How To Get AI Startup Ideas

I have news for you guys. YC is throwing our first ever AI Startup School in San Francisco on June 16th and 17th. Elon Musk, Satya Nadella, Sam Altman, Andrej Karpathy, Andrew Ng, and Fei-Fei Li are just a few of those confirmed, the world's top AI experts and founders who will teach you how to build the future. It's a free conference just for computer science grad students, undergrads, and new grads in AI and AI research. And we'll even cover your travel to SF. But you have to apply and space is limited. Link in the description to apply for a spot. Now, onto the video. Right now, if you're building a startup, working on like cutting edge AI, even if you haven't found the right idea yet, why give up and go back to Google or college or something? There's a high probability that your lucky break is just around the corner. Yeah, and this is kind of the most exciting thing you could be doing with your time. Something's happening, I'm not happy with that. Let me go all the way to the edge. Let me go into this outside world and from first principles, understand the root cause of this. And then you're gonna discover all kinds of things that software and especially AI, the current form of AI can actually solve. There is a very important question that all founders have to ask when they commit to an idea. And that is, if not us, then who? Welcome back to another episode of The Light Cone. I'm Gary, this is Jared, Diana and Harj and collectively we have funded companies worth hundreds of billions of dollars, often with just an idea. And that's what we're gonna talk about today. There are a lot of smart technical people out there right now who are following along with this AI stuff. They see incredible potential in the technology. They totally buy that this is a special time to start a company. And the thing that is holding them back from doing so is they just don't have an idea that they're really excited to go and work on. I think we should just basically open source all the tricks that we've learned that we've so far only discussed in office hours, but we should just tell everyone and hopefully this actually helps some people come up with great startup ideas. Well, one of the blueprints that seems to be emerging is that you can't just stay too close to where you're at. Basically the default bad startup idea is lazy in that it's almost like a hackathon idea. I read about it on X and a bunch of people are doing it and why don't I do that? Another version of it is just figuring out what is hot out there and jumping on a bandwagon. Jared, one of the things you caution founders against is actually just maybe running with their hackathon idea. Yeah, often when people are thinking about startup ideas, they tend to gravitate towards things that seem very easy to build the first version of, but most of the best startup ideas are actually at least somewhat hard to actually ship the first version of. And so that was sort of my meta lesson from doing this individually with founders is I'm always trying to push them in the direction of harder, more ambitious ideas and their subconscious is trying to push them back in the direction of things that they can build in like a weekend. I guess the interesting thing about it is you need to get out of the house. So rather than just do what is right in front of you, you have to either aggressively introspect and look within into your history and what you are uniquely great at, or you need to aggressively get out of the house into other places on the outside, being industry, being government, being other places that serve humanity in some way, and that's not in your house either, actually. So aggressively internal or aggressively outside. So why don't we make that real by diving as deeply as we can into some specific examples, what is within us already. There are lots of examples of people who already spent years and years to get all the way to the edge of understanding whether it's AI or some other field in the world, and they did it either in their studies, in their research, or where they worked. Diana, I feel like you have a few really good examples. Yeah, so basically the examples here are of founders who had this very unique experience in prior jobs they were working at. And one of them is this company that I mentioned before called Salient. They're basically building a AI voice agent that does loan processing for auto debt collection. That is a bit of an esoteric idea. And when I was working with the founders, it took a bit of time to land on a good one. And this turned out to be a good one when I heard that Ari, the founder, learned about it because he used to work at Tesla. And as part of being in the Tesla finance ops team, part of the problem with leasing Teslas was this whole process of getting all the payments back. And it was all done very manually with all these business operation units that were outsourced. And he thought, oh, why not build an AI agent for that? So that was really good. And that ended up working really well. And now they're servicing a bunch of large banks. That's a good one. Another one is, which I also mentioned previously in another episode, is the founders of Diode Computer. They're basically building the AI circuit board copilot. And their insight is that both of the founders were electrical engineers, but also software engineers. So that turned out to be a very unique gap, intersection of skills that has not been, is not that the world