Pod Insights Lightcone Podcast Startups Ideas You Can Now Build With AI About this episode There's never been a better time to start an AI company. Not just because there are new ideas, but because the tech finally makes old ones actually work. On the Lightcone, Garry, Harj, Diana, and Jared talk through the kinds of startups that are suddenly viable thanks to LLMs—from full-stack law firms to personalized tutors to recruiting platforms that can finally scale. They share the patterns they're seeing, the ideas they're excited about, and what it means to live at the edge of the future, where breakthroughs often look like second chances. If you've been waiting for the right moment to build, this is it.
Transcript Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.
0:00 Harj Taggar? There's all this tooling and infrastructure still to build. There's clearly still a bunch of startups yet to be built in just the infrastructure space around deploying AI and using agents.
0:12 Jared Friedman? If you're living at the edge of the future and you're exploring the latest technology, there's so many great startup ideas, you're very likely to just bump into one.
0:19 Garry Tan? You apply the right prompts and the right data set and a little bit of ingenuity, the right evals, a little bit of taste, and you can get just magical output.
0:39 Garry Tan? Welcome back to another episode of The Light Cone. Every other week, we're certainly realizing there's a new capability, a million token context window in Gemini 2.5 Pro. It's just really insane right now. And the thing to take away from that, though, is that we have an incredible number of new startup ideas, some of which are actually very old, and they can only happen right now. Harj, what are some of the things you're seeing?
1:06 Harj Taggar? Well, one thing I've been thinking a lot about recently is what are types of startup ideas that couldn't work before AI or didn't work particularly well that are now able to work really, really well? And one idea that is very personal to me would be recruiting startups, since I ran a recruiting startup, Triplebyte, for almost five years. And I think something that I've clearly seen is that there was a period of time when we started Triplebyte, it's around 2015, where recruiting startups were kind of like a really popular type of startup. And I think a lot of the excitement around those ideas back then was this idea of applying marketplace models to recruiting, because there were marketplaces for everything except how to hire great people, and specifically, great engineers.
1:49 Harj Taggar? And we started Triplebyte with the thesis of You don't just want to let anyone on your marketplace. You want to build a really curated marketplace that evaluates engineers and gives people lots of data about who the best engineers are. And this was all pre L L M. So we had to spend years essentially building our own software to do thousands of technical interviews, to squeeze out every little data point we could from a technical interview, so that we'd effectively built up this labeled data set that we could run machine learning models on. But we didn't even get to do that until like years three or four.
2:21 Garry Tan? And initially, it was actually a three-sided marketplace in that you needed to hire an interviewer in between to get that human signal.
2:29 Harj Taggar? We had companies hiring engineers. We had the engineers looking for jobs. And then we had engineers we contracted to interview the engineers. So it was like lots of things going on right now. And all of the evaluation piece of it, at least now with A I, is very, very possible. I mean, specifically with the A I code gen models, you can do code evaluation. And I think probably one of the hot A I startups at the moment is this company called Merckor, which is essentially similar to the Triple White idea. It's a marketplace for hiring software engineers. But I think what A I has unlocked for them is the evaluation piece of it. They could just do on day one using L L Ms that they need to build up this big label data set.
3:14 Harj Taggar? And they've been able to expand into other types of knowledge work quite easily. For us to have gone from like engineers to analysts to all these other things would have taken years because again, we had to rebuild the label data set. But with L L Ms, you can just do that on, you know, day one effectively. And so I think this whole, this whole class of recruiting startups that are trying to evaluate humans at being good at specific tasks or not is a really interesting space. That's much more exciting to find good startup ideas in now than it was five years ago.
3:44 Garry Tan? So that's a very powerful prompt for anyone listening. What are marketplaces that are three-sided or four-sided marketplaces that suddenly become, you know, two or three-sided, or now there are two-sided marketplaces like Duolingo that are, you know, a little bit under fire because they're sort of starting to say, actually, And maybe we're just gonna use AI for the person that you're gonna talk to in another language. That is totally a coherent thing that you could go to almost any marketplace in the world and say, what if? What what will LLMs do in that marketplace?
4:19 Harj Taggar? I think the other thing I really respect the Merkel founders for is there's also just a psychological element as a founder to when you enter into a space where there's been lots of smart teams and lots of capital that's flown into it. This was definitely with recruiting startups. I mean, Triplebyte raised something like $50 million. Our main competitor, Hyatt, raised over $100 million. I think in aggregate, hundreds of millions of dollars went into funding recruiting marketplace companies. And overall, as a category, did not do particularly well. And so I think going in, you face a lot of skepticism if you're gonna go out and pitch investors for an idea, even when you have the like, well, LLMs change everything. That pitch two years ago was still not as compelling as it is today. And so you have to be willing to sort of
5:00 Harj Taggar? push through a lot of sort of cynicism and people who are burnt out, who have lost lots of money on an idea to even kind of keep going to test it out and and make it work.
