The Truth About Building AI Startups Today [Lightcone Podcast Ep. 1]

Welcome to the very first episode of the Light Cone. I'm Gary. This is Jared, Harge, and Diana and we're group partners at Y Combinator and we get work with some of the best founders in the world. Jared, why are we calling it the light cone?

Well, in special relativity, the light cone is the path that light takes from a flash of light. You can imagine a flash of light and it spreads out in a cone shape. And in special relativity, you think about it spreading out in a cone both in the future, but also in the past. And in this podcast, we are here in the present, but we are going to talk about both the past and future of technology.

So that's how we came up with the name. And one of the things that we're all seeing is the encroachment of AI into almost every piece of, society at this point. You know, every business transaction, every, thing that we sort of use with computers, suddenly a new burst of technology is sort of entering.

everything we're doing. And we're seeing it in the startups that we're funding, which is why we're so excited about it. Think, you know, what what's the percentage of companies you've backed right now that have large language models in your summer twenty three was close to 50% of the batch. And that's pretty interesting.

Like, I think a lot of people, like, see that number, and they think, oh, YC must have funded so many AI companies because we have this thesis about AI. And, like, it's just easier to get into YC if you're an AI company because we just, like, love funding AI companies. And it's funny to us because we know how that's not true.

And yet that's probably what, like, 90 that's probably how 90 plus percent of people actually think YC works. Yeah. How does how does it how does it actually work? Should we tell people, like,.

how it actually works? Actually, it's interesting. The smart founders apply to us with what they want to work on, and we fund the smart founders, like Yes. Irrespective of what they wanna work on, actually. And and exactly. And so the fact that half the batch is working on AI.

says something much more interesting than just the YC partners think AI is cool. It's an emergent phenomenon of what the straw the smart founders wanna work on right now. It's like, where do they think there's the high beta to build the largest company? And I think the most ambitious and smartest founders.

are going after this because it's definitely, I think the exciting thing about right now with AI, I think it's, like, real. There's been a lot of waves for AI and multiple AI winters, but this one, actually, GPT 3. 5 and then four blew out of the water a lot of task and it impressed a lot of smart people.

When a lot of smart people start paying attention and building in this current idea maze, I think.

big generational companies are getting built as we speak. One thing I'm seeing that's interesting is I feel like a lot a lot more founders are dropping out of college to start working on AI. Because they don't want there's a FOMO. Yeah. There's like an actual, like and usually, it's so funny. My my interview question is almost always like, what's the rush?

Like, why do you wanna drop out of college? Like, why don't you just, like, graduate? Because it makes a lot more sense to graduate and then do a start up. And the reply is usually like, well, like, this might actually be, like, a once in a lifetime opportunity, and I I think I actually agree. And and the other cool thing is that this is an opportunity where college students.

are particularly well like, young founders are particularly well positioned to work in it because nobody has like, there's no one walking around with, like, four years of LLM experience. Yep. Yep. So, like, everyone is starting from the same playing field. And so if you can learn fast, you're gonna be at the same level as everybody else. That's right. And you know one.

an area I've seen that come to play is, like, developer tools for prompt engineering. I've been seeing like these sorts of tools are getting uptake. It's like ability to like chain together different prompts and test your prompts and see like the second order effects.

And actually a lot of college students are the people who are just like playing around with prompting models and seeing what comes out. And it's a really easy startup idea for them to just build the tools that they want. And the tools that they want are literally setting the standard for what every developer should want.

I know a lot of the headlines are all around AGI and all of the fancy stuff and then the really cool demos of multimodal AI, like AI generated video and this kind of stuff. The stuff that I've seen in the batches are actually taking off is a little bit more mundane. I probably say a lot of it's sort of like workflow automation.

It's finding things where there was a human doing some repetitive task usually involved searching for things or filling out forms and then using.

LLMs to replace that. It feels very obvious to us, the people who work at YC, that this is an amazing opportunity.

There are so many jobs in the world that are basically very mundane information processing, typically stuff that's hidden in some back office somewhere where there's somebody who's just reading stuff and summarizing it, reentering it from one system into a different system in a slightly different format. And it's such a perfect fit for LMs. LMs are perfect for this job.

