's any technological innovation where incumbents are gonna capture 100% of the value. It's just a matter of what's the ratio going to be. And because there's so much value to be created, I think there's a ton of value to be captured by startups compared to even other past technical innovations. Even if incumbents took 90%, which I don't think they will, of the value created, literally AI is probably 100x bigger in terms of value creation than many technological innovations we've seen. In that sense, there'll be plenty of value to be accrued by startups. So I think it's still a great time to be investing in startups, but also to look at public markets to some extent, I guess.

Nathan Labenz: (52:22) I don't know that I've ever heard anybody put it quite that way. Senior leadership at companies and rank and file developers both being inclined to embrace a technology at the same time. I don't know that I've ever heard it described that way, but I do think that is really apt. And I just had a little bit of an experience like this over the last day at Athena, and this is an executive assistant company where I'm the AI advisor. I was just struck in the last 24 hours that talking to founders and then talking to a person who was hired at the company as a software developer, not as an AI role, both are obsessed with AI and literally reading papers. You don't see people reading work out of academic labs in their professional life super often. And now we have people going way out of role to do it. It is really a remarkable moment. I think that's a fantastic observation.

Yohei Nakajima: (53:20) And then just to even add to that, I think, to some extent, just because you have an idea and want to do it doesn't mean you're going to do it well. So there are going to be a lot of large companies who are going to lean in and try to do it. And again, I don't know what percentage, but there's going to be a percentage of them that are not going to do it well, which means they're going to end up possibly buying a startup, buying from startups. And some of those people who learned about the technology and the industry and tried to build it internally, but ran into whatever issues are going to spin off and build their own company to tackle that market or something like that. So there's still going to be, even if incumbents are leaning in, just because they're leaning in doesn't mean they're going to do it right. And the people who don't do it right, that energy will shift into, I believe, the startup market. Yohei Nakajima: 53:20 And then just to even add to that, I think to some extent, just because you have an idea and want to do it doesn't mean you're going to do it well. So there are going to be a lot of large companies who are going to lean in and try to do it. And again, I don't know what percentage, but there's going to be a percentage of them that are not going to do it well, which means they're going to end up possibly buying a startup, buying from startups. And some of those people who learned about the technology and the industry and tried to build it internally, but ran into whatever issues are going to spin off and build their own company to tackle that market or something like that. So there's still going to be, even if incumbents are leaning in, just because they're leaning in doesn't mean they're going to do it right. And the people who don't do it right, that energy will shift into, I believe, the startup market.

Nathan Labenz: 54:01 Okay, so I give you super strong marks on your analysis on the incumbent versus startup split. Let's try the layer version of it. People are again, well, the chips are gonna make a lot of money, but maybe the applications do, maybe they don't. It's so easy to spin up applications. Maybe it's the tooling, maybe it's the model providers. Do you have a similar thesis on how you think about the layers of the stack?

Yohei Nakajima: 54:29 I guess so, to some extent. I do think, I mean, hardware, just because of scale benefits, there's probably not going to be too many. You're not going to see a world with tons of really big hardware companies building these chips. Would guess it would be somewhat consolidated. Models, I would say maybe a little bit less consolidated. It still takes a lot of money to build, but we're seeing more and more tooling to spin up models of your own, open source models, smaller, more specialized models. There's probably gonna be some variation, but as far as the core foundational models, it'll probably stay to some extent consolidated is my guess. Not saying there's no room for new ones, but as a pre seed investor, I'm noticing that all of these rounds are $100,000,000 kickoff rounds. It's one to keep an eye on, but not one I'm diligencing too deeply. Infra is the most confusing one where you have these extremely popular infrastructure projects that are fully open source and making no money that have raised a lot of money. And you're seeing that, and they're clearly creating a ton of value. But their ability to capture value, I think, is still in question. I'm not saying they can't, but I don't think we've seen them all. Think that it'll be interesting to follow that. And so I think because people aren't sure what's going to happen to some extent, I think there is opportunity if you can come up with conviction and invest. And if you're right, you'll probably make money in the infra layer. And then the application layer, I think there's going to be a lot of value created. There's going be a lot of companies. A lot of them are going to be niche specific use cases that are core AI. And then there's going be some that barely use AI, but because they use AI, they're going to beat everyone else who doesn't use AI. And I think there's a lot of different business models in the application layer, and I don't know which one's going to be the best, but I think a lot of them are good ideas. So one of the ways I'm thinking about it is diversifying, at least in the application layer, the types of business models so that I have some hedging on that as well.

