The wildest thing right now is you can start a company that can make tens of millions of dollars literally in 24 months, and you can do it for potentially, you know, 2 million, 5 million. A year ago, I remember many of the startups in the

batch would get sort of enterprise proof of concepts or pilots, in particular, and there was a lot of cynicism around whether any of those pilots would translate into real revenue. Fast forward a year, I think we have all firsthand experienced that these pilots have

turned into real revenue. It's still early days, honestly, like it, you know, we sort of breathe a sigh of relief right now in 2024, but it's anyone's game, honestly. Like these things are moving so [Music] quickly. Welcome back to another episode

of the Light Cone. I'm Gary. This is Jared Harge and Diana, and collectively we've funded companies worth hundreds of billions of dollars right at the beginning. So 2024, what a year! How are you feeling about this, Harge?

Pretty great. I think this is the year that everything broke in favor of startups. What I've been thinking about a lot recently is when ChatGPT launched two years ago now, the immediate consensus view was all of the value would go to OpenAI, and very specifically,

do you all remember when they announced the GPT or the ChatGPT store? Yeah, like I remember the consensus was everything that was built on top of ChatGPT was a GPT wrapper, and the App Store was just going to be released and crush every single person.

Trying to build an AI application, and OpenAI would be a ginormous company, but there'd be no opportunity for startups sounds kind of ridiculous to say that now because, um, who even remembers the ChatGPT store? Exactly. The ChatGPT store itself was a nothing burger, but

like, more importantly, what are the big AI applications today? Like, I'd say outside of ChatGPT itself, the breakout consumer application is Perplexity. The breakout enterprise application is probably Glean. Maybe, um, in legal tech you have Casetext, you have Harvey, you have PhotoRoom. Like, there's the

point being there are many, many applications that have been built, not by OpenAI. It's been a great time to build startups. Yeah, the wildest thing right now is you can start a company that can make tens of millions of dollars, uh, literally in 24 months from zero,

and, uh, you can do it for potentially, you know, 2 million, 5 million. That's sort of the story of one of these companies, Opus Clip, which never had to raise a real Series A, and that's something that we sort of see across the YC community as

well. Yeah, I think that's, um, a particularly important point, that you can do it as a startup without raising tons of capital. Because post the ChatGPT store launch, I then remember, uh, Anthropic and Claude emerged, and the consensus view for a while was all of

the value is going to go to one of these foundation model companies, and that the only way you can compete in AI is to raise huge amounts of money, um, either because you've got venture capital or you're Amazon or Facebook or Google, um, with tons of cash already. But that if

If you weren't one of the big foundation models, um, there would be no value, and the applications built on top of these things would either be built by the foundation model companies themselves or just not be that valuable. Again, something that turned out to be completely untrue, right? And in

particular, what drove that is open source. Like the weird series of events where, like, the weights being leaked and, like, Meta just, like, rolling out LLaMA. Yeah, I kind of forced the hand for Meta to launch LLaMA, which was funny. And people thought, "Oh, it was just this cool

open-source model," but it was 18 months behind OpenAI. And people started doing a lot of derivative work out of it, like Vicuna and all these other animals related to llamas that came out. And it took the—or LLaMA—one of the companies at

YC as well that enabled people to do, uh, local kind of Docker development, like models running on device. It was pretty cool, but people didn't think that, um, they were going to be able to catch up. And the thing that changed from 2023 to 2024 is that during the summer, it was a

turning point. It was the first time that the top foundation model in all the rankings and benchmarks was LLaMA, and that was a shock to the community. Yeah, so it turns out choice, uh, matters, and choice means that it's not as much about the model. I think the model still matters

quite a lot, um, but once you have choice in model, it means you can't have the sort of idea of monopoly pricing. And you have that model; your competitor also has that model, but all the other things seem to end up mattering a lot more, which is