5:09 Garry Tan? That's something that repeats actually all the time. I mean, Instacart was that story exactly. Like, Webvan was sort of this rotting corpse of a startup, just hanging in that doorway. And most people looked at that and said, oh, man, I don't want to walk near that. Like, there's there's gonna be more. But, you know, simultaneously, the iPhone and Android phones were everywhere and you could have a mobile marketplace for the first time. And I guess that's why we're pretty excited about this moment, because suddenly, all, you know, the IDMAs just moved, like, all of the walls to the IDMAs have shifted around. And the only way to find out is you've got to actually be in the maze.
5:48 Jared Friedman? It is very similar to Instacart and Webvan, if we go back in history, right? Because like the the big technology unlock for Instacart was the fact that everyone had a phone now. Enabled like the Webvan model to actually work for the first time. And like, it's the same thing with LLMs and recruiting companies now, and a whole bunch of other ideas.
6:06 Diana Hu? I think it makes focusing even on specific parts of the marketplace. To be great ideas to start with, even with this recruiting idea space. This is a company called Apreora that Nico, the other GP here at YC, funded back in winter 24. And their whole premise is to build AI agents that run the screening for technical interviews, where a lot of engineers spend a lot of time just doing a bunch of interviews. And the pass rate is so tiny. When I used to run engineering teams at Niantic, all that pre-screening was just so much work.
6:42 Jared Friedman? The engineers hate doing it.
6:43 Diana Hu? Yeah. And even that one piece, not exactly, let's say, marketplace. So what is the hardest part of it? And if you solve it right now, it works out. So Apreora actually does a pretty good job. It's being used by large companies and it's been taking off.
6:58 Harj Taggar? It's another one where we can actually expand the market because I think there are plenty of technical screening products pre-Apreora. But you could only use them to do fairly simple evaluations to like weed out people who weren't engineers at all effectively or very, very junior. But Apreora's product now with LLMs, you can do more sophisticated evaluations to kind of get more nuanced levels of screening. And so suddenly now companies will be like, oh, actually, I could actually give this to not just like my international applicants or my college students. I'll just give it to like senior engineers who are applying, which just opens up the opportunity.
7:34 Diana Hu? So you were talking a bit about education as well, Gary, about Duolingo. I think that aspect of doing hyper-personalization is one of the holy grails where has been difficult for ad tech companies to crack, right? Because every student, as they go through their learning journey, everyone is very unique and knows different things. And it sounds really cool to build like the awesome personal AI tutor that Harch did at RFS for, right?
8:00 Harj Taggar? Yeah. The thing I'm excited about is for as long as I can remember the internet's been around, like one of the like dreams of it was that everyone now have access to like personalized learning and knowledge and we'd all just.
8:13 Harj Taggar? You know, have these like great intellectual tools to learn anything. And clearly the internet's made it easier to learn, but we've never had really truly personalized learning or personalized tutor in your pocket idea, which is possible now for the first time. And I think we're definitely seeing smart teams applying to Y C who are interested in building that type of product.
8:32 Diana Hu? Couple of companies that we funded that are kind of working out is this other company that also Nico funded called Revision Dojo that helps students do exam prep and is sort of the version of flashcards, but not like the janky, just like boring going through content, but the version that actually students like and gets tailored for their journey. And that one has like a lot of DAUs and a lot of power users, which is super interesting. And I think Jared, you had worked with this other company called Edexia as well?
9:04 Jared Friedman? Yeah, Edexia does tools for teachers to grade their assignments, which is another example of work that like is like not people's main job, but is this other thing that they have to do, like engineers doing like recruiting that they generally hate doing. There's like a lot of studies that show that like the biggest reason that teachers churn out of the workforce is that they hate grading assignments. It's just like no fun at all. And so Edexia like is an agent that's like very good at helping teachers to grade assignments.
9:33 Garry Tan? Yeah, one of the interesting trends for some of this stuff is that it's private schools who are actually much more nimble. And, you know, I'd be curious what policy changes we need to make to actually support this in public schools because the public schools need it the most, actually.
9:48 Harj Taggar? I guess question for you, actually, Gary, I'm curious about this stuff is it's clearly possible to build much better products with L L M. So if we take the learning apps, for example, they can go far beyond anything you could do for personalized learning pre L L M's. But it doesn't necessarily mean that you instantly get more distribution, especially if you're going after the consumer market. So do you? How do you think that plays out? Do better products automatically get more distribution? Or will these startups have to work equally as hard to get distribution to be big companies as before?