Nathan Labenz: 56:29 Multiple follow-up questions there, but do you have a taxonomy for the business models?

Yohei Nakajima: 56:35 That's probably a good idea to write down. None specifically, but there's some using LLMs on the back end, for example, to take unstructured data or whatever data around the web to build clean structured data, but then the front end doesn't use that much AI. Those actually have much higher margins because they're just using LLM once to collect the data, and they're not using LLMs on every usage, so the margins are higher, so that's an interesting model. That being said, it might be less defensible to some extent because if the unstructured data is public, for example, someone else could, in theory, collect that structured data as well. The most common one is the wrapper, which is less defensible from a product standpoint. But if you can ingrain deeply into the user's workflow and be something they need and grow fast enough and get market share, then that market share could be defensible, especially if you can keep building value on top of it that makes the switching cost higher. That's probably the most common model. Scaled services is what I'm looking at. I think the idea that the reason SaaS is so popular amongst VCs and why it is popular is because SaaS is scalable, software is scalable. But now that we have with LLMs, we can make customer service scalable, we can make sales scalable, we can build companies where it feels like a service to the end user, but on the back end, they're using AI and automations to scale rapidly. I think that's an interesting business model and I'm seeing more of those that hint at that. Those are a couple examples. Oh, then I think one, this isn't really a business model, but more so I do think there's probably room for companies that just use LLMs well internally. And it's not even, if you look at it from the outside, it's not an AI company, it's no AI in the product, but the way the company operates, the way the leadership guides their employees to use LLMs and AI tools is just so efficient that they're just operating as a company 10 times more efficiently than their competitors.

Nathan Labenz: 58:22 Even down to the local business level, just having a decent chat on your website that nobody else in your immediate competitive set has huge advantages to some simple stuff these days. How about the enterprise kind of retail divide? I think you're probably, you may have even better information on this than I do, but my general sense on the retail side right now is that all these apps are posting really good growth numbers, but typically with pretty poor retention with maybe few exceptions. And I wonder if that's kind of what you see as well. And if so, does that lead you to think more about trying to invest in enterprise serving companies?

Yohei Nakajima: 59:08 I think it's a fair generalized observation. I think part of it is because people are interested. If there's 5 legal AI tools and a lawyer is really interested in AI, they might jump and try off eye, but they're not going keep using off eye, which means if the average user is trying 5 tools, then there's going to be 20% retention assuming they end up on one. To some extent, retention being low is just because people are interested in actually exploring and trying to find the right tools for themselves. That's also a good thing. Think it was Sarah who mentioned this in her TED AI talk. Sarah, she had a good one too. I'm excited to watch it again when it comes out. But she talked about minimum viable quality for AI tools. I do think that's a thing. It's not just about getting the tool out there, but making sure it has quality output. So I think with these companies, it's not just about getting distribution, but also about ensuring quality because people will switch, people will churn. But to answer your question from a more high level, I do think there's the retail is that. But then one aspect, one kind of theme I'm seeing is more kind of consulting first enterprise approaches. I mean, I've heard the name Palantir come up in a handful of conversations, probably not surprisingly, but they're kind of doing more stuff there. There's a company called Distill that's some ex Palantir guys doing similar. And I think, Rekha, there's the kind of enterprises want handholding and consulting more, and they have the money to throw down to it. So it actually is an interesting go to market right now to start enterprise consulting, to be able to bring in money, hire good talent, work on really good problems, and then maybe automate that process a little bit and decrease costs to target a larger audience as you go, which I think is a strong model if you can pull it off.