of software and hardware don't tend to talk to each other. So they had built circuits at Apple, at startups, and even custom processors. So they had a lot of experience building very high-end electronics. And they saw the gap and frustration of working with hardware engineers, and why didn't they do things like software? Why didn't the world of Git exist? Why didn't the world of, I have to parse through all these data sheets to verify all the components manually? That's the electrical engineer job. Why do I have to do that manually? Why not just get an LLM to parse and do all that verification, like in code, with like QA, right? So that was the insight. And that was so unique to them because they were the only ones that had the unique experience of one of them being a super strong software engineer and the other one being super strong at hardware. I think this is one of those cases where starting a startup that actually turns out to be successful requires you to be similar to a PhD or postdoc researcher in some sense, where you have to go all the way to the edge of what human beings know and understand. And then instead of like, you know, publishing research that, you know, sort of pokes that edge out a little bit. Instead, it's like you're creating a product or service that people really want. To that point of being the PhD level of expert in that world, when founders get into this, they have such a unique fit with founder market fit. They're the best in the world. Literally, there's no one like them that had that work experience and that happened to want to do a startup that happened to now be really interested in AI. And there's this moment in time that is only N equals one. That's only them that can do it, which is cool. Yeah, Ari being in a place like Tesla is always very interesting to think about because, you know, there is a very important question that all founders have to ask when they commit to an idea. And that is, if not us, then who? Do you wanna talk about Spur, Jared? Sure. So Spur is building an AI QA agent. So the way testing works now is if you have a large company, you probably have QA engineers who write tests to test your software, and they're just building an AI agent that writes the test for you. And the way they came up with the idea is one of the founders worked at Figma, which has a notoriously complex front end that's very hard to test. And she realized that the engineers were spending a ton of time testing the front end and writing and maintaining tests for it. And that AI just obviously would enable you to automate a lot of that work. I guess Figma is a really good place to sort of come out of in that if you're already at the edge of design and collaboration, plus this AI thing sort of happens, you already have exposure to the right customers and know what the people at the edge are doing because they're at the edge. I have a kind of crazier one, which might be heartening to some of the people in the audience, in that this is probably my youngest team I've ever funded. How old are they? They were 19. They dropped out of freshman year at University of Waterloo. It's a company called Data Curve. And actually, one of a few pivots, they actually came in as a company called Uncle GPT. And it was sort of this toy hackathon idea, really. Like, I think they literally won a hackathon with the idea. You're sort of standard chat GPT wrapper back when wrapper was the pejorative that everyone was saying. But the deeper problem was people didn't really want it. It was a really cool demo, but there weren't customers that were willing to pay for it and use it all the time. And then during the batch, they actually became AI for product managers to go back to what we were saying earlier. It's sort of, that again is a not leaving the house enough sort of idea for them because neither of them as 19-year-olds had been product managers. So it's actually very, very hard to make software or products for people and teams where you actually don't have direct experience or knowledge of it. And so luckily, and this is maybe a really good example of looking back within, the founder actually, she actually was an intern for Cohere, which was all the way out on the edge in terms of LLMs and Cogen. And so she had already been working on data tools and producing synthetic and real data for large language models for Cohere. And she went back to her old boss and they said, hey, this is what we need. And she said, oh, well, I could build that. And so now she's basically off and running. I mean, she had a great demo day and then she's making mid to high seven figures for a company that just started June of last year. I have noticed this pattern with a lot of startup founders. Whenever I have a team in the batch that's pivoting and they do the office hours where they're like, you know, lost confidence in my old idea. What should I go work on now? Like the first note in my decision tree when I'm trying to like help them find a new idea is like, are the founders experts in anything? Because if the founders are experts in anything, then like often that's the place to look for ideas first. And the thing that I've noticed is that it's often surprisingly hard for the founders to actually know what they themselves are experts in. And sometimes you kind of have to pull out of them their actual areas of expertise. This is why it's extra hard for 19 year olds. But at the same time, that's part of the reason why I really love this