10:21 Garry Tan? I guess one of the more awkward things that's still true is that, you know, intelligence is much cheaper. It's quite a bit cheaper than it was last year, but it's still enough that you have to charge for it probably. But that's something I would probably track. I mean, it seems clear that, you know, distillation from bigger models to smaller models is working. It seems clear that the mega giant models are teaching even the production model size of today to be smarter. The cost of intelligence is coming down quite significantly. So, you know, I know that we tease this sort of almost every other episode, but like consumer AI, it finally might be here soon.
11:05 Garry Tan? And I think the thing to track is, well, how smart is it such that like any given user incrementally only costs, I don't know, pennies or like 10 or 15 cents? Like then it becomes so cheap that you will just have intelligence for free. Maybe it'll be a return to the freemium model that we got used to during web 2.0. This idea that you could basically give away your product. And then for, you know, five or 10% of those users, there are things that they so want that, you know, you're going to sell them a $5 or $10 or $20 a month subscription.
11:41 Harj Taggar? That's basically what OpenAI is doing, right?
11:43 Garry Tan? Yeah.
11:43 Harj Taggar? Like that is like.
11:44 Garry Tan? Perplexity does it, OpenAI. You know, going back to education study with two Ds, they're doing it and they're seeing a lot of success. I mean, on average, the kids who use that actually get on grade level or, you know, can kind of go up even a couple of grade levels. Those are real outcomes for students. So, you know, right now you still got to pay for it, but maybe not for a while. And that's actually a really big unlock, you know, that that's the moment where you could have 100 million or a billion people using it. OpenAI might be furthest ahead with it, but the hope is that, you know, really thousands of apps like this start in, start coming out across all the different things you'll need. And that's something that.
12:30 Garry Tan? I know we'll keep saying it, like it's gonna happen.
12:32 Diana Hu? I mean, it's kind of happening already for EdTech. Speak, it's this company that got started a couple of years ago before LLMs were a thing at all. It was a team of researchers that really believed that you could personalize personalized language learning, which might have been a big contrarian back then because Duolingo seemed to be the game in town that was winning. And they really focused on really personalizing that whole language learning. And they got, they started taking off in Korea for a lot of learners that were trying to learn English. And when GPT-3 and 3.5, they were early adopters of it, started coming out. They saw that, wow, this is gonna be the moment.
13:12 Diana Hu? They doubled down and they've been on, on this trajectory now with lots of MAU's, EAU's that's really working out.
13:21 Harj Taggar? I think one thing, going back to the consumer thing that we haven't talked as much about.
13:26 Harj Taggar? We've seen a lot with the startups that are selling to enterprises or companies about how the budgets become so much bigger. When companies are willing, when companies stop thinking about you as software as a service, but they start thinking about you as replacing their customer support team or their analytics team or something like that. They'll just pay way, way more. So the same thing will apply in consumer, right? Like if you think about a personalized learning app, often EdTech companies struggle with who's actually the buyer and who's gonna pay. And if you go for like younger children, for example, it's like you've got to get the parents to pay. But the parents aren't gonna pay that much for an app that their kids like don't retain or complete like some sort of online course that they're disengaged with. But we know that parents would definitely pay for like human tutors and like, you know, that's like actually probably quite a big market.
14:11 Harj Taggar? And so if your app goes from being like a self-study course that doesn't get any completion to actually being on par with the best human math tutor for your 12-year-old, parents will pay a lot more for that. And so those, like, it's possible that like the product now just become, has a business model that you didn't have before. And that alone means you don't necessarily need millions of parents using it. But even a hundred thousand parents using it, paying you a significant amount, means you now have like a much bigger business than was possible before.
14:39 Garry Tan? Yeah, I feel like we have to talk about moats a little bit. I mean, it's pretty clear, a company like Speak or almost any of these other companies that could have durable revenue streams. Like, what you need is brand, you need switching costs. Sometimes it's integration with other technologies that are sort of surrounding that experience. Like, in a school, it'd probably be being connected to Clever, for instance. Like, login is, authentication is pretty obvious. So, yeah, I feel like Sam Altman has talked about this a bunch. You know, it's, it's not enough to drop AI in it. You know, you still have to actually build a business. I don't think OpenAI is necessarily, you know, out to get all the startups. Like, I actually think that on the API side, they very much hope that a lot of them do really, really well.
15:23 Garry Tan? And certainly, we want that, too.
15:26 Harj Taggar? They did just hire, like, the Instacart CEO as their CEO of application.
15:30 Garry Tan? That's right.
15:30 Harj Taggar? So it does kind of seem like they are definitely paying more attention to the application layer.