Nathan Labenz: 1:00:51 It's definitely a good time for consultants in general, because there's a lot of people with a lot of questions. And it is one of these rare moments where, and I'm not a huge believer in consultants in all cases by any means. But there is this just huge knowledge gap where I find so often, I'm sure you have this too, where people will come out of the woodwork and ask me a question. And often it's pretty easy to answer. I can just kind of answer pretty quickly and be, oh, you want to set up the simplest possible chatbot that has access to your documents? Boom, go to chatbase.co. You'll be up and running in 15 minutes. I've tried 10 of these and this one really is really simple and good. And people are, oh my God, you're a genius. And I'm, honestly, there's not even that much genius there. I'm honest, I usually do tell them. But yeah, it's kind of just a knowledge gap that will close, but for the moment, definitely makes a consultant often pretty valuable. I had one more. Oh, new things versus kind of old things. I mean, this is another Sarah comment. She said that they're looking at some of these kind of fundamentally new behaviors, like talking to an AI bot all the time. We've had Eugenia from Replica on the show. Character obviously is a leader in this space to the point where Facebook is potentially going to try to do their Facebook thing and ship it to 3,000,000,000 people. How much do you find yourself thinking about ways to apply AI to current things versus just fundamentally new experiences or products or ways to spend your time that are kind of only possible now with AI?

Yohei Nakajima: 1:02:42 I think the latter excites me more. But I spend my time thinking about, I think both equally to some extent. Because to some extent, I think the first one is easier to at least make money in the short run. And it's a fuzzy question because an old thing with AI can look like a very new thing.

Nathan Labenz: 1:03:04 Coding, honestly, going back to one of your things at the top where you said you have all these projects and you go pull out a snippet from whatever. Find myself doing that these days when I'm coding for the purpose, not even so much of me to adapt it, but you may have a better term for this, but I've been calling it coding by analogy. And basically what I'll do is take whatever I have to GPT-4 and say, hey. Here's what I have. It helps if it's working. Here's what I want. And it is damn good at making that jump. So I think that is a fundamentally new experience. I basically don't code manually anymore almost at all. My fingers barely work.

Yohei Nakajima: 1:03:51 Especially on mobile replit, it's hard to copy paste specific lines. So I'm usually copy pasting whole files back and forth.

Nathan Labenz: 1:03:57 Yeah. I mean, that's a pretty profound shift, I think, in what that experience looks like. How about out of your portfolio, anything that you would want to highlight as either, I mean, really examples of any of these things we've talked about, but I'd love to hear some of the things that have sparked enough conviction in you that you've actually invested in them.

Yohei Nakajima: 1:04:16 Obviously, AI was one of my first AI investments. Actually, that was back in 2020 during the no code time. They had no code AI tool, and they built out this incredible tool that can do both traditional AI in terms of predicting columns based on past data, but also LLM tools. But the approach they've taken recently is actually just acting like a service. So they have a data scientist as a service for 1000 bucks a month, and they have internal people who know how to use their internal tools really well and really efficiently. From a company perspective, they just work with their data scientists with an open, obviously AI, who will do all the cleaning of data, all the predictions, all the setting up whatever APIs or whatever you need. But really, it's just team members internally who are good at using a tool they've really robustly built out. So that falls into this kind of scaled services model that they've kind of pivoted into. I think that approach is really clever and they're doing well.

Nathan Labenz: 1:05:11 And not to ask you for any numbers that they haven't disclosed, but you got to have pretty serious leverage. I guess it obviously depends on what your SLA is in terms of what service commitments you're making. But $1,000 a month doesn't buy you a big fraction of a data scientist in today's market. Kind of sounds like inherently, they've got to have one data scientist supporting at least 20 companies, right? Probably more than that.