Product, your ability to sell, your ability to actually adjust to user feedback, your ability to get to zero churn—all of those suddenly become far more important than capturing a light cone of all future value through the model. Right? A very specific way I... I felt this is... I remember a year ago working

with startups in the batch that are essentially building model routers, just like an API to call a specific model. And I remember the, um, a lot of the motivation for that at the time was reducing cost. It was like, "Oh, like you

don't want to just burn up all of your ChatGPT calls; you want to spread them out across various different models." And the argument against that was, "Oh, like the cost of all this stuff is going down to zero anyway; like there's no value to be had in being

like a model router, and no one wants to build their applications with a model router; they're all just going to call whatever is the best model." I think fast forward a year, like that's totally not true. Like, from what I can tell, the model router was actually a really great

entry point into just building a new stack for building LLM-powered apps, and most of the applications we're seeing, I think they just don't want to be beholden to a specific model. Does that map with what you've seen?

Yeah, actually one of the things we've seen now in the fall batch that just presented at Demo Day, which was one of the trends that shifted from Summer '24 and Winter '24, was precisely what you're saying: companies started to use multiple models for their applications—like the best one for speed at some point, because

Sometimes you need to parse a lot of input very quickly. It's fine if it's a bit more lossy, and then you need the bigger model to handle the more complex task. So a lot of companies in Fall '24 have this, actually, a multiple-model

architecture to use the best one for the best task, which is similar to the concept of a model router, but the idea evolved. Instead of being more of routing, it was more of an orchestration. I think a concrete example we gave a couple of episodes ago was

Camper, a company we work with. They use the fastest model for parsing PDFs, and for more complex ones, they use GPT-3.  That's how it's done. Other companies doing fraud detection have this concept of a

junior risk analyst, where they just use a fast and easy model like GPT-4 mini, and then they use the bigger one with GPT-3 or similar. Another example is—I think Cursor talks about it in their episode with Lex Fridman—they also have this

complex multi-architecture with multiple models. This is why it works well. They have one very specific model for predicting what you're going to type next, but one for understanding the whole code base—very different tasks. So that's definitely happening now. Yeah, the

other thing that popped up for Fall batch—there's a company I'm working with called Variant—and what they're trying to do is take basically state-of-the-art open-source LLMs that can do code generation and then teach

the aesthetics. So, uh, starting with icon generation, and so they built this huge sort of post-training workflow that should work on, you know, as the open-source models get smarter and better at, broadly, they can just, you know, take the next version of that and

then, uh, take their post-training architecture and data set and then basically teach a given model aesthetics; so what a certain thing is supposed to look like, and uh, not in a diffusion sort of way, but actually at the SVG level. And we think SVG will actually translate

into all kinds of aesthetics. So it's an interesting approach and like one of the newer ones, and that post-training is a whole coherent way to sort of skip the whole idea that, um, all the value is like accrued into the model, especially because

of open source. To your point, the other thing I, I've been having flashbacks to is a year ago. I remember many of the startups in the batch would get sort of enterprise proof-of-concepts or pilots, in particular, and there was a lot of cynicism around whether any of those

pilots would translate into real revenue. Um, lots of parallels to crypto and how anytime there's some new interesting technology—blockchain, more specifically than crypto—but anytime there's a new technology, enterprises always want to run pilots and POCs because it's someone's job to like check off, yeah, we did the

like hot new technology thing. The chief innovation officer must have his V. We've spoken about this in one of our episodes, I think. And yeah, fast forward a year, I think we have all firsthand experienced that these pilots have turned into like real revenue and

If anything, the startups in the Y Combinator batch now are going to sell into real enterprises faster than they have before and are ramping up revenue and reaching milestones like a million dollars ARR faster than I've ever seen. Yeah, the fall batch just did this actually.

Again, which is actually the first time I think we noticed it. It was actually the summer batch of this year, and one of the funnier things that we realized was, do you remember when Paul Graham would tell us how fast you needed to grow during the Y Combinator batch? 10% a week.