15:35 Garry Tan? That's right. Yeah, I mean, you'd be crazy not to, right? Like, by all accounts, OpenAI is highly likely to be a trillion-dollar company at some point. And, you know, as powerful as a Google or an Apple or any of them. The interesting thing right now is like, they're still on the come up. And then, if anything, the big tech platforms are actually still holding back a lot of the AI labs. And the most profound example of this is, why is Siri still so dumb? It makes no sense, right?
16:07 Harj Taggar? Totally.
16:08 Garry Tan? I mean, I think that points to something that we actually really need in tech today. We actually really need platform neutrality. So, in the same way, you know, 20, 30 years ago, there were all these fights about net neutrality. This idea that there should be one internet, that ISPs or big companies should not self-p preference their own content or the content of their partners. You know, that's what sort of unleashed this giant wave of really a free market on the internet. The other profound example of that is actually Windows. If, you know, if you open up Windows, you actually have to choose your browser. And then you also need to be able to choose which search engine you use.
16:50 Garry Tan? And these are things that, you know, the government did get involved in and said, "Hey, you, you know, you cannot self-p preference in this way." And, you know, if you remember the moment where Internet Explorer had a majority of web users, like, that could have been a moment where Google Couldn't have become what it became. So we actually have a history of the government coming in and saying, this should be a free market and for that free market to create choice and then therefore prosperity and abundance. And so I would argue like, you know, why doesn't this exist for voice on phones? Like you should be able to pick not, you shouldn't be forced to use Google Assistant. You shouldn't be forced to use Siri.
17:33 Garry Tan? You should be allowed to pick and you know, it's been many, many years of having to use a very, very dumb Siri.
17:40 Harj Taggar? On the moat topic, something I just find fascinating is I saw some numbers recently about how Gemini Pro models, like just their usage, particularly from consumers, is just an insignificant fraction of chat GPTs. And I think at Y C, we've been doing our own internal work, building agents and actually being at the cutting edge of a lot of the AI tools. And we found that Gemini 2.5 Pro is like as good and in some cases a better model than O3 for various tasks. That hasn't trickled down into public awareness yet, right?
18:12 Jared Friedman? Which is fascinating since Google already has all the users with their, with their phones.
18:16 Harj Taggar? And I don't think anyone would say OpenAI is not a startup anymore, but relative to Google, it essentially is. So there is clearly some sort of intangible mode around being the first in a space and sort of staking your claim as like the best product for a specific use case. And I feel like some of-
18:32 Garry Tan? And actually making it good.
18:34 Harj Taggar? Yeah, yeah, yeah, yeah. But at some point, maybe it doesn't even necessarily need to be like objectively the best. It just needs to be good enough.
18:41 Garry Tan? I mean, that's the bet that I think a lot of the big tech companies are trying and failing at. I mean, there's, Microsoft has a copilot built into Windows now that is still quite inferior to anything OpenAI puts out. Gemini itself is very, very good and I use it quite a lot. It's probably, I don't know, 40% of my agent, you know, sort of if I need to especially summarize YouTube videos, it's very, very good at that.
19:07 Diana Hu? For multimodal is really good.
19:09 Garry Tan? Yeah, a lot of the Gemini integrations into Gmail or, you know, Google Drive are not-
19:15 Jared Friedman? They're totally useless.
19:16 Garry Tan? Yeah. It's like, is there someone at the wheel over there? I don't get it, you know?
19:20 Diana Hu? I mean, I think that's even confusing for us is even using it as a developer, there's actually two different products. There's Gemini, where you can consume Gemini and Vertex Gemini. And I think they're like different orgs. I think it's suffering a little bit from being too big of a company and essentially shipping the org. There's like these two APIs you can consume to use Gemini. And we're like, why two? One is from DeepMind and the other one is from GCP?
19:47 Garry Tan? I think that comes from the culture of Google, though. I mean, there's definitely the sense that if two orgs are competing and fighting, normally in a normal org, you go up and in a functioning startup, for instance, you know, it goes up to some level and then ultimately the CEO or founders. And then they just say, okay, well, I see the points over here. I see the points over there. We're going this way. But, you know, having lots of friends from Google, it doesn't seem like that's the culture there. Like, there's a layer of VP and sort of management that is actually like, you guys just fight it out. And so then you ship the org.
20:21 Diana Hu? I think the crazy thing about Google, they probably should have won a lot of the experience of the best model. There's almost like, I don't know where all this Game of Thrones analogy could be used. They might be a little bit like Dennis Targaryen. Because they secretly have dragons. The dragons are the TPUs. And this is one of the reasons why I think they could be the one company that could get a lot of the cost of intelligence to be very low. And they also have the engineering to be able to do a cost-effective, large context windows. I think one of the reasons why the other labs haven't quite shipped as big of a context window is cost, actually.
21:03 Jared Friedman? Is it actually the hardware? Like, it's just like you can't actually do it without?