example. That founder just had to reach back into her internship from the prior summer and there was something lying in plain sight there. I think what you're saying about that, I've seen it a lot. I think a lot of times when founders come in into these office hours with us, it's almost a bit of a allergic reaction to what they were doing. By definition, it's because they were experts on work and grinded out years and years. They're like, oh, I don't wanna do another decade on this thing that I put all this time. It's like, this is so boring. And they wanna chase some shiny object that they don't know anything about. Sort of the grass is greener, but then they sound so much smarter when they talk about that particular domain. And then when you kind of reflect it back to them, they're like, oh yeah, you're right. It's like, I never heard anyone go so deep into this as what you eloquently have said versus going over the shiny idea, which is very surface level of insight. And internships is another interesting meta point here. I mean, some huge percentage of YC's billion dollar companies can be traced directly back to not just a job, but specifically an internship that one of the founders had. And so maybe like a meta point is like, if you're like in college and you want to be in a place to have good startup ideas, like do internships at like really cool companies that are on the bleeding edge of something. Cause that's like a really like tried and true path to get you a great startup idea. I think the other meta point is also being picky at where you end up working. I mean, the example with the founder of Data Curve working at Cohere. Cohere is at the bleeding edge. The founder of Cohere was one of the authors of the All Attention You Can Need paper, which is the seminal paper that pretty much created this whole AI boom now. She was working there. Another good example I have is this other company called David AI. The founders were working at Scale and Scale is at the bleeding edge of providing all data sets for right now the AI boom as well. And David AI found this niche where Scale wasn't going into, which is the scarcity with data sets around multimodal data with speaker separated audio and going deeper into that because Scale got very deep into more the LLM world. So that turned out to be good. It's kind of to your same point kind of Cohere and here in this case with having worked at Cohere slash now Scale, working on the bleeding edge, you get to find high quality problems that are gonna be the future. So that's not the only way to look within. Maybe the one that people really look to and is a little bit more obvious is what are things that you wanna see in the world that you could see yourself just working on for the rest of your life? You know, there'll be dragons for this, but on the other hand, we have some really noteworthy examples of companies that have really found something and have made something people want. I have one story in particular that I'd share. This story really like stuck with me. It's about a company called Can of Soup. We funded Can of Soup and the founder, Gabriel, had been an early engineer at Substack, which is a company we funded years earlier. And very early on, he like lost confidence in the idea that we had funded him for. And then he kind of wandered in the wilderness in sort of a pivot hell period where he was like trying to come up with a new idea in sort of an artificial way, as often happens when founders are pivoting. And he was looking at these various like B2B SaaS ideas that were all totally plausible, but he just wasn't really excited about any of them. And he went on a walk with his old boss, Chris, the CEO of Substack. And Chris gave him a piece of advice that's really stuck with me. Gabriel pitched him one of these B2B SaaS ideas and Chris was like, who cares? Work on something that captures the human imagination. And that was the prompt that got Gabriel to just start thinking about like a much bigger idea that he would actually be excited to work on for like a really long time. And that's what sort of led him down the path of coming up with Can of Soup, which is this like AI Instagram-like thing that's like a totally new kind of social network. And it's a really big, crazy, ambitious idea. We don't know if it's gonna work yet, but it's like super interesting and like so much cooler than the sort of like manufactured B2B SaaS ideas. I mean, social networks seem like it's pretty ripe for things that people really wanna work on. One of my favorite AI companies right now is called Happenstance. The founder was a Apple AI researcher, sold his last startup. And then he started realizing, especially once like Word2Vec and vector databases started coming out, that when you use things like LinkedIn, how often is it that you're like typing something that you're looking for? And it just, I think it's just still using plain old plain text search. Like I think it's just using indices from MySQL for all we know. It's literally not smart. And the thing about LLMs and especially, LLMs plus vector search now mean that the search engine itself can be so much more intelligent. And so, I'm always trying to connect people in the batch to people who could buy their thing or who could help them with access or all of that. And then Happenstance now is just this wild thing where I can type almost anything more or less in a fuzzy way. I can even describe the people I'm trying to help. I can even describe