And yet, we actually don't get that many applications for people working on this. And there's a lot of founders out there who are searching for a great idea. So if you're out there and you're looking for a great startup idea and you want to do something in AI, this is a good place to look into. I'll give you an example. So last batch, I had a company I worked with called Sweet Spot.

And we funded them.

The idea was something about, like, food ordering from food trucks, something, like, random. And they pivoted immediately looking for a new idea. And the idea they found was using LLMs to automate searching for government contracts to bid on. Contracts to bid on. Oh my god. Such a good idea. Yeah. And submitting the proposal.

That sounds so boring. What could be more boring than searching through a list of all the government contracts? You know how they found it? Is exploring startup ideas, and then they realized one of their friends, his job was to work for one of these, like, government contractors.

And his whole day was just spent, like, refreshing this government website to, like, find things and then submit a proposal. And they're like, what? Like, that's, exactly that. That's so boring. Like, wouldn't you like a tool that did this for Yeah.

And they launched and like pretty much straight out of the gate got like a pretty decent amount of traction because they're like opening up the people who would actually do it. Like it becomes easier to like find government contracts to bid on when it's all automated away and like software does it for You know, obviously,.

we all know that you know something that's boring is actually kind of awesome. Yes. But why is that? That's like you know, just off the bat, know, we have a sense that something that's boring might actually be an incredible business. There's an old PG essay where he talks about this, and he says he he quotes a phrase,.

where there's muck, there's brass. Nice. It's like it's as far as so it's almost like old English. Do you wanna explain it, Haraj? Just means that you you can find treasure in surprising places. Yeah. And and I think the cool thing is you have to go.

deep and vertical and solve a very concrete problem. Like, some of the problems with let's maybe talk about AI tarpets.

What a tarpet idea is is it's an idea that from the outside looks really shiny and attractive. It looks like a great startup idea. And so lots of founders go and they start working on it. And then you realize once you're in it that it's actually not a good startup idea, but but by the time you're there, you're, like, stuck in it.

And so it just attracts founder after founder, and they just get stuck in the tarpid idea. And we see this a lot YC because we see all these applications. And so it's really obvious to us when, like, 500 people apply to a YC batch for the same idea. But they don't know that 499.

other founders are also stuck in the same tarpid. What's tricky, I think, about tarpid ideas for AI is, like, we know something's a target idea in hindsight once enough people have been stuck in it. So with AI, it's so new we don't know yet. So I have a couple that I'm actually keen to get your thoughts on. A very common one is AI Copilot. So it's like, hey.

I'm gonna make it easy for people to, like, build an AI Copilot for their product or or service. It's this really unusual type of phenomenon where there's so much interest from potential customers to, like, want a copilot that it's actually quite easy to start getting, like, inbound leads if you pitch this. And if it's even easy to get people to pay you money upfront.

But what's really hard is to get them to actually, like, use the copilot because they don't actually know what they want it for. Like, just heard that AI copilots might be changing the feature of software, so we should have an AI copilot, but they don't actually know what their customers will use it for. I think for me, and maybe I just have a.

a mental block around chat interfaces, but I've never been that big a fan of chat because it puts so much of the emphasis on the user knowing how to speak to a computer.

And, you know, while in the next five or ten years, I think we will all get far more used to using it that way, I think the the low hanging fruit right now is just using the large language model to actually do the sort of knowledge work that a human being could do and then package it into the UI that, you know, whether it's a mobile app or a web app that is just familiar, like sort of what people use to do their work right now.

And it's, you know, basically the LLM is better used as sort of this, like, I don't I mean, it's almost like, you know, this thing that's sprinkled in that, you know, it it the software suddenly does something really powerful,.

but you don't have to change the way you would wanna use the software as it is. It's sort of like a an example of a phenomenon that, like, I I think we have seen in the past when, like, some technology gets really hot and all of a sudden, like, all these companies are like, they're being asked by people, like, what's our AI strategy? And they're like, oh, well, we better get an AI strategy.