21:07 Diana Hu? I think you can do it, but I think it's just very expensive and not cost-effective. But I think they done it so well and they got TPUs, which I think is smart for Sam. If you saw his little announcement, he's still the CEO of compute, quote unquote. So I'm sure they're probably working on something around there, too.
21:25 Harj Taggar? This is just classic innovator's dilemma. It's like if Google replaced google.com with Gemini Pro, it would instantly presumably be like the number one. Chatbot LLM service in the world, but.
21:39 Jared Friedman? That it would give up eighty percent of its revenue.
21:42 Harj Taggar? You would probably need a pretty strong founder CEO to do that. It's the kind of thing I can imagine Zuck doing, right? Like being like, yeah, you just, you can't imagine a hired CEO is going to do that.
21:53 Diana Hu? He's done that. He renamed the company to Meta.
21:54 Harj Taggar? Yeah.
21:55 Garry Tan? Yeah, Meta has its own issues too. Like, I'm so surprised, you know, I mean, you have Meta's AI and WhatsApp, it's in the Blue app, it's everywhere. But, I mean, who actually uses it? I don't think any of us.
22:08 Harj Taggar? I started using the Meta AI in WhatsApp. It's very classic. It makes me feel like Zuck is clearly still in charge of product because I don't think anyone else would launch it that way. You just, you now you have an AI assistant that's just in all of your chats. And you sort of, it comes with a, you can just at it and they will just start talking in a group chat. And it feels quite invasive, actually, as it's.
22:30 Garry Tan? Well, it's not that smart and then it can't do anything.
22:33 Harj Taggar? Yeah.
22:33 Garry Tan? And then you have, I mean, most people are surprised that it's in there.
22:37 Harj Taggar? It's just, it's just like, it feels like having someone from Facebook just in your chats.
22:43 Harj Taggar? And so it's just like, it's, I don't know, remind me of like the original News Feed launch or something. It's just like the classic Meta style of like, this is sort of, I don't know, objectively optimal. Like, I'm sure people will love it.
22:55 Garry Tan? You need to add a little bit of design taste into these things. I mean, it blows my mind that I can go to the Blue app, which I still kind of, you know, it's probably people watching this are like, what the heck's the Blue app? This is like facebook.com, which maybe nobody uses anymore.
23:09 Diana Hu? It's very millennial.
23:09 Garry Tan? Yeah, but you know, you have this Meta AI and you ask it, hey, who are my friends? I'm going to Barcelona next week. Who are my friends in Barcelona? And it's like, sorry, as an AI, I actually don't have access to that. It's like, what, you know, what is the point of this?
23:24 Jared Friedman? Our partner Pete Cuman wrote this really great essay where he talked about the Gemini integration with Gmail. And he really like broke down in great detail why Google built this integration all wrong and how they should have built it. It's almost like he was a PM at Google. Oh, wait, he, he was a PM at Google.
23:46 Garry Tan? It was very profound in that one of the things he pointed out was that, you know, you have a system prompt and a user prompt. And. If you are actually going to empower your users, you actually allow your user to change the system prompt, which is the part that normally is like above, you know, to use the Venkatesh Rao's idea of like sort of the API line. It's sort of like the system prompt is actually what is exerted, it applies like sort of imposed upon the user. And so Gemini follows this very specific thing. I think the example is actually an email saying that Pete's going to be sick to me.
24:30 Garry Tan? He's like, sorry, I'm not going to be able to come in. And he asks the agent to write this letter and it's very formal. And of course it is because there's no way to change the tone. It's actually one of the best blog posts in that I think he had to vibe code the blog post itself because you can actually try the prompts yourself on that webpage.
24:50 Jared Friedman? Yeah, it's super cool. It's like a, it's in this like interactive
24:54 Garry Tan? Yeah.
24:55 Jared Friedman? template language.
24:56 Garry Tan? Which made me think it's time to start an AI-first vibe coding blog platform.
25:02 Jared Friedman? Oh, like an AI, like a, like an AI posturist?
25:04 Garry Tan? Yeah, basically.
25:05 Jared Friedman? Is it time for posturist 2.0?
25:07 Garry Tan? Yeah, with all my extra time, that's what I'm going to work on. But that's a free idea for anyone's watching. We'll fund it.
25:14 Jared Friedman? There's another class of startup ideas that I'm, I'm particularly excited about that I think are like, perhaps the time is now, which is, do you guys remember the tech-enabled services wave?
25:26 Garry Tan? Yep.
25:27 Jared Friedman? Yeah, so for folks who, who, who didn't follow this, in the, in the 2010s, there was this huge boom in companies called tech-enabled services. Triplebyte was one, actually.
25:38 Garry Tan? Yeah, very much.
