Startup Ideas You Can Now Build With AI
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. If you’re living at the edge of the future and you’re exploring the latest technology, there are so many great startup ideas, you’re very likely just to bump into one. You apply the right prompts and the right data set, and a little bit of ingenuity, the right eval, a little bit of taste, and you can get magical output.
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. 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? 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 well. One idea that is very personal to me would be recruiting startups since I ran a recruiting startup, triple bite, for almost five years. I think something that I’ve clearly seen is that there was a period of time when we started Triple Bites around 2015 where recruiting startups were kind of like a really popular type of startup.
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. We started Triple Bite with the thesis that 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.
This was all prelim, 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 build up this label 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. Initially, it was actually a three-sided marketplace where you needed to hire an interviewer in between to get that human touch. We had companies hiring engineers, we had the engineers looking for jobs, and then we had engineers we contracted to interview the engineers.
There’s lots of things going on right now, and all of the evaluation piece of it, at least now with AI, is very possible. With the AI codegen models, you can do code evaluation, and I think probably one of the hot AI startups at the moment is this company called Meror which is essentially similar to the Triple Bite idea. It’s a marketplace for hiring software engineers, but what AI has unlocked for them is the evaluation piece that they could just do on day one using LLMs.
They need to build up this big label data set and they’ve been able to expand into other types of knowledge work quite easily. For us to have gone from engineers to analysts to all these other things would have taken years because, again, we had to rebuild the label data set. But with LLMs, you can just do that on day one effectively. I think 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.
So that’s a very powerful prompt for anyone listening. What are marketplaces that are three-sided or four-sided marketplaces that suddenly become two or three-sided? Now there are two-sided marketplaces like Duolingo that are a little bit under fire because they’re starting to say maybe we’re just going to use AI for the person that you’re going to talk to in another language.
That is totally a coherent thing that you could go to almost any marketplace in the world and ask, “What will LLMs do in that marketplace?” The other thing I really respect the Meror founders for is there’s also just a psychological element as a founder when you enter 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, Triple Bite raised something like $50 million; our main competitor raised over a hundred 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. I think going in, you face a lot of skepticism if you’re going to pitch investors for an idea. Even when you have the “LLMs change everything” angle, that pitch two years ago was still not as compelling as it is today.
You have to be willing to push through a lot of cynicism and people who are burnt out, who have lost lots of money on an idea to even keep going to test it out and make it work. That’s something that repeats all the time. Instacart was that story. Exactly. Web Van 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.”
Simultaneously, the iPhone and Android phones were everywhere, and you could have a mobile marketplace for the first time. That’s why we’re pretty excited about this moment because suddenly all the walls in the idea maze have shifted around, and the only way to find out is you’ve got to actually be in the maze. It is very similar to Instacart and Web Van if we go back in history, right? The big technology unlocked for Instacart was the fact that everyone had a phone now. It enabled the webvan model to actually work for the first time, and it’s the same thing with LLMs and recruiting companies now and a whole bunch of other ideas.
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, there’s a company called Apriora 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. The engineers hate doing it. Even that one piece is not exactly, let’s say, marketplace-oriented, but if you solve it right now, it works out. API actually does a pretty good job. It’s being used by large companies and it’s been taking off.
It’s another example where you can actually expand the market because I think there are plenty of technical screening products pre-Apriora, but you could only use them to do fairly simple evaluations to weed out people who weren’t engineers at all or were very, very junior. But a prior product, now with LLMs, you can do more sophisticated evaluations to kind of get more nuanced levels of screening. Suddenly now, companies will be like, “Oh actually I could give this to not just my international applicants or my college students. I’ll just give it to senior engineers who are applying,” which just opens up the opportunity.
You were talking a bit about education as well, Gary, about Duolingo. I think that aspect of hyper-personalization is one of the holy grails that has been difficult for edtech companies to crack, right? Because every student, as they go through their learning journey, is very unique and knows different things, and it sounds really cool to build like the awesome personal AI Twitter that Harj did an RFS for.
The thing I’m excited about is that for as long as I can remember, the internet’s been around, and one of the dreams of it was that everyone would have access to personalized learning and knowledge, and we’d all just have these great intellectual tools to learn anything. Clearly, the internet’s made it easier to learn, but we’ve never had truly personalized learning or the 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 YC who are interested in building that type of product.