Or, like, with crypto, they was like, oh, everybody needed a blockchain strategy. And even before that, it was like, everybody needed a mobile strategy. For a moment in time, it's like easy to sell them something that like placates their desire to check some box. But in the end, you've gotta actually make it successful for them. Like, otherwise, it's not gonna stick. I agree.

And so perhaps with this AI Copilot thing, maybe it's too early to call. Perhaps they actually will find product market fit, maybe with something that's not a chatbot UI. They'll keep iterating on the UI until they find something that's an AI Copilot people actually want. Or maybe it's just going to fizzle. Just like turns out most people don't need an AI Copilot.

Some of the advice I've been giving those those specific companies is.

the another old PG essay about if you if you're trying to sell technology to someone and they're not buying, like, see if you can just build a competitor.

And so it's like, hey, if you're trying to sell like fintech company a copilot and they're not buying it, well, like if you are convinced they should have a copilot, like why don't you just like build the company with the copilot as the main experience and see if you can outcompete them or not? I like that. That I like that. I think getting people to focus on the use case.

I think the problem is the whole thing with.

kind of the gold rush, people selling more the shovels and the tools. And even then in this case, it is a bit of that, but a lot of people aren't digging gold yet. Like, the reality is this is such a new technology and even the end applications that apply AI, the reality is they're so early. They don't have product market fit.

So it's sort of bit of a the blind leading the blind in here is like, what do I even know what the pattern is for Copilot? I mean, it sounds cool just to join the cool kid club of we're doing AI and we're gonna check mark. So I think that's the danger for a lot of these startup.

It's like it seems that they're getting traction as you mentioned, but then when you we poke them closer, is anyone actually using you? What are the actual use case? And then the founders come back and they start a blank at us. But look at all the sign up. Look at the revenue.

But then they're not really using your product. I mean, we're seeing even the second order effects. Right? So a bunch of us are funding dev tools companies that sell to AI companies, and they're selling tooling.

But then they might, you know, they might sell enterprise contract to someone who also upstream has a Fortune one hundred that said that they'd pay a hundred thousand dollars a year for that contract.

And then six to nine months later, that, you know, Fortune one hundred went back to the incumbent, you know, some other leading, you know, IBM, Salesforce, like, something like that because they ended up adding large language model technology to what they they were doing, and people just switched back.

And suddenly, the dev tool company suddenly realizes, oh, I had five contracts, but three of them went away because my customer actually lost their customer. So it's actually, like, sort of remarkable how fast this is evolving, you know, right now in 2024.

A specific type of idea I'm curious to get thoughts on here as well is offering fine tuning open source models sort of as a service broadly. That's a very popular idea I think over the course of 2023. Here's what I've seen. So like why do people want, like why is there any demand for a fine tuned like open source model at all? It tends to be initially I think the big driver was cost.

Like OpenAI, like ChatGPT was expensive and people wanted a cheaper version of it. And so I think it was very easy to get customers with the pitch of, hey, like we can fine tune an open source model and it's just gonna be much cheaper.

What I think a bunch of the companies in the space are seeing is that, like, that's not enough to keep the customers, especially because, like, OpenAI like, the cost of all of the models is just going down. And that's gonna keep happening with the.

OpenAI has a plan for all of those. So there's something more that all these fine tuning companies need to do. Yeah. It has to be better, not just cheaper. I think where it's exactly that, where I think is having more legs is when these companies need to customize it to private datasets.

So you have the open general big foundation model, but then you have to tune it up to specific datasets that, for example, a health care or fintech can't give out can't give out. And they don't have the team of experts to do it. So I think the one company that I think Brad worked with was Credle that kind of was doing that.

What are you seeing about so the concern around data privacy is another big reason.

Are you seeing that as being enough? Are people worried about giving these data sets to OpenAI?

It's really interesting. Whenever you have something so new like this, it actually sort of resets the clock on the competitive landscape again. So you almost can expect all the same things will happen again. Just as ten, fifteen years ago, cloud was brand new. And then you had cloud cybersecurity and Cloud strike and all these companies sort of come out.

You know, we're seeing the first wave of cybersecurity companies, you know, like PromptArmor. So they sort of wrap your API calls. And what they actually have figured out is that for a lot of large language models, if you do any sort of fine tuning or training with private data, you can actually just speak to the model and get it to spit out your private data again.