A week, and the wild thing is, in aggregate across both summer and fall batches, that's what those batches did. Wow! So, which I don't think ever happened.  It's 3X over the course of Y Combinator, which I

don't think has ever actually happened on average. On average, it was only the best companies that did that, which is the top quartile or something, right? So the companies are better. The general thing that is true is just that the time it's taking to reach

$100 million in annual revenue is trending down. Yeah, and not only that, um, we had dinner with Ben Horowitz recently, and remember he was saying when they started Andreessen Horowitz, the common understanding was that in any given year there'd only be 15 companies that year

that would even make it to $100 million a year in revenue. And they said they ran the numbers the last 20 years, and every decade the number of companies that could actually make it to $100 million went up by 10x. So what was 15 per year,

Maybe 20 years ago, I mean we're talking about 1,500 companies a year that have a real shot at actually making that number. And when you combine that with what we're seeing in the summer and fall batches, it's not that surprising.

Jared had a really good argument on our last episode about how vertical AI is going to enable this, to have 1,500-plus companies bloom. Yeah, that's why it's growing so fast; it's because the value proposition to companies of these products is so incredibly strong that

like they're just flying off the shelves, because companies are smart and they can do an ROI calculation. And when the ROI is fantastic, all those truisms that people believe about enterprise sales cycles and like, quite a taste to get big enterprise deals, go out

the window, because companies are smart and they'll make rational decisions. You know, there's another way that this broken favor of startups that I was thinking about... um, it's hard to even remember now, but a year ago, one of the things that people said a lot was that

these LLMs are not reliable enough to deploy in the enterprise. They hallucinate. That was why a lot of people said, like, these pilots and POCs won't translate into real contracts, because, yeah, it's too risky a technology for people to actually

deploy. Yeah, and not only is it translating into real revenue, but translating into real deployments that are like being used at large scale, you know, doing thousands of tickets a day. And I think it's because we've learned how to make the agents reliable via the kinds of

techniques that Jake talked about when he was here, and just all this infrastructure has grown up around the models. It's enabled people to make them reliable. Yeah, that's actually a big trend this year: this concept of thinking of AI more

as agentic. That's a term that kind of bubbled up a lot this year. It wasn't in the conversational space last year. Last year was more about a lot of things that were very chat-like—chatbots, I mean—that was kind of the riff on

it, but now remixed into a bunch of agents for XYZ. And we just—I mean, you just put out—Gary just put out a great explainer video about computer use from Claude—but just the capability of the models keeps pushing in the direction of just being able to do complex

multi-step things and actually take over your computer, call other applications, and perform complex tasks that just didn't seem possible a year ago. What about regulation? Seems like we sort of dodged a bullet there with 1047, and

it looks like some of the Biden EO is not that likely to survive a Trump White House. TBD what that means in the longer term, but certainly one of the things that we were very worried about was that some certain amount of AI, beyond a certain level,

would suddenly become illegal or require registration at your local office. It's certainly been a weird time to be in tech because I've never experienced software and technology intersecting with politics so much, and in particular, I'm not used to genuinely caring about national politics affecting

Startups in a YC batch, or just companies less than a year old, it did, for a moment, worry us. It wasn't clear whether startups would build innovative AI applications or suffer from regulatory capture.

From OpenAI and a few big players. We're obviously glad it broke in the favor of startups. It seems like we're still in the early game.  The platforms themselves could resemble the

Win32 monopoly, right? Windows has access to the APIs; they know the stats about what's working on their platforms, and guess what? They can build it into their platform. We sort of breathed a sigh of relief in

2024, but it's anyone's game. These things are moving so quickly; I wouldn't totally breathe your last sigh of relief yet. We have to keep working on this.

It's clearly been a great year for startups. What else has been happening? Who else has it been a great year for? There have certainly been some big funding rounds, like OpenAI unsurprisingly raising huge amounts of capital.

Even within YC, we've seen Scale AI really break out this year. $6 billion for OpenAI, $1 billion for Scale, $1 billion for Cohere. Scale AI is worth talking about because it's such a classic startup story. You were there in the early days, right?

I interviewed them for YC. Um, what, um, tell us what the idea was that they interviewed with and how they ended up landing on what you know is probably one of the best startup ideas of the last 10 years. The fun thing about the Scale AI story is that it is the sort of epitome of the classic YC startup story.