Yohei Nakajima: 1:05:37 I'm not going go into specific numbers like you said, but I mean, there's definitely scaled benefits to having a platform that can do a lot of the work. I mean, if you ask a data scientist at a slightly old company, I mean, the amount of time they spent on cleaning data and doing stuff that they don't need to do is pretty high. So if you can if you have built tools to automate that part, then you can definitely be more efficient.

Nathan Labenz: 1:06:00 Yeah. That's fascinating. I do think that sounds like a pretty compelling offering at a minimum.

Yohei Nakajima: 1:06:05 I've spent a lot of time in media. I used to work with the Disney Accelerator. I've consulted for Nintendo a lot, so that's an area I just get kinda attracted to. I have 2 companies and, I guess, 3 companies in the media space. Have Oggx Labs that does prompt to video. They started really more around grabbing media clips from around the web to put them together for you based on a prompt or based on transcribing a video, but they've layered on LLM so you can kind of just describe the video on. It'll write a script. It'll create an audio output. It'll chunk it up and find media from online to basically put together a full video for you. Space I know you're familiar with. Spext is kind of in the podcast slash YouTube space. They started as more closer to kind of what's the one that a lot of people use? Descript. But they've really layered on clever LLM. So for all the videos, and you can have it in Discord or Zoom, they kind of bake, you can bake it into Discord or Zoom to record stuff. It'll do chapter summaries, it'll do takeaways, and you can ask questions directly on the video. More recently, they've expanded to being able to do full library. So if you just give the YouTube channel, it'll just go transcribe all the videos, you can ask it a question, and it'll give you all the short clips that are most relevant to that question that you can just jump to from anywhere in their library. And also, can just kind of ask general questions against the YouTube library. That one's pretty powerful. And then they've also built out video editing features on the back end. The video creators, once the library is connected, can say, hey, can you find a clip about AI doomerism and clip out a 10 second clip and then add our opening title to it. And so just kind of natural language engagement against the whole video library is what they've been building out recently.

Nathan Labenz: 1:07:46 It seems like there's a tremendous possibility, although I'm not sure it's necessarily what people want. But the opportunity for sort of semi dynamically generated content experiences, whether it's games or video, the kind of choose your own adventure style media. Do you think that is gonna be a thing? Nathan Labenz: 1:07:46 It seems like there's a tremendous possibility, although I'm not sure it's necessarily what people want. But the opportunity for sort of semi dynamically generated content experiences, whether it's games or video, the kind of choose your own adventure style media. Do you think that is going to be a thing?

Yohei Nakajima: 1:08:10 I do. I mean, I thought the Black Mirror one was pretty interesting, especially with LMs. I mean, to some extent, there's a blurred line between a choose your own adventure media experience and a video game. Video games to some extent is that, I think. So blurring the line is interesting. Especially from an education angle, if designed correctly, I think it can accelerate somebody's learning if it can keep them both engaged and interested and layering on content at the right level for that person in the right way could really accelerate learning or just learning about anything.

Nathan Labenz: 1:08:48 One that I noticed in your portfolio is this company SeniorSign. And I don't know if they have an AI component to what they're doing right now, but I love the fact that you invested in a technology company specifically focused on the senior market, because that's been a long standing thesis for me, especially as demographic shift is underway in some parts of the world and getting underway here in my part of the world as well. It seems like this disconnect between the technology that we have and what seniors actually could benefit from is just so huge. AI seems like a great way to bridge that gap. Have you been seeing any deals or do you have any kind of ideas as to what might take shape there?

Yohei Nakajima: 1:09:25 That's fascinating. Senior center was more my no code phase. I mean, fact that people are onboarding on the senior care centers and huge paper digital binders today is just mind blowing to me that that still exists. Yeah, I mean, I think AI can be very powerful there. I'm excited to see what kind of products people build around that, I think. There is some hesitation, I think, from older people against technology, so it'll be interesting on the right go to market and approach on creating AI that seniors would want to talk to.

