EVERY SPOKEN WORD
35 min read · 7,142 words- 0:00 – 0:22
Apply to F24 batch
- GTGarry Tan
Hey, everyone. I have some pretty crazy news to share with you today. YC is doing the first ever fall batch. Applications are due August 27th and we fund you for $500,000. All you have to do is apply on ycombinator.com/apply. Now, let's get on with the episode.
- 0:22 – 1:05
Coming Up
- DHDiana Hu
NVIDIA became the most valuable company in the world. Who would have thought? There's been a lot of concerns with the different articles online thas- are saying it's, AI's over-invested.
- GTGarry Tan
The AI darlings today that might have already been blessed by the, you know, world's biggest VC funds, but they're looking at a balance sheet that has $100 million or $200 million or $500 million and absolutely zero revenue.
- HTHarj Taggar
I think this has captured, like, everyone's sort of imagination, but there's also just fear that it's unsustainable and everything's gonna pop and crash at some point.
- JFJared Friedman
Well, is it? Is it going to pop and crash at some point?
- 1:05 – 1:21
Intro
- GTGarry Tan
Hey, everyone. Welcome back to another episode of The Light Cone. I'm Gary. This is Jared, Harj, and Diana. And collectively, we have funded companies worth hundreds of billions of dollars, but when they were just, you know, one or two people, sometimes just an idea.
- 1:21 – 2:38
What the media is saying about AI
- GTGarry Tan
One of the things that we're doing a lot of these days is funding AI companies. These are some of the things that people are saying about AI now, that it's a hype cycle, that nobody's ever going to make money out of this, can you look at how much money is being put into NVIDIA, into data centers, there's just no way numerically that this whole space could possibly make money, this is dot-com all over again, this is the crypto boom and bust all over again. You know, doom and gloom, decelerationist type stuff. One of the memes that comes to mind for me when I think about market hysteria is this hilarious cartoon. It's, uh, pretty famous. I've got a stock here that could really excel. Sell? Excel? Sell? Sell? Sell? Sell? Sell! And then, the next frame is, this is madness, I can't take it anymore. Goodbye. Goodbye? Bye? Bye? Bye! It's just like this sort of madness that overtakes the market. We're also really familiar with the Gartner hype cycle, or we have our own version of it, which is sort of the life cycle of a YC startup, which is sort of the wiggles of false hope and then like long trough of sorrow. And then, you know, finally actually getting to the promised land.
- 2:38 – 9:32
Where are we in the AI hype cycle?
- GTGarry Tan
Where are we right now? You know, a lot of people, uh, who we saw who were just starting out in their careers asked us at startup school just a few, just a week ago, "Should I actually even be working on AI right now?" Which is, like, the craziest question from what I can tell. (laughs)
- DHDiana Hu
Where's the fear coming from, from all the found- founders that are looking for an idea? They're, like, looking at this, is this real or is this a hype?
- GTGarry Tan
When you're early in your career, you actually, you know, have maybe read about, uh, the hype cycles in the past, the- the $200,000 parties every Friday or Saturday night, just, like, free-flowing booze and just crazy bacchanals in San Francisco in 1999, 1998. Like, you read about it and you hear, oh, like, all of those companies died. And you sort of worry, is that, you know, where we're at now?
- JFJared Friedman
Yeah, I'd say it's been kind of a surprising experience for me because where we live in Silicon Valley with, like, our set of, like, friends and peers who we hang out with, like, people are basically just talking about AI, like, all the time, and there's a strong consensus that, like, this is an incredible moment in history. But when we all went to Cambridge a few months ago and met with college students, um, for startup school east, not that many of them were working on AI. Very few of them were actually thinking about AI. They were just, like, going about doing the same, like, regular old startups that we've seen college students start for, like, 20 years. And I was surprised at how big a disconnect there was between those two worlds.
- HTHarj Taggar
The thing that's unusual to me or novel about this current hype cycle is that I feel like startup world always goes through periods of, like, ideas are hot and you feel like everyone's, like- like, it's this sort of, that meme you just talked about, it's like people start getting into a- a particular genre of idea and then suddenly everyone's working on, like, Uber for X or their social mobile local app. This time around though, that's both happening for AI and startup world, but then if you look at, like, public stock market, it's al- like, AI is having a huge impact there as well. Like, the stock market is up this year, but all the gains are coming from, like, the big tech companies.
- GTGarry Tan
The magnificent seven.