And they have a solution that stops it. So it's so interesting because it's entirely possible. They're basically creating a new industry again of cybersecurity for LLMs sort of in the same way that cloud opened up that space and created cybersecurity for the cloud. Yeah. I definitely think that a whole world of.

controlling within an enterprise in particular, like controlling who has access to, like, which LLM has access to, like, what data and who has permissions is, like, a really ripe space for building interesting software. I think the other exciting area that a lot of dev tools are getting built is getting more.

This is like a step further fine tuning, but more purpose trained models that are smaller. So take, for instance, a LAMA and getting those to run locally in machines for inference. And when you customize them, train on a specific domain and target data, it's gonna perform better than the general model. The general model was kinda trained on all of the human language for all of the task.

But if you wanted to build like the best, let's say, language model for parsing SQL queries, you would then parse very specifically just the set for SQL query. And I think some of those that are interesting companies that we fund is like Olama that you funded that's trying to make the development process for running all of these locally a lot faster.

And I think we're also funding some of these that are custom for coding. The thing that was surprised learning from some of the startups that are building, coder type of copilots, which I think is is a use case that's working out, making a lot of the workflow for programming a lot faster. It's kinda like autocomplete and copilot type of thing. They're training on older models of GPT.

They don't even need the newest one. And then I asked, like, why is that? And even for, like, one of the companies who fund that last batch, metalware for hardware, they're not using the state of the art model. Like the older GPT, I forget which one was like the older 2.

5 or three was sufficient in actually creating good enough results because the vocabulary for a specific domain for hardware or software is a lot smaller than the human language. So this is other world where the open model that's customized, I think, is gonna win and compete versus the big one for specific domains. So there's lots of companies with this. Yeah. That's what Toby Lutke from.

Shopify actually still dabbles with this stuff. I think he actually built the internal copilot for Shopify. And what he was saying is the best way to use whatever GP four, the, you know, latest closed source models that are most expensive and have the most parameters, just think of it as a prototyping tool.

Anything you do with those prompts, you can get your own model to do with a little bit more training. It's kinda like when people build hardware, you have the analogy of prototyping.

with FPGAs, which are very expensive. Right? And then when you have the right architecture for hardware, then you do the circuit path and actually do the custom SOC. So right now, for some of these tasks, the large language model is sort of like your FPGA, whatever, GPT-four.

And then when you customize it, you do like a super efficient one, coding path for, I don't know, Shopify for coding assistance and hardware, software, etcetera.

That becomes your SOC that you train and customize, which is cool. I think that pattern is emerging. It's like as I hear you talk about that, Diana, what's interesting, I just think it's just like so many different startups that could be built. It just feels like we've never had this moment.

At least I didn't feel like I've never experienced a moment where there's just so many potential startup ideas to be built, like, all at once. Yeah. There there absolutely hasn't. We we definitely saw this in the last batch with all the pivoting companies. Oh, yes.

People don't always realize this, but, like, many of the companies get into YC, within a month after we fund them, they're looking for a new idea because the old thing didn't work or they lost interest in it or something. And it's normally not actually that easy to find a great startup idea for a team to work on. But man, was it easy last Oh my god.

It was just like great startup ideas just lying on the ground. You'd trip over them. Yeah. That was a fast I think you actually had a tweet about it. That was it went pretty viral that talked about this is the.

batch the batch ever in your whole career working at YC where founders got to good ideas the fastest ever. Hardest been here even.

longer. Yeah. No. It definitely feels unique. I've never had so many successful pivots. Yeah. And Gary, to your point about the chat GBT rapper, I think back like, I feel like that meme really came out, like, just about a year ago. Yeah.

Let's talk about GPT rappers. Yeah. Like like, I feel like the first sort of group of ideas I saw in the batch were also generative AI ideas built on top of chat GPT. So it's stuff like, hey. Like, automate your marketing copy or automate, your creative creative content or something like that. And that term got thrown out, oh, these things are all just like wrappers on top of chat GPT.