Nathan Labenz: 1:10:00 Yeah, that talk to, I think, is the key phrase because my grandmother, for example, has an Alexa device at her side table at the end of the couch, and she'll use it to play music. And she does kind of personify it quite a bit, and she's fully with it. But she'll say, Oh, she wasn't playing it for me today, and stuff like that. And I'm excited. And she spends most of her time alone.

Yohei Nakajima: 1:10:25 She actually

Nathan Labenz: 1:10:26 has pretty good support network as individuals in their 90s go, for sure. But still spends most of her time alone. And I do think just a little more conversation. And you can imagine kind of creating a profile in the background that the AI could kind of. It doesn't have to be that customized. The smallest gesture from the staff where she lives is something that really makes her day. And I also think there's something interesting about seniors where they're not at all cynical about things that were in fact kind of contrived. Sometimes you'd go to this kind of faux downtown or some historical building or whatever, and my grandfather, who was especially this way, he's like, This is so fake, phony, pat on. But he loved it. He didn't care if it was made for this purpose or not. He would still find great pleasure in it. Old people, in many cases, are really endearing in that way. And maybe in some ways could be more receptive to these AI conversation partners because they're just not jaded.

Yohei Nakajima: 1:11:32 I have this mixed emotional reaction to this though. On one hand, I'm like, Yes, I want people to feel heard and connected. The idea that someone doesn't have to sit there lonely, especially I think that's happening more and more, is good. But at the same time, there's also probably the old school part of me that's like, I want to make sure people have human relationships. If there's an AI that's too empathetic, are they going to not want to connect with humans anymore? Is that good for us? Is a question. And maybe the answer is it is fine, but I don't know.

Nathan Labenz: 1:12:02 I think it's a very, very good question to ask. And I felt this way about it when I spent a decent amount of time actually using the Replica app when I had the interview with Eugenia, the CEO there. And it was fascinating in many ways, but one of the ways that was, honestly, maybe the most interesting was this was still pre GPT-4 and pre a big price drop. So she didn't have the ability to do super advanced conversation as cheap as it is today. And my actual experience of the app was, it wasn't super sophisticated in conversation. And yet, she'd grown a big user base, and lots of people really cared about it and were genuinely emotionally invested in this. And my takeaway from that was a lot of people really need this kind of stuff. And for the audience that she had at the time, I came away feeling like this has got to be very positive for a lot of people because, it's a tough world out there, a lot of people don't have a lot of things that I'm maybe privileged to take for granted. But then I also, just like you, I was like, but as this stuff gets really good, and it becomes the norm or it starts to. It's one thing to use this as almost like a treatment for people that are struggling, and it's quite a different thing for it to be the norm among

Yohei Nakajima: 1:13:25 There are going to be couples who break up because the girlfriend is upset at the boyfriend for sharing too much with his AI partner.

Nathan Labenz: 1:13:31 I bet that's already out there. Yeah.

Yohei Nakajima: 1:13:34 Right? It's going to get weird. And then I mean, this I feel like this is going to just go deeper, but this was an interesting conversation that came up with actually, I had a panel, the awesomeness panel the day after Ted with Jim Fan from Voyager, June from Stanford, Smallville, and then Noam, who did the AI for diplomacy game. So it was a gaming for AI panel. The conversation did go at some point, it did touch on if you have an emotional connection with an AI, they also inherently, to some extent, have some influential power over you. And so if you give companies the scaled ability to control what a chatbot might nudge a person to think, then it does kind of take that whole, there's a lot of responsibility on that to nudge humans in the right direction or not take advantage of it. Just something to keep an eye on, I guess, from the safety standpoint.

Nathan Labenz: 1:14:26 Yeah, absolutely. There's a great article on Less Wrong from early this year by a person who had a pretty consuming and borderline deranging experience of kind of falling in love with a character on Character AI. And the fact that this was posted on Less Wrong is remarkable. And the person upfront was like, I'm someone who should have known better because I have a pretty good understanding of what these systems are and loosely how they work at least. And yet, it was so compelling experientially that he kind of lost himself in it for a while and eventually somehow pulled himself out of it and wrote this post. But I would definitely encourage anyone and certainly, if you doubt the potential for these sorts of things to happen or think it would only happen to someone who is dumb or whatever, read this one, and you'll hear an account from an intelligent person who was not super naive going in and still kind of ended up in a strange place.