- HTHarj Taggar
Yeah, the magnificent seven. It's never been this concentrated in history, I believe. And if you think of what's driving all of the gains in, like, the magnificent seven big tech companies, it's essentially all, like, AI hype. So I think, (laughs) -
- DHDiana Hu
(laughs)
- HTHarj Taggar
... like, I've never seen, like, the- those two things be in sync to this degree before where, like, the startup stuff and the YC batches are increasingly become, like, trending towards 100% AI at this point, and then the public market returns are essentially 100%, like, AI driven as well. Which is why I think this has captured, like, everyone's sort of imagination, but there's also just fear that it's unsustainable and everything's gonna pop and crash at some point.
- JFJared Friedman
Well, is it? Is it going to pop and crash at some point?
- DHDiana Hu
There's been a lot of concerns with the different articles online thas- are saying is AI's over-invested, and even in the public markets a lot of the pundits are concerned what are you gonna do with all these mega investments on AI chips? I mean, NVIDIA became the most valuable company in the world. Who would have thought, right? And then that's the, like, the bottom layer of the infrastructure. And then a lot of folks are wondering, okay, you invested in all this infrastructure, then something needs to happen to pay dividends for it, right? Sort of like in the early days-... the analogies with railroads. So you lay down the roads. Are there trains gonna come?
- GTGarry Tan
And I mean, one very extreme version of whiplash that I feel like, uh, is just racking all of us right now is sort of, remember about maybe a year ago, uh, when it seemed like there were just a few foundation models, uh, that were gonna run away with it, and there was the threat of becoming not just AGI but ASI, Artificial Superintelligence. You know, there's sort of this idea that, oh no, like, what if there is no opportunity at all left for literally anyone else? It, like, might destroy society. And then I think now with 405b out, Anthropic, uh, Claude, you know, sort of 3.5 Sonnet is really quite competitive, you actually have choice. And we're sort of swinging to this other moment where we're sort of wondering like, well, how does value accrue to foundation models versus, um, sort of the, the hosting companies? I think they're the big winner here. And then we hope that actually the software companies themselves, whether they're startups or in- incumbents are, uh, also just getting the benefit of these foundation models.
- HTHarj Taggar
Yeah. I feel like because everything is moving so fast, it's- it's easy to underestimate how important that point is. Like, if we roll ourselves back to the start of 2023, a few months after ChatGPT had just launched and we were starting to see the first trickles of, like, AI ideas in the batch, the- the ChatGPT rapper meme was really out there, right? Like, everybody was talking about how all these startups were gonna get crushed because ChatGPT and OpenAI would just own everything. Fast forward a year and a half later, it's like, it's clearly not gonna be the case. There's multiple models. You've now just got like the first, like, true frontier open source model coming out of Facebook, which we would never have predicted. Um...
- DHDiana Hu
(laughs) That's cool.
- HTHarj Taggar
Right?
- DHDiana Hu
Who would have thought the beating and best model was going to be the open source? Because it was trailing, uh, six months to a year behind what OpenAI was doing, right?
- HTHarj Taggar
I remember before this last, um, LLaMA release, like the four of us were just talking about it as, wouldn't it just be great if just the lag between like the frontier models and open source, like every time there was a new release, that open source could catch up, like within X months, and if we could just make X smaller and smaller, it would be very exciting. But to be at, like, essentially parity, I don't think any of us even saw that coming a month ago.
- GTGarry Tan
Yeah, we're there. I mean, there's that graph that the open source models were on an exponential and ostensibly the frontier models look like they're on an S-curve.
- DHDiana Hu
And what you're saying, Gary, about the difference in terms of what the current batch right now in summer is using for models versus, let's say, six months to a year ago, is very different. I do remember the numbers. Roughly, in the previous batches, I want to say like 90-ish percent, 80% of folks were using OpenAI models because that was the best and that's simply what was working. And now, uh, we did a casual survey and Sonnet, Claude 3.5, there's actually quite a lot of folks using it. Back- back then it was only like one or two folks, companies, and now there's like a couple dozen of them using it. And also LLaMA, it's like a lot more. So I think we're seeing the OpenAI model usage, uh, because all of these are becoming competitive and losing market share in the current batch as we speak.
- 9:32 – 15:15
Where does the value come from?
- JFJared Friedman
Yeah, Gary, I think you make a really good point about it being not clear where the value accrues. Like, even if you believe that AI is going to be massive and it's going to be like trillions of dollars of value created, there's still a great deal of uncertainty over who will capture the lion's share of that. Is it the GPU makers? Is it the hosting providers? Is it the model developers? Is it the application developers? Like, like which pieces get commoditized and which pieces become incredibly valuable? And it reminds me a lot of like both Web 1.0 and Web 2.0, you had the same phenomenon where there were, a lot of people were very bullish about the overall space, but it wasn't clear where you wanted to live in the space. Like, even if you go back to like Web 1.0, there was a massive hype over owning the browser. Like for a long time, it was believed that like the way to like become super rich in the internet was to like own the internet browser because that was the gateway to the internet, right? And like, Netscape was valued at like, it was like, uh, you know, I don't know, billions of dollars at the time. Turns out that wasn't the place to play. (laughs) But that wasn't obvious for years.