sort of the level or area inside the company I think I want to connect this founder to. And it'll just figure all of that stuff out. It'll write the SQL queries and then use its own, a mix of vector search, LLMs and SQL to find those people in a way that like LinkedIn search just fails 10 times out of 10 for some of these really complex queries. Part of our job is kind of helping founders to think bigger because the whole process of starting a startup is already scary. And sometimes founders start with a very small idea that could be inconsequential, but if you 10x it, then how could the world change? And I think Jared, you had some really good example for this one. I do. And incidentally, if you're looking for a startup idea, one thing you should definitely do is you should go and read or reread Paul Graham's essay called How to Get Startup Ideas, which is really kind of a definitive work on the topic. And he talks about this concept called blinders, where if you're looking for a startup idea, you tend to have blinders on where your subconscious doesn't even allow you to see certain ideas because they seem too ambitious and too scary. And so they don't even consciously double to the surface for you to be able to like decide if you want to work on them or not. And a great example I have of this is a company called EasyDubs. EasyDubs is building the universal translator like from Star Trek. So imagine you go to Japan, but you don't speak Japanese and you want to have a conversation with someone who only speaks Japanese, you can use EasyDubs and it'll translate like simultaneously and in real time. So you can have a real time conversation with somebody who speaks a different language. So one of the common things people run into when they go through this path of idea is just what we said before, like I really don't have any expertise. I've mined everything I can. I've mined all of my experiences and I can't generate a good startup idea that way. And that then takes you to like Gary's point around like you have to get outside of the house and you have to start putting in the work to like build the expertise. And so I feel like actually our advice to startups during the batch when they're going through this path will often change. It's like, stop thinking about kind of what your like two week revenue goals are and start treating yourself as researchers and just try and like build expertise in something in the hope that you find startup ideas. I have a story on that type of way of generating idea. It's a company called Egress Health and they spent a while pivoting and trying to find an idea weren't kind of landing on anything that worked really well. And so I think one of their parents I think it was one of the founder's mothers like a dentist who ran her own small dentist office and he just went to work with her for a day just to kind of see like, how does a dentist office work? And like, is there anything that software could do better? And he realized a lot of the admin work involved in sort of insurance, processing someone's insurance and pre-authorizing them. And all of this work was just like routine that could really be processed away by an LLM. And so they just started working on that. They started building an LLM powered back office for dentists and it's working like really, really well. That's so cool. Yeah. I love it when founders end up in this branch of the decision tree to find a startup idea because through them, I get to learn about all these corners of the world and like think about where there might be interesting problems. There's a couple parts about that story that I want to kind of pull out hard. One is like using family connections, like a lot of our best startup ideas, like a lot of YC's like billion dollar startups, literally it was like the founder's parent or uncle or cousin or brother or some like old college roommate or just some like random connection that was like just enough of an opening to like lead them to an interesting place. It's surprising how important that is. You know, basically you could cold email a thousand people sometimes and get literally zero responses. But if you have someone who you're going to see every Thanksgiving, I think they're going to give you some access and access is all you need actually. Like that's sometimes like right at the beginning moment, you just need access into and some underserved place, some place where no good software engineer or AI engineer has ever seen or has seen yet. Like it works now better than ever because what we said in our previous episode about just how AI agents are going to be so much bigger than SaaS. I think take the salient example, the egress health example, probably five years ago, building software for just like car loan lenders or just for like a small dentist practice was probably not a big enough opportunity by itself. And so people probably felt like my connections and my expertise are not that valuable. But now it's like any one of these things with AI, it's so much more valuable than just building like a CRM for a dentist's office. Do you like actually replacing a human? The human's probably being paid like 60 to 80,000 bucks minimum like per year. And so just the value of your software just went way, way, way up. And I also love that they actually went on site and spent a day in the dentist's office, just being like a fly on the wall. Like anytime you can find anyone working in any industry who will let you do that is