A couple of companies that we funded that are kind of working out is this other company that Nico funded, called Revision Dojo, that helps students do exam prep and is sort of the version of flashcards, but not like the janky, boring going through content. It’s the version that actually students like and gets tailored for their journey.
That one has a lot of DAUs and a lot of power users, which is super interesting. I think Jared, you had worked with this other company called Adexia as well. Adexia does tools for teachers to grade their assignments, which is another example of work that is not a person’s main job but is this other thing that they have to do, like engineers doing recruiting that they generally hate doing. There’s a lot of studies that show that like the biggest reason that teachers churn out of the workforce is that they hate grading assignments.
Adexia is an agent that’s very good at helping teachers to grade assignments. One of the interesting trends for some of this stuff is that it’s private schools who are actually much more nimble, and 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.
I guess the question for you, Gary, is I’m curious about this stuff. It’s clearly possible to build much better products with LLMs. If we take the learning apps, for example, they can go far beyond anything you could do for personalized learning prelims. But it doesn’t necessarily mean that you instantly get more distribution, especially if you’re going after the consumer market. 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?
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. That’s something I would probably track. It seems clear that 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 consumer AI finally might be here soon. The thing to track is how smart is it such that any given user incrementally only costs pennies or like 10 or 15 cents? Then it becomes so cheap that you will just have intelligence for free. Maybe it’ll be a return to the premium model that we got used to during web 2.0. This idea that you could basically give away your product, and then for five or 10% of those users, there are things that they so want that you’re going to sell them a $5 or $10 or $20 a month subscription. That’s basically what OpenAI is doing, right?
The complexity does it. OpenAI is 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 can kind of go up even a couple of grade levels. Those are real outcomes for students. Right now, you’ve still got to pay for it, but maybe not for a while. That’s actually a really big unlock. 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 really thousands of apps like this start coming out across all the different things you’ll need.
That’s something that I know we keep saying is gonna happen. I mean, it’s kind of happening already for edtech. Speak is 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 language learning, which might have been a bit contrarian back then because Duolingo seemed to be the game in town that was winning. They really focused on personalizing that whole language learning, and they started taking off in Korea for a lot of learners trying to learn English.
When GPT-3 and 3.5 came out, they were early adopters and saw that wow, this is going to be the moment to double down, and they’ve been on this trajectory now with lots of MAUs; EAS is really working out. I think one thing going back to consumer stuff that we haven’t talked about as much is that we’ve seen a lot with the startups that are selling to enterprises or companies about how the budgets become so much bigger when they 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 way more.
So the same thing will apply in consumer, right? If you think about a personalized learning app, often edtech companies struggle with who’s actually the buyer and who’s going to pay. If you go for younger children, for example, you’ve got to get the parents to pay. But the parents aren’t going to pay that much for an app that their kids don’t retain or complete, like some sort of online course that they’re disengaged with. But we know that parents will definitely pay for human tutors, and that’s probably quite a big market.
So if your app goes from being 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. It’s possible that the product now just 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 a much bigger business than was possible before.
Yeah, I feel like we have to talk about modes 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. What you need is brand; you need switching costs. Sometimes it’s integration with other technologies that are surrounding that experience. In a school, it would probably be being connected to Clever, for instance; login authentication is pretty obvious.
So yeah, I feel like Sam Altman has talked about this a bunch. It’s not enough to drop AI in; you still have to actually build a business. I don’t think OpenAI is necessarily out to get all the startups; I actually think on the API side they very much hope that a lot of them do really well, and certainly we want that too. They just hired the Instacart CEO as their CEO of application. So, it does seem like they are definitely paying more attention to the application layer.
That’s right. I mean, you’d be crazy not to. By all accounts, OpenAI is highly likely to be a trillion-dollar company at some point, as powerful as a Google or an Apple. The interesting thing right now is they’re still on the comeup. The big tech platforms are actually still holding back a lot of the AI labs. The most profound example of this is, why is Siri still so dumb? It makes no sense, right? Totally.
I think that points to something that we actually really need in tech today. We really need platform neutrality. In the same way, 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-preference their own content or the content of their partners. 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 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. These are things that, you know, the government did get involved in and said, “Hey, you cannot self-preference in this way.” If you remember the moment where Internet Explorer had a majority of web users, 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. For that free market to create choice and, therefore, prosperity and abundance. I would argue, why doesn’t this exist for voice on phones? You should be able to pick; you shouldn’t be forced to use Google Assistant. You shouldn’t be forced to use Siri. You should be allowed to pick. It’s been many years of having to use a very dumb Siri.