- HTHarj Taggar
I think the time is an important part 'cause if you think of, um, again, in our own world and our careers, the, some of the biggest companies YCs funded would be like DoorDash and Instacart being two examples, um, which were really driven by us all having smartphones and being, wanting to do things on our phone. Uber, obviously, like another huge company. Um, but those were all released, or at least like the first versions of those came out like almost I think around four years post the iPhone, like, coming out. Like these things take a while before you really know which ideas are gonna take off and where all the value's going to go.
- GTGarry Tan
I- I think, uh, one really important fact is that there are all of these other parts of the value chain. Obviously, there's the foundation model, there's the hosting provider, there's a chip maker, and then way over here are the startups we're funding, which is the application layer. And, uh, the important thing to note is that you do not need $100 million to start an application layer company. You just need you and sometimes just you, usually a co-founder, and then if both of you know how to code, you can literally take sort of these hyper powerful things that are now basically off the shelf and then go into some other market and you can create a product, solve a real problem, literally get money from people who are willing to pay you forever if you can actually solve their problems using the technology at hand. And you can do it all from the desk that or, you know, literally the computer that you're looking at this video from.Like, you don't need any permission other than literally a working internet connection and the laptop you have.
- JFJared Friedman
Which is exactly the story of Instacart and DoorDash, right? Those are application layer companies enabled by, like, the mobile phone technology ship, but they didn't need to, like, make their own phones. It was just exactly what you said, Garry.
- GTGarry Tan
Yeah, and every other part of it is like, yeah, maybe you do need, you know, $50, 100 million to get your foundation model company off the ground. Maybe you do need that much for fabs or hosting or all these other things. But even then, like, I guess that's not entirely true (laughs) . Like, it's just that much harder.
- HTHarj Taggar
I think also... So if we go back to this, like, are we in AI hype cycle question and try and, like... If we define it more precisely, no one's saying that AI has literally zero value. Like, it's pretty clear that it does. I think usually when people talk about hype cycles what they, what they're reacting to is they're seeing the price of things go up very, very quickly. Like, whether that's, like, public market stocks like NVIDIA, like you mentioned, or even in startup world where you see, like, you know, companies get to a billion-dollar valuation within, like, six to 12 months of starting, um-
- DHDiana Hu
Which has happened with some of these, uh, very famous AI researcher teams that used to work at DeepMind, let's say, or at OpenAI. They leave and start, and without product market fit, six months later they have this giant valuation. Uh, I don't know, right?
- HTHarj Taggar
Yeah.
- GTGarry Tan
It kind of reminds me of, um, Professor Coins back in the crypto boom.
- HTHarj Taggar
(laughs)
- GTGarry Tan
Like, if you had distributed systems experience and, you know, you could just walk into even, like, the first person you meet who is a crypto VC, and you would walk out with, you know, a billion to $5 billion market value, like, just without having even a line of code, even, like, a white paper. Like, nothing.
- HTHarj Taggar
Crypto is the perfect example, right? Like, that, that was the last time... And it hasn't even been that long ago, it's been two years. The last time it felt like there was a hype cycle specifically defined as, "Hey, it seems to us like the value of these companies is increasing at, like, an unsustainably fast pace," that was all these crypto companies that were, like, doing their token launches and just, even just raising equity rounds and seeing valuations, like, double and triple every three to six months. So I think there's definitely some scar tissue from that.
- GTGarry Tan
Yeah.
- HTHarj Taggar
On the flip side, that is also just the startup world. Like, I mean, there have always been startups that at every phase feel like they're overhyped and the valuation is too large and they'll never grow into it. But, like, I remember people saying that about Stripe when it first launched, because Stripe had raised this big round from Sequoia, valued at like $100 million, and I don't think they'd even publicly launched yet. It was all belief in the founders and the idea and the market. And so that is partly just the business, is investors invest, like, hoping to make money, and the way they make money is that whatever price they pay today is... Like, there's profit to be made by how much the company's gonna grow in the future.