fricking gold. Like you're going to discover something cool. I think it's this concept of going undercover as an undercover secret agent to learn all the deep secrets about a industry which is all kept kind of behind closed door for good reasons outside of the outsiders. But because you have this special connection with a family member or someone like that, or sometimes founders are very charming and they get in through that as well. I've had one example like that. And you can kind of learn a lot about these esoteric industries. One example is this company called Happy Robot. They're basically building AI agents for coordinating logistic for truckers. They don't come from trucking, such an esoteric world for them. The founders are Spanish, which is like very far removed from this and PhD students. And the way they landed into that is just the founders are very personable. They're very friendly. And when you talk to them, you want to be friends with them. Well, the good news is even if you aren't connected by family or friends, you might also not be friendly and extroverted enough to make it work the way Happy Robot did. There is still another way. And I'm gonna not put this company on blast because they are an AI billing company that is doing well. But the way they came to the idea was not through connections per se. One of the co-founders actually got a job doing medical billing as a biller, as a remote person for a New York-based automatrist office. And he did not actually disclose that he was using software or building software, but that's what he did. He got a job. It was like an undercover job. Like it wasn't like he happened to be working as a medical biller. He was like, I want to automate medical billing, but in order to do that, I need to understand how it works. So I'm gonna get a job as a medical biller in order to understand how it works from the inside. Am I saying it right? Exactly. He actually got a real job and was paid as a medical biller. I have a founder who did the same thing in a different industry. It's wild, right? It's totally wild. But it works if you don't have connections and you can't walk in and sweet talk people to get access. There are just jobs that are knowledge work jobs that you can do. And then this is actually my pitch sometimes to regulators that open source is actually a very important piece of this. Because the reason why this person was able to do it legally was he was building his own software to automate the work all locally on his own computer. So he was like building his own AI agent robot using Llama 3 to replace himself on two MacBook Pros at the time. And there was no violation, no laws were broken. It's legal to use your own computer and use Zoom and use those things to actually go and work with an external party's thing because it was a remote laptop job and you could do that really easily with it. So I think that was one of the crazier examples and it sounds like that's something that- More founders should totally do this. They should totally just go get random jobs, like working at random industries and learn about them from the inside. It doesn't take that long. It's not like you have to get like an MD or something in order to become a medical biller. I think it's like a two to four week training program or something like that. And those are actually sort of the ideal things to get automated right now. Sort of, you know, laptop, remote laptop jobs that you can get very easily. It turns out LLMs are very good at doing those jobs these days. And laptops are very powerful and, you know, synthetic data down to lower, you know, smaller parameter models is also very good. So this is kind of the golden age of truly, truly going undercover. I think one of the themes here is how do you go all the way to the edge of, you know, what people know, especially if you yourself are an engineer or AI engineer. And this one was a kind of intense one. So AblePolice actually does work with police departments. And the reason why that found, the founder is actually growing Daniel on Twitter. So the way he found out about this problem was actually, you know, pretty serious. Like one of his friends was the victim of a crime. He started doing research and discovered, as I discovered in San Francisco, that many police officers are actually just drowned in paperwork. You know, you might do an eight, 10 hour shift, and then you're spending two or three hours of those eight hours to 10 hours just filling out paperwork at the end of it. San Francisco has this crazy law where if you stop anyone to even talk to them, you have to fill out as much paperwork as if you had arrested them. So how can you do police work when you have a police commission that drowns you in paperwork like that? And this is not merely a San Francisco thing. This is like almost all over the United States. And so, you know, this is a very good example of literally going, you know, undercover in that he went on ride-alongs, he investigated this thing that was very upsetting to him in society, discovered the root cause, and then LLMs were happening. He's like, why would anyone sit in front of a web browser filling out multiple hours of like click, click, click, like enter data, literally like, you know, transcribe people's driver's licenses. Like this is, you know, why are we turning police officers who are supposed to look out for public safety and a lot of the job is clerical work? And, you know, of course you can use LLMs plus