There are other kinds of startups that get started, you know, like Stripe, for example, that's not a typical YC startup story where some very well-established people raise a billion dollars with a PowerPoint pitch. But like Scale AI is like the classic story of how young programmers can gradually build a 10-billion-dollar company over time by being smarter and harder working than anybody else. So yeah, when Alex

uh, interviewed at YC, he wasn't working on anything related to AI. It was a completely different idea, and the idea for Scale AI kind of got pulled out of him by the market. And it's it's actually still like several pivots, because the

original idea at YC didn't have anything to do with AI. And then for a long time he was basically doing data labeling for the self-driving car companies. They applied, as I remember, they applied with a healthcare-related idea. Yeah, it was a website for booking

doctor's appointments. Okay, cool. And then they pivoted during the batch. Um, do you remember how they came up with the data labeling idea? Because this is supposed to be in, was this 2016? Yeah, the way they came up with the data labeling idea was that Alex had

I worked at Quora, and Quora had to do some data labeling for moderation and stuff. So, at the time, the big data labeling service was Amazon Mechanical Turk, and they were deemed unbeatable because they were run by Amazon, and Amazon could throw infinite money at it.

It was always at scale; it was at a quite large scale already. But Alex had a unique insight: he'd actually used Mechanical Turk at Quora, and he knew that it kind of sucked to actually use it. So he had this

sort of unique insight on the world, and so he just tried to build a better Mechanical Turk, basically the version he would have wanted when he was at Quora. And as I remember it, like, really their early traction came almost entirely from one customer: Cruise.

Yes, which needed to do tons of data labeling on all the images that the cars were taking as they were driving around San Francisco. You got to draw a circle around the traffic light and things like that. And the cool thing about Scale is that they actually

caught two waves. So they, you know, accidentally caught the first wave of all these self-driving car companies because ML took off at that time in computer vision. There was just an unprecedented demand for labeled data for training sets that just hadn't existed before, and so they were able to

ride that wave. And then, as that wave was cresting, LLMs got big, and all of these companies needed to do that at very large scale, and Scale was just perfectly positioned to move into that business as well. Yeah, I think the

Scale's story is so interesting.  Before LLMs, it was clear it was a multi-billion dollar business. Now, with LLMs—it's called the LLM wave—it's probably going to be a hundred billion dollar plus company.

And I'm seeing that at the ground level too, where many companies that maybe finished their batch even pre-batch didn't have an idea, pivoted into an AI idea that's taking off. I'm just seeing much more success in founders who waited and could find

an idea that they just couldn't before. I have a company from a year ago; they proved the whole batch, they couldn't find a great idea. It actually took them six months after the batch until they realized one of their parents ran a dentist office, so he

just decided to go hang out at the office to see if there was anything he could automate. Uh, and they just ended up building an AI back office for dentist offices, and now it's just like their week-over-week growth is fantastic. It's doing really, really well, and I'm

seeing lots of cases like that spring up. Definitely seeing that as well. I think there's something about the advantage of having all these very hardcore, young, technical founders that are willing to kind of just bet the farm and go all in on just a little bit of a glimmer of, "Oh,

this is where the future is going to be; let me just try it." And then it actually ends up working. I just started with the dentist; I have a lot of teams that pivoted as well into different spaces where they kind of found that glimmer.

Like, oh, computer use came out, and I have a couple companies that are working and betting and going in that direction, and it's like working well. I mean, it's still early. I mean, this is just a fall batch, but that's cool too. Okay, so what are

some of the actual trends that we've seen? What are some of the specific trends and waves that startups have been riding coming out of the batches? Voice AI is something we've talked about. It's clearly, um, maybe the most promising vertical for AI right now in terms of

just raw traction. Do you think, uh, voice is a win-or-take-all, or will it be something that has sort of a hundred different verticals that are very tailored to those specific verticals? That's literally one of the questions I get from some of our, uh,

voice AI startups themselves. They're like, should I be going horizontal or should I just continue to grow within my vertical? It feels to me like voice itself, VO, is, um, I just like, it touches everything, and there's so many different applications for it that

um, you can—there's probably infinite applications to build where voice is the interesting element of it. I mean, things that just spring off the top of your head, like language learning applications. I'm sure there's not going to be just one really cool voice AI-powered language

learning application; probably going to be multiple of them. Remote work, like, um, teleconferencing, there's probably like a whole other area where there's interesting things to do with voice AI, and even within customer support, we highlighted a number of companies we talked about last time—um, company Parel