And OpenAI is going to take all of that. It's just going build all of these things, and then we're going to release their app store. And it's just going take all their value, these things will die was sort of the meme. All of SAS software is just MySQL wrappers. Exactly. I think this is a great analogy.

You can think about any SAS product as basically a database wrapper. You can imagine like negging any SaaS product. Because like, the first version of a SaaS product is basically just a crud app. And it's just like you took MySQL and then you like built like a website on top of it. And I think people are going to look back on this term GPT wrapper similarly how.

think of how we would look at the term database wrapper, which just seems silly. I mean, is why I think the chat interface is wrong. I actually think there is value accrued to really great UX, good copy, good interaction design, information hierarchy, being able to approach a product and say this is the job to be done and for users to come in just sort of naturally understand what to do.

There is a craft to building software that is timeless and that sort of transcends whether or not you're using a large language model. And so that that I think is what I mean by these things are not SAS software is not a MySQL wrapper.

Well, here here'd be a question I'd interested in in in everyone's thoughts on. Suppose you're a new founder and you really wanna build a big company and you wanna do something on top of LMs.

How would you differentiate between an idea that could be a great foundation for a billion dollar company and an idea that is likely to get run over by g p t five and is probably, like, not a good starting point?

I think if a founder's working on something too general and not solving a specific need for a user, they can actually go talk to another use case. So I I worry about the ones that are too generic and building going after some kinda abstract.

It will solve all the things. Yeah. If it's like, hey. Like, throw your data in here, and we'll do, like, automations on top of it, for everything that's probably hard to compete with whatever one of the foundational models might offer. But if it's like, hey, we're give us your sales log data, and we'll.

spit back, like, suggested next actions that you can like, for salespeople to make them better at sales, that's probably gonna work better. Or give us all your compliance checklist to pass HIPAA compliance and process that. It's like, that's very specific and lots of business logic or give us all of your data for processing government forms. Right? Yeah. So it's a lot of custom business logic.

So the same thing with the SaaS era, a lot of the applications and how you build applications in there, there's always the separation of business logic and the crowding a lot of architectures for this app. And a lot of the value of the company is accrued on that business logic that is so custom per company. And there's a whole pattern of programming patterns on how people separate those.

Yeah. It gets as this all goes multimodal, this is gonna get really interesting. So early days, yeah, we've seen companies work on voice AI apps to be like a sales rep.

And I think it's an interesting example of the kinds of ideas that might be possible now with AI is where you take something like a Salesforce and you try and reimagine, like, what would Salesforce do if it were started today with all the power of AI? Well, it almost certainly do more than just be like a CRM. Right? Like, it would may like, it would find who your leads might be.

Like, maybe now it can make the calls for you. It could, like, set them up. Like, maybe it goes all the way to start, like, implementing, like, the first version of the product for them. Like, I think it just like, the scope of software you can build with AI now is so big. I think it's another good way to find ideas. Like, look at software today and reimagine it.

with the power of AI today. We should fund a number of companies that effectively are AI voice agents for small businesses because they receive I don't know. If you're like a flower shop or a AC repairman in the middle of The US, there's a lot of calls for you to schedule, and you don't have a lot of stuff automated.

And there's these YC companies that are using they're building these AI voice agents to basically be their receptionist.

I know one of our partners, Paul Buchheit, is quite worried about this, actually. He's worried about there's gonna be a world of just sort of, like, all these AI agents that are out trying to do malicious things and that we're gonna need, like, our own, like, good defensive AI agents out there making sure we don't get.

scammed out of all of our money. I mean, is actually why I'm so an advocate for open source AI because these things are sort of real considerations. You know, can you imagine there only being one hyper dominant AGI and it's totally closed source, it's owned by one company, and, you know, it's only available to the highest bidder.

And imagine you being someone who just had to go to the doctor and on the other end of it is some health insurance company that access and blocked it out from everyone else. And you getting on the phone, you're not able to navigate or go against the impenetrable AGI that is able to get around anything that your side might throw at it.

Like, actually want some form of actually equity at the AI level. Like, we actually want not merely the biggest companies to own the most capable AIs. We want all consumers to be able to have from the bottom up the same access to that same technology. And that's, you know, the best insurance against tyranny.