computer vision to take that two hours or three hours down to five or 10 minutes, especially because you already have camera data all day from the person doing the work. Can I leak a alpha trick about the previous topic, finding jobs? This is something that I suggested that Founder do. You can literally go to indeed.com and search for jobs that have keywords like this, like remote analyst, remote like clerk, things like that. And just look at all of the jobs that people are hiring for. Like some of them are probably weird jobs that like most people haven't even heard of. There's all of these jobs like this out there. You just go get one of those jobs. There's another trick. Another trick is if you don't want to work the boring job yourself, is think of if you have any friends who have very boring jobs and go follow them and shatter them at work for a day. So I have a company- You have an example like this where this actually happened, right? Yeah, it's a company called Sweet Spot. And when we accepted them into YC, I think the idea was something about payments for taco trucks or something totally random. And then they started searching for ideas and they had a friend who just worked. His full-time job was to sit in front of a computer on the government website, looking at the government posts like contracts are available for bidding. And so his whole job was just to keep refreshing that page. And every time there was like a bid that was relevant for the company he worked at, he just like pasted the link somewhere. And they were like, well, that seems like something that could definitely be like automated with AI. And this was about a year and a half coming up to almost two years ago now. And that insight actually turned into a really exciting idea where now it's a AI platform for just all of your government contracting and procurement. That they both like find the thing for you, they generate like all of your bid, they give you advice on how to like sharpen it to have the best chance of landing it and price it optimally. All of these things like a single package and it's growing like incredibly fast. I think that category of indeed jobs is also this jobs that AI is very good at automating that are temporary. I think Jared, you work with one of those companies. Any category of jobs that is being outsourced to a low wage country is like a strong signal that there is a startup to be built right now in the current era. Like that's just like an amazing place to look. One company like that, that I worked with last year is a company called Lilac Labs. And what Lilac Labs is doing is they are automating the person at the drive-thru who takes orders. They were snooping around other startup ideas that didn't end up panning out. But in the process of looking at the startup ideas, they realized that for a lot of the drive-thrus in America, when you drive up to the drive-thru and you like place an order, the person on the other end who's like listening to your order and like transcribing it into the point of sale system lives on the other side of the world. Because that job which was always done by the person sitting at the drive-thru has now been outsourced to like DPOs in like low wage countries. And that was a clue to them that like this was an amazing like target to go after. You have one too, Diana, right? It's a little bit different flavor. So this is more for looking into spaces where there's a product that is not 100% working and there's a bunch of consultants that are making a bunch of money to just get that product to work. And one example is UiPath. It's this giant company that went IPO that does robotic process automation. Maybe a lot of the audience. It's macros. A lot of macros to automate workflows on desktops that enterprise buy it. But in order to get UiPath live, it's just so much work. It takes like consultants and consultants that need to be certified. What these founders found is like, hey, what if we actually build a way better product that actually works and actually does RPA without the need of expensive consultants? So this is a company that a founder called Automatt that basically is this solution, but a lot better that actually works and it's only possible now because of AI. Was this in part inspired by like recent advances in getting LLMs to browse the web and use desktop applications? Totally. I mean, they actually went through YC before computer use was launched. But the interesting thing about this team, they have always been living at the edge of technology. They actually were ex-Googlers that got access to BART. It used to be called BART, now it's called Gemini. So they actually prototyped a bunch of apps on AI before it was even cool. So this is during the pandemic and before it. So it had a lot of experience building with AI before everyone did. And they saw this is a place that's gonna go. I think that's a cool meta lesson. Like PG has this line in his essay, live at the edge and notice what's missing. You're in a way better position to have a great AI startup idea if you're constantly trying the latest stuff and like actually personally consuming and developing on the very latest stuff because then you're one of the first people to realize that something new is possible. And also if your friends are doing the same. Like actually, I think this is an example if you have friends who are working at interesting companies or companies themselves that are pushing things on the edge, it can be really