Cap. Yeah, it turns out that customer support isn't really one vertical; there are many different flavors of customer support, and they're very different internally once you get into the details.  I think there are very specific types of workflows

you need to do per industry, and that's the point of why vertical AI agents are going to really flourish. I mean, the same thing for bots; it's just very different workflows. If you're building, I don't know, the voice agent to do customer

support for an airline, it's very different than doing it for a bank, very different than doing it for a B2B SaaS company, etc. Yeah, I guess that question of, 'Is there going to be um pure horizontal integration?' is sort of like saying, 'Will there only be one website?'

or it'd be like saying, 'There are just going to be both.' There will be horizontal infrastructure companies that do really well and vertical applications because to say otherwise would be like saying, 'Oh, like Stripe powers payments on the internet and it's also just going to have all the most valuable applications

that accept payments on the internet.' It's just not how it works. There's enough value in just being the horizontal infrastructure layer, so I'm sure there will be great voice AI companies that just make it really easy for you to build your own voice AI

application, while there will also be hundreds of really valuable vertical apps. What are... what are other trends that we've seen besides uh voice? We were talking about robotics earlier. There are certainly... we are certainly working with more founders building robots this year than I think any year

Ever uh, what's driving that? I have an X team that's called Weave Robotics. They're going to try to ship a real robot in 2025. It costs about $65,000-$70,000, but that's actually what it costs to have the actuators and the

safety needed to actually have it work in your home. I think it's actually driven by this idea that um, the LLM itself can be sort of the consciousness of the robot, like, "Am I doing this thing that, you know, my owner needs me to do?"

You know, how do I actually interact with them and the other people in the household? Uh, but it's funny because then the voice language action model that might actually do a certain thing, like fold laundry, that's uh, almost tool use inside of the broader LLM consciousness.

So I feel like that's one of the things that I'm excited to see. You know, will it really work? And I think we're going to find out this year. I guess the way I think about it, robotics is basically half AI and half hardware; half of the

part of the equation is starting to work well. The hardware is still hard; the hardware is still very expensive. Yeah, there's still some evidence that uh, being able to actually do laundry, for instance, like that might be one of the first things that gets shipped, I think.

The dream case for startups is going to be that you can build just the AI or the software piece of it and run it on commodity hardware and do really great things. The opposite case would just be actually if the two things—like if you need to be good at the hardware and the

Software and hardware are coupled together and need to be produced together. You'd expect Tesla to be the obvious winner in this space, but it remains to be seen. I'm pretty optimistic; we have multiple companies trying to be creative on how to run the models on commodity hardware for specific use cases. It still feels early; it feels like robotics hasn't quite hit its 'ChatGPT moment' yet.

Maybe the moment is self-driving cars. They've been working in San Francisco. I don't think it's talked about enough; people who don't live in San Francisco often don't realize the extent to which these are fully deployed in San Francisco, and regular people are using them every single day.

Yep. I saw Tony from DoorDash recently, and he said he exclusively uses Waymo everywhere. I live in Palo Alto and have no option for it, but I would love to; it'd be amazing. I mean, the wild thing is there are only a few thousand of these deployed right now in the entire world, in San Francisco.

Yeah. What about big flops for 2024? Uh, you know, I seem to remember that we started one of our Light Cone episodes all wearing Apple Vision Pros and Quests, and uh, we haven't talked about AR since then. What happened?

2024’s Biggest Startup Trends

Videos to help you build a successful startup. Subscribe for startup advice, founder stories, and a look inside Y Combinator.

What is Y Combinator?
We invest $500,000 in every startup and work intensively with the founders for three months. For the life of their company, founders have access to the most powerful community in the world, essential advice, later-stage funding and programs, recruiting resources, and exclusive deals. 

Visit ycombinator.com to learn more.

Video Team:
Zach Both - Video and Content Lead
Ryan Loughlin - Senior Producer