On the moat topic, something I just find fascinating is I saw some numbers recently about how Gemini Pro models, particularly from consumers, are just a significant fraction of ChatGPTs. I think at YC 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 as good and in some cases a better model than GPT-3 for various tasks. That hasn’t trickled down into public awareness yet, right? Which is fascinating since Google already has all the users with their phones.
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 moat around being the first in a space and staking your claim as the best product for a specific use case. And I feel like actually making it good. Yep.
But at some point, maybe it doesn’t even necessarily need to be objectively the best; it just needs to be good enough. That’s the bet that I think a lot of the big tech companies are trying and failing at. Microsoft has a co-pilot built into Windows now that is still quite inferior to anything OpenAI puts out. Gemini itself is very good, and I use it quite a lot. It’s probably 40% of my agent if I need to, especially to summarize YouTube videos; it’s very good at that for multimodal stuff.
Yeah, a lot of the Gemini integrations into Gmail or Google Drive are totally useless. It’s like, is there someone at the wheel over there? I don’t get it. I think that’s even confusing for us as developers. There’s actually two different products: there’s Gemini, where you can consume Gemini, and Vertex AI Gemini. I think they’re like different orgs. It’s suffering a little bit from being too big of a company and essentially shipping the org. There are 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.
I think that comes from the culture of Google. There’s definitely this sense that if two orgs are competing and fighting, normally in a normal org you go up to some level, and then ultimately the CEO or founders say, “Okay, well, I see the points over here; I see the points over there; we’re going this way.” But having lots of friends from Google, it doesn’t seem like that’s the culture there. There’s a layer of VP and sort of management that is actually like, “You guys just fight it out,” and so you ship the org.
I think the crazy thing about Google is they probably should have won a lot of the experience of the best model. There’s almost like a Game of Thrones analogy that could be used. They might be a little bit like Daenerys Targaryen because they secretly have dragons. The dragons are the TPUs. 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 do 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. Is it actually the hardware? You can do it, but I think it’s just very expensive and not cost-effective. I think they’ve 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, so I’m sure they’re probably working on something around there too.
The classic innovator’s dilemma: if Google replaced google.com with Gemini Pro, it would instantly presumably be the number one chatbot LLM service in the world, but that would give up 80% of its revenue. Yeah, you would probably need a pretty strong founder CEO to do that. It’s the kind of thing I can imagine Zuck doing, being willing like, “Yeah, you just can’t imagine a hired CEO who’s going to do that.” He’s done that; he renamed the company to Meta.
Meta has its own issues too. I’m so surprised, you know? I mean, you have Meta’s AI in WhatsApp. It’s in the blue app; it’s everywhere. But I don’t think any of us actually use it. 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 now have an AI system that’s in all of your chats, and it comes with a, you can just add it, and it will just start talking in a group chat, and it feels quite invasive, actually.
It’s not that smart, and then it can’t do anything. Most people are surprised that it’s in there. It feels like having someone from Facebook just in your chats. It reminds me of the original newsfeed launch or something. It’s just the classic Meta style of, “This is sort of, I don’t know, objectively optimal; I’m sure people will love it.” You need to add a little bit of design taste into these things.
It blows my mind that I can go to the blue app, which I still kind of, you know, people watching this are like, “What the heck’s the blue app?” This is like facebook.com, which maybe nobody uses anymore. It’s very millennial. But 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 them.” It’s like, what? What is the point of this?
Okay, our partner Pete Cuman wrote this really great essay where he talked about the Gemini integration with Gmail. He really 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 was a PM at Google. It was very profound that one of the things he pointed out was that you have a system prompt and a user prompt. If you are actually going to empower your users, you allow your user to change the system prompt, which is the part that normally is above; to use the Venice show’s idea of the API line, it’s like the system prompt is what is exerted upon.
Gemini follows this very specific thing. I think the example is actually an email saying that Pete’s going to be sick. He asks the agent to write this letter, and it’s very formal. Of course, it is because there’s no way to change the tone. It’s actually one of the best blog posts, and I think he had to vibe-code the blog post itself because you can actually try the prompts yourself on that webpage.
Yeah, it’s super cool. It’s in this interactive template that made me think it’s time to start an AI-first vibe coding blog platform. Oh, like AI-osterous. Yeah, basically, with all my extra time, that’s what I’m going to work on. But that’s a free idea for anyone watching. We’ll fund it.