25:38 Jared Friedman? That was like tech-enabled services for recruiting.
25:40 Garry Tan? Yeah.
25:41 Jared Friedman? Right? We also had Atrium, which was tech-enabled services for. Law firms.
25:46 Harj Taggar? It started with Balaji's blog post about full stack startups, if you remember. Like, the concept was just that software eats the world means software just kind of goes into the real world. And so this is not the success example, but an example of it was, hey, like, instead of just having an app to deliver food, you should also like have a kitchen that cooks the food and software to optimize the kitchen and you just do everything. And that like the full stack startups in theory would be more valuable than just the software startups because they would do everything.
26:17 Jared Friedman? Yeah, because instead of just selling like software to like the restaurants and capturing like 10%, you could just own the restaurant, you could capture 100%.
26:24 Harj Taggar? This is exactly what Tripwire was because we were like, we're going to be a recruiting agency effectively. We're not selling software to recruiting agency. We're actually just doing the whole thing. Like, we're gonna, we also had recruiters on staff that were just there to help people negotiate salaries and match them to the right companies. And yeah, it was very much in that wave of do everything.
26:41 Garry Tan? But that wave of startups generally forgot that you need gross margins.
26:45 Jared Friedman? Yeah. What happened? Like, I mean, like, fast forward, basically, the short version is like, it didn't really work. And the full stack startups actually were not more valuable than the SaaS companies. And the SaaS companies sort of like won that round of the like, derwining competition of different business models.
26:59 Harj Taggar? I think fundamentally, it's just what Gary says, it's just they were actually not great gross margin businesses. But it was actually, I think what it, it was just hard to scale them. At least in Tripwire's situation, we actually got to like $20 million annual run rate, $24 million annual run rate within a few years. So like, if you compare this to like a regular recruiting agency, it was like super fast. But if you compare this to like the top software startups, not that like impressive.
27:29 Jared Friedman? And it became harder and harder to scale because you had more and more people.
27:32 Harj Taggar? Yeah. Yeah. Basically, like, the margins didn't work out particularly well, so then you need to keep raising more capital. And so if you were like a fearsomely good fundraiser, you could sort of do it and kind of push yourself. But even in those cases, I think most of those businesses, at some point, it just caught up with them. Like, at some point, like, actually, we have to figure out a way to scale the business and have good margins and make this like profitable and not just rely on the next fundraising round is what I felt hurt a lot of the.
27:57 Garry Tan? You could argue Zenefits was one of those for insurance and a bunch of different HR-related things. It was actually, they basically relied too much on hiring more salespeople and more customer success people instead of actually building software that then would create gross margin. And so Parker Conrad said, well, I'm not going to do that again. And I'm also going to force all the engineers to do the customer support so that they go on to build software that doesn't require so much support. And thus there is gross margin. And that was a whole lesson that I feel like the whole tech community learned collectively through the 2010s. If we learned one thing, it's gross margin matters a lot. You cannot and should
28:41 Garry Tan? not sell $20 bills for $10 because you're going to lose everything.
28:46 Harj Taggar? I think a sort of non-financial reason why the gross margins matter is low gross margin businesses usually mean you have some ops component and then you have to like run the ops component. So if I think of my like Tripwire experience, there was like a lot of brain power spent on like, how do we like manage this team of like contracted engineers, this team of like humans looking after the, like essentially the human recruiting team, like lots of pieces of the business where actually the existential issue we had is how do we get to like millions of engineers across the world, like all like on like on our platform and all locked in, i.e. like how do you just get lots of distribution? And I think something that's nice about a high gross margin business is another way of saying it's just a simpler product or a simpler company to run. And you can actually
29:28 Harj Taggar? just spend all of your time focused in on how do I make the product better and how do I get more users and get more distribution so that you can keep that like exponential growth for a decade. And I think a lot of full-stack startups partly plateaued out because it's just the complex businesses to run.
29:44 Diana Hu? Maybe a very famous example of that was like WeWork, right?
29:48 Harj Taggar? Yeah.
29:49 Garry Tan? Yeah.
29:49 Diana Hu? Which is very, took it to the limit. Margins were not there. It was not, didn't have the tech margins, right?
29:57 Garry Tan? It had community adjusted a bit, which was... Very creative.
30:03 Jared Friedman? What I've been excited about recently is like, I think you can make a bull case that like, now is the time to build these full stack companies because like, you know, like you were saying, like the triple by 2.0s won't have to hire this huge ops team and have bad gross margins. They'll just have agents that do all the work. And so like now actually like full stack companies can look like software companies under the hood for the first time.
30:23 Harj Taggar? And you gave a great example. So Atrium, started by Justin Tan, full stack law firm, didn't work out for all, I think, a lot of these same reasons.