Yohei Nakajima: 1:15:30 I'll check it out.

Nathan Labenz: 1:15:31 Couple of last questions, and then we can let you go. You've been super generous with your time. Building in public, you've done it and you've had great success with it. Do you advise your companies to do it? I'm struck that there are some. I'm thinking of Harvey. Lindy is a former guest who basically are allowing no users outside of, in Lindy's case, just I think it's almost all internal still. Harvey has big customers, but they basically don't say anything. In a world where a thousand people created derivative projects of your BabyAGI in 2 seconds, Do you advise people to still put it all out there in public or maybe it's just situational, but what's kind of your guidance on how much to build in public if you're building an AI product today?

Yohei Nakajima: 1:16:16 I think it's situational, absolutely. I mean, if you have a strong technical team that you're confident can build the solution you want to build and you have connections, you have deep connections and intros, ways to connect into large enterprise customers who can pay you a lot of money, the incentive to open source actually, I think, is pretty low. You should just build the product and sell it. On the flip side, for example, if you're, if you grew up sometime outside of a major hub, you don't have any connections, don't know anybody, and you want to build a dev tool, then open solution makes a ton of sense. Because if you can get people excited about your open source projects, suddenly you've built a network of people who are familiar with you and your work, and they're even probably contributing to you, and suddenly you have friends that you can build alongside. And then a lot of companies fall into the middle of somewhere between that, but not a lot of people starting companies are like, I have all my customers handy, ready to go, a lot of people are trying to sell to customers. So open source is a decent mark, it's a good marketing technique. Especially if there's a technical aspect where you need developer buy in from your buyers. Then if you can have a popular open source project, then you've already gained credibility in some of the key players of your potential buyers. So it can decrease your sales cycle, right, if you decide to go commercial, open source and build a commercial layer on top. But there are a lot of approaches. I don't, I'm not going to say I know what the right way is. I'm observing and learning a lot. Actually, interestingly, if you are interested in this open source commercializing question, group that follows OSS Capital. JJ there, they run the commercial OS COS, I think, open source conference. So he's put out a lot of content around approaches to commercializing open source. So I've learned a lot from him. And actually, he was the person who convinced me to open source BabyAGI.

Nathan Labenz: 1:17:59 Okay, cool. That's a great recommendation. 2 more questions. One, kind of big super big picture, one a little more personal. I used to ask this question every interview, and then I kind of went away from it. But just recently, at the AI engineer summit, there was a survey result presented where seemingly in a, pretty neutral, although I took the survey. Otherwise, pretty neutral survey, which was about what tools do you use and, do you like LangChain or Lama Index or whatever? Those were kind of most of the questions. Then there was this question which was simply put p doom question mark. And people filled it in, and then they showed the results. And the results were all over the spectrum. Basically, you could call it the uniform distribution to a first approximation. I thought that was extremely striking because it was like, there's definitely no agreement on this. And it's obviously super important if there is a serious chance of p doom. What's your take on that? Do you spend time worrying about it? Do you think about it?