valuable. I have another story of a company that searched and pivoted an idea during the batch called PrayDB. And they had friends who had sold their startup to another startup. That was just, this is just during the era where sort of Pinecone and vector databases were becoming more and more popular because AI was just taking off. This is around like late 2022, early 2023. Their friend told them, hey, like the big problem we have is like, we want like a good quality real-time sync between like our Postgres database and Pinecone. And no one's really built that. And we're gonna have to build it all internally in-house. So PrayDB found us, okay, well, we could just build that for you. But then as they actually got deeper into the idea, they realized, oh, well, like PG vector is relatively new extension within Postgres that you can actually just do a lot of the Pinecone, the stuff that you want to do on Pinecone, you can actually probably just do in Postgres. And so they just started pushing the limits of PG vector to see how much of the functionality they could replicate. And they got like surprisingly far. And now they have sort of enterprise customers that are using them, not just as sort of a replacement for a separate vector DB, but also just as a replacement for elastic and various things for search and semantic search across all of the stuff in their database. And I think that's, they got that idea just because they were in the world of technical founders doing things and building things and telling them, oh, hey, like, this is our issue. You can just do things. Yeah, you can just do things, good advice. And to that point, I think the thing that makes this category of looking outside in is to hang out with very smart people. Like your example, I have another one, is this company called Reducto. And because they were in YC, they got to be friends with a lot of other founders that are building at the edge of AI. And they found an interesting problem for a lot of the rack applications in order to get them to work well. You need to be able to extract the chunks on a very well, but this kind of problem, you would only find out if you end up working with the top percentile of builders are really at the edge of really building the next generation of applications. So that's how they found the idea of Reducto, which extracts perfect chunks and data from PDFs. Another way I've seen people build the expertise to get an idea, I wouldn't even call it quite expertise, but this is something, just getting going and building any product makes you an expert in whoever the users of it are and puts you in a good spot to find ideas. And an expert in building products like that. So you build expertise in multiple areas. Wait, I can't just become an expert from being an ex-influencer? You're saying I have to go talk to users. You actually have to build something. You have to ship code. A company I went through that I worked with twice in, yeah, 2022 again, so almost three years ago, searching for ideas. The idea they landed on was a freelancer marketplace. And honestly, they didn't have, the company's called Juicebox. They didn't really have much differentiation for that idea, but they were just excited about it. They'd done some freelance work themselves. They just wanted to start building it, go ahead. And they worked on that idea for a decent chunk of time and it didn't really take off. But in the course of sort of working on freelancer marketplace and working with companies who are hiring freelancers and hiring other people as the LLMs developed, they realized, oh, there's actually like a real need for just like LLM-powered people search for recruiters specifically. And they started building that. And that has really taken off. It's called PeopleGPT. And it's just like the really effective search, especially for recruiting teams who just find exactly who they want, sort of a fuzzy prompt. And it will just give them a list of all the dream candidates to go and ping. And you can see how like they wouldn't, they're not, they've never run, they've never been recruiters themselves. They never actually really hired people, but the expertise they built was because they just launched something. There's this funny thing during the batch, especially around fundraising, which is a very interesting phenomenon. I'm sure you guys have seen this too, where we have founders who are out there, they ship a product, they're talking to real users. Those users turn around and give them their credit card number or sign on the dotted line, like big enterprise contracts for 10 or $100,000 a year. And then fundraising rolls around, they start getting the first nos, and then they get, it's like just getting gut punch, but gut punches after gut punches, and they come back in office hours. And they're like, investors don't get it. And the thing that I find myself saying over and over again is like, yeah, investors don't get it because they're trying to do it the way like a founder would trying to be an ex-influencer, trying to understand, just reading feeds from like literally their toilet, like in shit posting, right? Like literally, that's not how you figure out what's going on. Why would you, the person who's outside of the house, not on the toilet, outside of the house, out there talking to people, shipping software and doing things, like why would you be taking any cues at all from the person who's still sitting on the toilet, like scrolling an ex-feed? Like it