30:32 Garry Tan? I heard him say that before. It's like, look, we went in trying to use AI to automate large parts of it. And it wasn't, the AI was not good enough at that moment, but it's good enough now.
30:43 Harj Taggar? If you look at within YC, we have Legora, which is like this, like one of the fastest growing companies we've ever funded. And it's not building a law firm, but they're essentially, you know, building AI tools for lawyers. But you can see where that's going to extend out to, right? Like eventually their agents are just going to do all of the legal work and they'll be the biggest law firm on the planet. And yeah, I think that's a kind of full stack startup that just wasn't possible pre LLM.
31:08 Garry Tan? I think this started right when Uber and Lyft and Instacart and all of these companies were happening. And the thing is now, I mean, you can actually have LLMs do a lot of the knowledge work. And then, I mean, increasingly, it could actually have memory. I mean, this is one of the RFSs. It's literally, you can have virtual assistants, but they become less and less virtual if they can also hire real people to do things for you.
31:35 Jared Friedman? Virtual assistant marketplaces was definitely like a whole category of companies for like 15 years, including exact where you build like a marketplace of like people in the Philippines and like other, other countries. And then you like can expose to sort of like Airbnb UI. I don't think any of them ever like really got, really became like amazing businesses, though.
31:54 Harj Taggar? Going back to Pete's post, I think the other thing that's interesting about the, the points he made around sort of the system prompt and user prompt, and maybe we want to expose the system prompt to users a little bit more.
32:06 Harj Taggar? It's an example of just how we're still so early in just using AIs and building agents. There's all this like tooling and infrastructure like still to build. You have to do evals, you have to run the models, like a whole bunch of stuff to build still. And so there's clearly still a bunch of startups yet to be built in just the infrastructure space around, you know, deploying AI and using agents. And Jared, you know, it's interesting, something that struck me about when I first came back to YC in 2020 is I remember a class of idea we weren't interested in funding was anything in the world of like ML, machine learning operations or ML tools. And I remember reading some applications and like, ah, like another ML ops like team, like they sort of never go anywhere.
32:52 Harj Taggar? Clearly, if you were working on ML ops in 2020 and you just stuck it out for a few years, you're in the right spot.
33:01 Harj Taggar? Any content you can share from that?
33:02 Jared Friedman? Yeah, I remember I got so frustrated after years and years of funding these ML ops companies with really smart, really like optimistic founders that just like didn't go anywhere that I ran a query to count. And I remember finding that, I think this was around 2019, we had more applications in 2019 for companies building ML tooling than we had applications for like the customers of those companies, like, like anyone who's like applying ML to like any sort of product at all. And like, I think that was the core problem is that like these people were building ML tooling, but there was no one to sell it to because like the ML didn't actually work. So there just wasn't anything useful that you could build with, with, with all this ML tooling.
33:46 Garry Tan? People didn't want it yet. I mean, directionally, it was absolutely correct. Like from a sci fi level on a 10 year basis, it was beyond correct.
33:53 Jared Friedman? Yes.
33:54 Garry Tan? It was just wrong for that moment.
33:56 Diana Hu? You actually have a team that stuck it out. I mean, part of the lesson is sometimes it will take a bit of time for technology to catch up. And this company called Replicate that you work with stuck it out. It was from that era.
34:08 Jared Friedman? Yeah, Replicate was from winter 20 and they started the company right before COVID. And during the pandemic, it was going so poorly that they actually stopped working on it for several months. And just like, didn't work on it because like, it wasn't clear that the thing like had a future at all. And then they picked it back up and just started like working on it quietly. But it basically was just like, they were just building this thing in obscurity for two years until the image diffusion models came out. And then it just like exploded like overnight.
34:36 Diana Hu? Ollama is another good example.
34:37 Jared Friedman? Oh, yeah.
34:38 Diana Hu? You work with.
34:39 Jared Friedman? About Ollama?
34:40 Diana Hu? So the Ollama folks were also from that pandemic era. And similar story to, uh, to replicate. They were kind of trying to do different things around here too. And they were trying to work it out to make open source models deploy a lot better. And they were also quietly working on it for a while. Things weren't really taking off. And then suddenly, I think the moment for them was when Llama got released. That was like the easiest way for any developer to run open source models locally. And it took off because suddenly the interest to run models locally just took off when things started to work. But not before that, because there were all these other open source models that were in hugging face.
35:22 Diana Hu? And especially the ones from like BERT models, those were like the more used deep learning models. They were like just okay, but not many people were using them because they weren't quite working.
35:34 Garry Tan? What's the moral of the story? I mean, some of it is like, uh, be on top of the oil well before the oil starts shooting out of the ground. But is that actionable?