There’s another class of startup ideas that I’m particularly excited about that I think are perhaps timely now. Do you guys remember the tech-enabled services wave? For folks who didn’t follow this, in the 2010s, there was this huge boom in companies called tech-enabled services. Triple A was one, actually.
That was tech-enabled services for recruiting, right? We also had Atrium, which was technical services for law firms. It started with Balaji’s blog post about full-stack startups. The concept was just that software eats the world means software just kind of goes into the real world, and this is not a success example but an example of it was, “Hey, instead of just having an app to deliver food, you should also have a kitchen that cooks the food and software to optimize the kitchen, and you just do everything.”
The full-stack startups, in theory, would be more valuable than just the software startups because they would do everything. Instead of just selling software to restaurants and capturing 10%, you could own the restaurant and capture 100%. This is exactly what TripleA was because we were going to be a recruiting agency effectively. We weren’t selling software to recruiting agencies; we were actually doing the whole thing.
We also had recruiters on staff that were just there to help people negotiate salaries and match them to the right companies. It was very much in that wave of doing everything. But that wave of startups generally forgot that you need gross margins. What happened was like, fast forward basically, the short version is it didn’t really work. The full-stack startups actually were not more valuable than the SaaS companies, and the SaaS companies sort of won that round of the Darwinian competition of different business models.
I think fundamentally it’s just what Gary says: they were actually not great gross margin businesses. It was just hard to scale them. At least in TripleA’s situation, we actually got to like $20 million annual. Run rate, $24 million annual rate within a few years. So if you compared us to a regular recruiting agency, it was super fast. But if you compare it to the top software startups, not that impressive, and it became harder and harder to scale because you had more and more people. Yeah. Basically, the margins didn’t work out particularly well, and so then you need to keep raising more capital.
If you were 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 just caught up with them. At some point, we had to figure out a way to scale the business and have good margins and make this profitable and not just rely on the next fundraising round is what I felt hurt a lot of the… you could argue ZS was one of those for insurance and a bunch of different HR related things.
Basically, they relied too much on hiring more salespeople and more customer success people instead of actually building software that would create gross margin. And so Parker Conrad said, “Well, I’m not going to do that again. 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.”
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 not sell $20 bills for $10 because you’re going to lose everything. 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 run the ops component.
If I think of my tripey experience, there was a lot of brain power spent on how do we manage this team of contracted engineers, this team of humans looking after essentially the human recruiting team, lots of pieces of the business where actually the existential issue we had is how do we get to millions of engineers across the world all on our platform and all locked in, i.e., how do you just get lots of distribution?
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 just spend all of your time focused 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 exponential growth for a decade.
I think a lot of full-stack startups partly plateaued out because they’re complex businesses to run. Maybe a very famous example of that was WeWork, which took it to the limit. The margins were not there; it didn’t have the tech margins. Right. It had community adjusted AIDA, which was very creative.
What I’ve been excited about recently is like I think you can make a bullish case that now is the time to build these full-stack companies because, like you were saying, 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.
Now, actually, full-stack companies can look like software companies under the hood for the first time. And you gave a great example. Atrium started by Justin Khan, a full-stack law firm, didn’t work out for all I think a lot of these same reasons.
But now 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 that AI was not good enough at that moment, but it’s good enough now. If you look at within YC, we have Legora, which is one of the fastest growing companies we’ve ever funded.
It’s not building a law firm, but they’re essentially building AI tools for lawyers, but you can see where that’s going to extend out to, right? Eventually, their agents are just going to do all of the legal work, and they’ll be the biggest law firm on the planet.
Yeah, I think that’s a kind of full-stack startup that just wasn’t possible pre-LLM. I think it started right when Uber and Lyft and Instacart and all of these companies were happening. The thing is now, I mean, you can actually have LLMs do a lot of the knowledge work, and increasingly it could actually have memory.
I mean, this is one of the RFS’s; 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. Virtual assistant marketplaces were definitely a whole category of companies for like 15 years, including Exec, where you build a marketplace of people in the Philippines and other countries, and then you expose it to sort of an Airbnb UI.
I don’t think any of them ever really became amazing businesses, though. Going back to Pete’s post, I think the other thing that’s interesting about the points he made around sort of the system prompt and user prompt, maybe we want to expose the system prompt to users a little bit more. It’s an example of just how we’re still so early in using AI and building agents.