Yohei Nakajima: 1:19:01 Somebody actually asked the question during dinner, I had an answer. I think I think about it slightly differently. I think just asking pdoom actually as a question to me doesn't make sense without a specific timeline. Although if you gave me a timeline, would tell you I don't have an answer. I think of it more as, p doom, which is p doom not specific to AI. I think it was a 100, meaning, I don't think humans will be around in a billion years. And so the question just becomes, what is the likelihood the doom comes from AI? Because there's a lot of other reasons for doom. It could be a comet, it could just be time, it could be a bigger pandemic with the worst disease or whatnot. And then when you start thinking about it that way, you do quickly realize AI does touch on a lot of other things. But ultimately, it's a combination of AI and human. Right? I think the most likely scenario is some sort of combination of human using technology wrong, right? Whether it's nukes or creating diseases or whatnot, and maybe AI is part of building that solution. So I guess my answer becomes p doom a 100. But then if you're going to ask that question, I think it's also important to think about the benefit of AI, which is that it can actually solve against some of the other doom scenarios. So it actually if it decreases the likelihood of other other doom scenarios at least the way I'm thinking about it, it increases the p doom of AI because if the other likelihood goes down, the likelihood of AI goes up. It's interconnected is what I'm trying to say, and it's a hard question to answer, but that's the framework I think about it, but I don't actually have a good answer.

Nathan Labenz: 1:20:23 Yeah. Do you have a sense for which direction we're traveling? Relative to 5 years ago, do you think that there is a, let's just say our natural expected lifespan, do you think there is a notably higher or lower p doom in light of all the AI developments?

Yohei Nakajima: 1:20:43 I believe we are on a path to a much better future. But I say that knowing that I'm an optimist. And I believe the pessimists against it are also just that voice is just as important. So I say I'm hopeful, but that's how I feel personally. But I think the discourse is extremely important amongst both sides.

Nathan Labenz: 1:21:06 I think I sort of shared that perspective. I don't know that I would say I'm an optimist necessarily, but I do see, and I get so excited about the tremendous upside potential of AI. And I guess maybe my answer would be it feels like the distribution has shifted toward more of a bimodal, where things could go really amazing or we could have some really crazy shit that could be really bad. But definitely both are very much in play in my mind.

Yohei Nakajima: 1:21:33 But again, it's like saying, Oh, should we do healthcare research? Because what's the likelihood that if we continue to research healthcare, we're going to eventually create a disease that can wipe up humanity?

Nathan Labenz: 1:21:45 Well, I would shut down some of that gain of function research. I do take a position on that one.

Yohei Nakajima: 1:21:50 Right. I think there are certain health type of research that maybe you should consider, but that is a good differentiation between calling it health research versus specific type health research. And I think saying, should we stop AI progress? Or if it is a specific type of AI that's dangerous, let's talk about what that should be. But I think broadly saying, should AI slow down is similar to saying, should we broadly slow down healthcare research? But I think the answer I would say is wrong.

Nathan Labenz: 1:22:13 What would you say you have learned about yourself in the process of creating all these AI projects and mini Yohei and all those things? Nathan Labenz: 1:22:13 What would you say you have learned about yourself in the process of creating all these AI projects and mini Yohei and all those things?

Yohei Nakajima: 1:22:23 I think I realize how not complicated we probably are to some extent. We are very complicated, but at the same time, a lot of the way we work especially is simplified, our work selves, because of, to some extent, I think, the industrial revolution and making jobs ones that you can replace one person with another. So the way our industry has evolved, just industry in general has evolved is, let's make jobs that we can easily replace one person for another, which means simplifying the job and task to a point that I actually think is pretty easy to map. So what I figured out, I guess to some extent, I've been fascinated at how simple we are when it comes to how we work actually to some extent.

Nathan Labenz: 1:23:02 Yeah. How much of what we do is in fact pretty routine?

Yohei Nakajima: 1:23:06 Yeah. How much of what we do can be automated is mind blowing. And it's not easy to get to a point where we've automated everything, but a large portion of work in most jobs can be automated with the tools we have today, if we just built it and integrated it the right way, which is the difficult part.

Nathan Labenz: 1:23:21 Yeah. Well, you've taken a nice little bite out of that grand challenge personally, and been someone that I've really enjoyed following and inspired many with your projects, including BabyAGI. So we will look forward to the release of the TED Talk coming soon. And for now, I'll just say, Yohei Nakajima, thank you for being part of the Cognitive Revolution.

Yohei Nakajima: 1:23:43 Thank you for having me. That was fun.

Nathan Labenz: 1:23:45 It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.