And certainly, that's actually what a lot of also not just founders, but smartest researchers who are really at the cutting edge. So I went to near Ips this past December, which was incredible to see the energy in there. The conference has grown so much. I think it's like over 10,000 attendees. There were 3,000 papers more than 3,000 papers accepted.

And I think back in 2017, it was only around 600 papers. When I went back in 2010, it was just in a ski lodge and maybe like a hundred papers. It's crazy, the kind of exponential growth. And one of the big topics of interest was a lot around AI ethics and regulation and how do we measure that. So that that was interesting.

But the thing that's different about typically, that was interesting in this conference is the amount of interest from researchers wanting to start companies too. One interesting data point is a lot of this era with GPT came about from one foundation paper. It's all attention you can need. It was this paper that got released got launched in New York Ips back in 2017.

It was a team at Google who was trying to figure out how to make machine translation between languages more cheap because English translation to any language is actually pretty good. But if you wanted to do, I don't know, German to Japanese, there's not enough data.

So they figured out this way to compress data, which became the transformer models for GPT, and it was, like, groundbreaking, and this is the foundation for LLMs.

That paper came out in 2017, and the fun fact, I was just looking this up, out of all those author eight authors, seven of them started different companies, and all of the companies in total, their rate, their worth valuation more than 6,000,000,000. Oh.

And now, people are seeing, oh, these like industry pioneers did this and it's creating this new crop of, I think, founders that I don't think would have started because I talked to a lot of AI researchers and I don't think they wanted to be founders. And I got a lot of this question, how can I turn my paper into a company?

Which I think is cool because this is like going back to the root of why I see a fund funding hardcore technical founders, and I think it's cool to see that energy there. So when we went and host our event, we I didn't plan, and it was, like, three x oversubscribed.

Nice. Standing room only. Yeah. Yeah. It's that sounds like really the new Homebrew Computer Club. So NeurIPS in December. Yeah. We gotta mark it on the calendar.

We'll come back. Yep. Diana, I love your point about how this is sort of, like, returning YC to its roots. It definitely felt that way last summer. Because when YC got started, the internet was really new. And the people who were building stuff on the internet were mostly technologists. It was actually pretty hard to build websites back then and pretty hard to build good software.

And as building software and building websites got commoditized, a lot more people came into the space. And this is a cool reversion back to the origins, where the people who are building the most interesting stuff are mostly really hardcore researchers and technologists because there's actually real new technology being invented.

It's not just innovating on business models, but like commoditized technology.

And again, just like every great technology is being dismissed. Right? So going back to like the chat GBT rapper meme. Again, I actually think that was great for YC because it meant we only got the people who are like tune who could tune that out. And we just say, hey. Like, either I'm just so interested in this technology.

I don't care, like, what the memes are or I'm just too busy building it to pay attention to the meme on Twitter, which is also great. But, like, the this I feel like this has always been the case. Right? Like, Homebrew Computer Club, like PCs are, like, dismissed as, like, toys, like, the Internet is dismissed as a toy, like, all all of these things. So it feels like that moment again.

Yeah. There is a a classic essay that I love that I saw off Hacker News. Do you guys remember this? It's geeks, mops, sociopaths in subculture evolution. And, you know, I I think that that actually is the one thing that's quite durable and, like, keeps returning. Right? It's always the geeks who are gonna be into the tech no matter what. They're on the cutting edge.

You know, I always think of Steve Wozniak talking about like, you know, we started Apple Computer with no idea that it would ever be a company. Like, we just wanted computers for ourselves and our friends. And so, you know, at some point the, you know, sociopaths come along and they start sort of monetizing the people who, you know, come to the scene and then the cycle returns and repeats.

So that's why I like being at the beginning of a new cycle and clearly AI is exactly that. So don't don't count it out, don't write it off. It's one of the most interesting things that are is happening out there. But you know there are clearly things to be careful of like don't be attracted to the new shiny thing. Instead look for the muck because where there's muck there's brass.

So that might be a great place to call it for the very first episode of the light cone. We'll see you next time.