There’s all this tooling and infrastructure still to build. You have to do evals, you have to run the models, a whole bunch of stuff to build still. There’s clearly still a bunch of startups yet to be built in just the infrastructure space around deploying AI and using agents.
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 ideas we weren’t interested in funding was anything in the world of ML machine learning operations or ML tools. I remember reading some applications and thinking, “Ah, like another ML ops team. These sort of never go anywhere.”
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. Any context you can share from that? I remember I got so frustrated after years and years of funding these ML ops companies with really smart, really optimistic founders that just didn’t go anywhere.
I ran a query to count applications, 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 customers of those companies, like anyone who’s applying ML to any sort of product at all.
I think that was the core problem: these people were building ML tooling, but there was no one to sell it to because the ML didn’t actually work. There just wasn’t anything useful that you could build with all this ML tooling. 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. Yes, it was just wrong for that moment. 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.
This company called Replicate that you worked with stuck it out. It was from that era. Yeah, Replicate was from Winter 20, and they started the company right before COVID. During the pandemic, it was going so poorly that they actually stopped working on it for several months and just didn’t work on it because it wasn’t clear that the thing had a future at all.
Then they picked it back up and just started working on it quietly. Basically, they were just building this thing in obscurity for two years until the image diffusion models came out, and then it exploded overnight. OAM is another good example.
The Olama folks were also from that pandemic era, and similar story to Replicate. They were 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. 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 skyrocketed when things started to work, but not before that. Because there were all these other open-source models that were in Hugging Face, especially the ones from BERT models.
Those were the more used deep learning models. They were just okay, but not many people were using them because they weren’t quite working. What’s the moral of the story? I mean, some of it is like be on top of the oil well before the oil starts shooting out of the ground, but is that actionable?
It’s kind of the classic startup advice of follow your own curiosity. Most of these teams, or 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. They 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, there are so many ways to do it. I mean, we were just sitting with Verun from Windsurf, and he pivoted out of MLOps into codegen. 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 parallels with string theory. It was exactly what you said, Harge. They found the mathematics to be elegant and interesting. That’s really the origin, and so they started working on deep learning before anybody really.
They built this speech-to-text stuff, and it just didn’t really work that well for a long time, so nobody really paid attention to this company. It wasn’t famous. The founders, to their credit, just kept working on it. When the voice agents took off, they all needed speech-to-text and text-to-speech, and most of them were actually using Deepgram under the hood.
They’ve just exploded in the last couple of years. I guess essentially the whole AI revolution is built on Ilioskava following his own curiosity for a long period of time. We need more of that.
Actually, this is maybe a meta point on this whole conversation. We were at colleges; Tyenne 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 best idea space had been picked over for like 20 years.
A lot of the startup advice that people would hear would be like you really need to sell before you build. You have to do detailed customer discovery and make sure that you’ve found a real new customer need, like the lean startup. Yeah, exactly. Fail fast, all this stuff.
That is still the advice that college students are receiving for the most part because it became so dominant. But I would argue in this new AI era that the right mental model is closer to what Harge said, which is just use interesting technology, follow your own curiosity, figure out what’s possible, and 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, there are so many great startup ideas, and you’re very likely to just bump into one.
I guess the reason why it could work extra well today is that if you apply the right prompts and the right data set and a little bit of ingenuity, the right evals, a little bit of taste, you can get magical output. And then that’s still a secret, I think.
I mean, you can tell it’s still a secret because you could look at there are hundreds of unicorns out there that still exist and that are doing great. Growing year on year, have plenty of cash, all of that. But the number of them that are actually doing any sort of transformation internally is not that many; a shocking few number of companies that are 100 to 1,000 person startups that are going to be great businesses.
But that class of startup, by and large, they are not entirely aware; there isn’t a skunk works project in those things yet. The extent of it is maybe the CEO is playing around with it. Maybe some of the engineers who are forward-thinking are doing things in their spare time with it. Maybe they’re using Windsor for 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. I think Bob McGru came on our channel, and he was just shocked. He was one of the guys as chief research officer building what became 01 and 03 and all these things, and then he releases it, and like who’s using it?
He expected this crazy outpouring of intelligence; it’s too cheap to meter, this is amazing, and it’s like actually like people are mainly just still on our quarterly roadmap unchanged from even a year ago. Yep. Pretty wild.
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. The best way to find them is to just follow your own curiosity and keep building. Thanks for watching. See you on the next show.