35:44 Harj Taggar? It's kind of the classic startup advice of follow your own curiosity. Like, most of these teams, almost all these teams were working on it because they were just interested in ML. They wanted to deploy models. They were frustrated with the tooling. Probably weren't necessarily commercially minded and trying to pick the best startup idea they could possibly work on. But I know sometimes you get lucky, sometimes.
36:01 Garry Tan? There's so many ways to do it. I mean, we were just sitting with Varun from Windsurf. And he pivoted out of ML Ops into CodeGen.
36:09 Jared Friedman? Deepgram is another one. Deepgram was one of the first companies I worked with back in 2016. And it was these two physics PhDs. They had done string theory, so they weren't even computer scientists. And they got interested in deep learning because they saw like parallels with string theory. And they were, it was, it was exactly what you said, Harj. They, they found the mathematics to be elegant and interesting. Like, that's really the origin. And so they started working on deep learning before anybody really. And they built this speech to text stuff. And it just like, didn't really work that well for like a long time. And so like, nobody really like paid much attention to, to this company. It wasn't famous. The founders, to their credit, just like kept working on it.
36:53 Jared Friedman? And then when the voice agents took off, they all needed speech to text and text to speech. And most of them are actually using Deepgram under the hood. And so they've just like exploded in the last couple of years.
37:04 Harj Taggar? I mean, I guess essentially the whole AI revolution is built on Ilya Sutskevra following his own curiosity for like a long period of time.
37:12 Garry Tan? More of that.
37:13 Jared Friedman? Actually, this is maybe a meta point on this whole conversation. So we were at colleges, Tayyan and I went on this college tour. And we spent several weeks speaking to college students. And I realized that there's this piece of startup advice that became canon that I think is outdated. Back in the pre-AI era, it was really hard to come up with good new startup ideas because the, like, the idea space had been picked over for like 20 years. And so a lot of the, the startup advice that people would hear would be like, you, you really need to like sell before you build. You have to do like detailed customer discovery and make sure that you've like found a real like new customer need.
37:55 Garry Tan? It was like the lean startup.
37:57 Jared Friedman? It's a lean startup. Yeah.
37:58 Garry Tan? Fail fast.
37:59 Jared Friedman? Fail fast, all this stuff. And that is still the advice that college students, I think, are receiving for the most part because it became so dominant. But I would argue that in this new AI era, that the right mental model is closer to what Harj said, which is just like, use interesting technology, follow your own curiosity, figure out what's possible. And like, if you're, if you're doing that, if you're. Living at the edge of the future, like PG said, and you're exploring the latest technology. Like, there's so many great startup ideas, you're very likely to just bump into one.
38:29 Garry Tan? I guess the reason why it could work extra well today is that you apply the right prompts and the right data set and a little bit of ingenuity, the right evals, a little bit of taste, and you can get like just magical output. And then that's still a secret. I think, yeah, I mean, you can tell it's still a secret because you could look at, there are like hundreds of unicorns out there that still exist and that are doing great. You know, the growing year on year, have plenty of cash, all of that. But the number of them that are actually doing any sort of like transformation internally is not that many. Like a shocking few number of companies that are,
39:13 Garry Tan? you know, hundred to thousand person startups that, you know, they're going to be great businesses. But that class of startup, like by and large, they are not entirely aware. Like, there isn't a Skunk Works project in those things yet.
39:27 Jared Friedman? Yep.
39:27 Garry Tan? Like, you know, the extent of it is maybe the CEO is playing around with it. Like, maybe some of the engineers who are really forward thinking are doing things in their spare time with it. Maybe they're using Windsurf or Cursor for the first time. And it's like, you look down and you're like, what year is it? It's a little bit like, hey, you know, get on this. Like, I think Bob McGrew came on our channel and he was just shocked. Like, he was one of the guys as chief research officer, like, building, you know, building what became O1 and O3 and all these things. And then he releases it and like, who's using it? Like, he expected this, you know, crazy, you know, outpouring of like, intelligence is too cheap to meter. This is amazing.
40:10 Garry Tan? And it's like, actually, like, people are mainly just, we're just still on our quarterly roadmap unchanged from, you know, even a year ago.
40:18 Jared Friedman? Yep.
40:18 Garry Tan? Pretty wild.
40:19 Harj Taggar? Okay, cool. I think that's all we have time for today. My main takeaway from this has been there's never a better time to build. So many ideas are possible today that weren't even possible a year ago. And the best way to find them is just follow your own curiosity. Keep building. Thanks for watching. See you on the next show.
Previous Windsurf CEO: Betting On AI Agents, Pivoting In 48 Hours, And The Future of Coding Next State-Of-The-Art Prompting For AI Agents Page updated October 25, 2026 · transcript and summary are machine-generated