doesn't make any sense, right? You have direct knowledge of the world out there and you're coming back into Plato's cave. And this person is like saying, well, I don't see the shadows yet on the back wall. And it's like, let me tell you it's out there, right? Like you literally have seen it with your own eyes. Here's another instance that I've seen where founders like psych themselves out, Kerry, that's related to that, which is founders who psych themselves out because spaces seem too competitive. And they end up like shying away from going after ideas that are actually really good because like two competitors launched on TechCrunch and like raised seed funding or something like that. Harjit, you have a good example of a company. Yeah, it's a company both Kerry and I worked with, actually we mentioned before, like GigaML. They originally applied with a tech idea. It was an idea to help Indian college students apply to US colleges, Indian high school students apply to US colleges. Then they pivoted into fine tuning as a service so around the time where that was just like the open source models had just been released, they couldn't build a sustainable business. So they didn't quite crack that nut. And they were looking for applications of the fact they had become experts in fine tuning like models for specific purposes. And they were trying to find a vertical application. And the one that they were most excited about was customer support. But they felt that it was very crowded. Like there were lots of people doing customer support, but they went for it anyway. And specifically they really focused on one company, Zepto, which has really been willing to be like a real early cutting edge adopter. I mean, another meta point here is like Zepto themselves really want to be like the most operationally efficient delivery company in the world. So they were looking for these really like high quality pieces of tech. Rumors of IPO-ing later this year. Who knows? We don't know. What I will say about the GigaML founders is they are incredibly smart engineers and not natural sales people at all. And just one thing I'm used to pre-AI, I'm curious if you have opinions on this, is it often feels, especially with B2B SaaS things, like if you're entering a crowded space, it's often as much about how can you differentiate on just sales versus like necessarily like your first product. And so you gravitate towards, ah, like I need to feel that this team can really like sell in order to get anywhere if they're gonna launch, you know, a new payroll product, for example. But with GigaML, what I noticed is that so many of the things just actually don't work very well. Doing like AI that can really replace your customer support team of humans is a hard technical problem. And so although lots of people are pitching that they have it, very, very few people can actually deliver to the level that customers want. And it just turned out that GigaML were like, their technical strength meant that they could actually deliver what no one else could. And that got them this deal and now it's sort of snowballing from there. It's like a huge enterprise deal. Did they close up to during the batch? No, I mean, they were pivoting for quite a while, actually. This is one of those stories of where I think it took them about a year, if I don't recall, to find the right idea. And that's quite normal, actually. It's quite normal, actually, yeah. A lot of our best companies in recent years have been that, yeah. Yeah, which flies in the face of, I think, what everyone believed, you know, ancient history five to 10 years ago. Like, you know, it's hard to believe, but 10 years ago, there were entire seed funds that would say, like, we never do seed extensions. Either you're going to be great and you're great immediately, or, you know, seed extensions are a sucker bet. And these days, I'm pretty glad that we're in the days where that's just not true anymore. Like, you can see that people are getting product market fit. My speculation on it is AI moves so quickly that every few months, there's just a new set of possible ideas that are generated. I also, this is like a more anecdotal thing, I also feel just like, because it's so exciting to work on a startup and work on AI right now, that the teams just have morale reserves for longer. Like, it's like, why would I, if you're building, if you're right now, if you're building a startup, working on like cutting edge AI, even if you haven't found the right idea yet, like, why give up and go back to Google or college or something? Yeah, there's a high probability that your lucky break is just around the corner. Yeah, and this is kind of the most exciting thing you could be doing with your time. There's all these product releases that keep changing the whole space, the whole time, which is crazy. Well, that's all the time we have for now, but I think that that's a pretty great thing for everyone out there to keep in their mind. You know, you can't stay in your house or sit on your toilet, you know, doom scrolling X. You have to either look very deeply within you to find that you're already on the edge because of something you've done, or you need to radically get out of the house, go into, you know, other people's real businesses and the real problems that humanity faces, and then get first principles understanding of what's going on out there. And then you can build a billion dollar business using AI. We'll see you guys next time.