- 15:15 – 17:51
Valuing a Tech Company vs. Speculative Assets
- DHDiana Hu
But there's a big difference that is maybe a bit nuanced between how to value technology company versus things that are more, like, asset speculative, right? Because in the crypto world there's a lot of that, so maybe we can speak-
- HTHarj Taggar
Yeah.
- DHDiana Hu
... to that versus true technology.
- GTGarry Tan
Yeah, I wouldn't say that it's purely speculation though. Like, I would say that when you have investors with billions of dollars to deploy, you have a hammer and all you see are nails. And what you're trying to do is, uh, put as much money as possible behind, uh, people and teams that seem like shelling points for other intelligent people. So that's probably what happened with Stripe. You could argue that it is actually somewhat rational, maybe not to, like, the, you know, billion dollar market values that people got. But it is somewhat rational that if you are a professor in distributed systems with a certain network, the other people you need to actually create, uh, you know, a working cryptocurrency that people trust and build apps on, that's actually the set of people who could do it. And then the money closes the door behind you. And I would s- argue that, like, that's some of the reason why you have things like Cognition Labs and Devin. You have, uh, you know, things like Harvey. That is the market trying to figure out where is the talent shelling point for really smart people. And like, if we all sort of go all in, it actually, you know, sort of buffers that company from, uh, frankly, competition. Like, and you, you know, it's sort of like the classic Peter Thiel, uh, you know, sort of speech about competition is for losers. I think it's literally on this Y Combinator channel actually.
- DHDiana Hu
(laughs)
- GTGarry Tan
And so it's not totally irrational, but it is, uh, worth lampooning at some level (laughs) .
- HTHarj Taggar
I- I think that's a fair way to look at the crypto stuff, incidentally. I think during that period you could divide it out into the asset speculation, which was essentially people who were really focused on launching these tokens very, very quickly and trying to get the price of tokens up. But there were also legitimately creden- like, very intense technical teams that were trying to work on building new protocols and trying to take, like, existing services and build decentralized versions of them. And I think the investor mindset was kind of what you're saying. It's like, hey, actually, like, we shouldn't bet against smart technical people trying to work on really hard technical problems. And there were lots of those within crypto, and so let's invest ahead of product market fit because, yeah, like, these people will attract the next, like, smart people. And if there's something here, these are the people that will figure it out.
- 17:51 – 21:41
Comparing the crypto hype cycle to AI now
- JFJared Friedman
So I- I think, I think this crypto analogy is really good because when Diana and I spoke with a lot of the best students at, at Harvard and MIT, one of the consistent things that we heard from them was that a lot of them felt burned by the crypto hype cycle, either personally or, or, or th- or they had friends. Like, they were, they were old enough to, to, to, to have been around for the 2020, '21, 2022, like, crypto hype cycle, and a lot of them were now a bit skeptical of AI and, like, w-... whatever the, like, new hotness thing is as a result. And so, like, you two were the- the founding investors of Coinbase and some of the most successful crypto investors in the world. What, how would you compare what happened with crypto in 2021 with what's happening with AI now? How is it similar? How is it different?
- GTGarry Tan
Coinbase, I think was remarkable in that it was not, um, actively a cryptocurrency per se, it was the enabling technology that you would need, uh, in order for it to actually happen. And they're still on that journey, I would argue. I mean, Brian Armstrong in earnings just this week started talking about, you know, we are actually working with or talking to every financial institution in the world to incorporate blockchain technology into, like, the actual norms of finance. And I think that that was always sort of the promise that I was excited about when I met Brian, uh, first when he was, uh, you know, an anti-fraud engineer coming out of, uh, Airbnb actually.
- HTHarj Taggar
Yeah, like Coinbase was essentially a marketplace, and I think it just- it didn't look too different to any other startup where it was like, oh, hey, there's this thing that people want to buy and sell to each other. There's no great marketplace for doing that, so why don't we just build, like, a trusted marketplace. Like it was-
- GTGarry Tan
And it was fringe and niche. It was n- I don't think people at the moment thought it was going to go the way it did. (laughs)
- HTHarj Taggar
That was it. The unknown with Coinbase was how big is the market, but it very clearly had product market fit with the people that cared about it early on. Like there were absolutely people who wanted to buy and sell Bitcoin that needed a better way to do it. So there was never any doubt that Coinbase had utility in the sense of, hey, there's this group of people that want to do something that Coinbase is making easier for them to do. I think what feels different about cre- uh, AI now versus crypto is just that like sniff test. So when you look at the products, a lot of the Web3 stuff in particular for most people just never really passed the sniff test of like, I don't get why I would use this. But when you look at like AI products, it's very, very clear that being able to like summarize a 50 page PDF market analysis report and actually pluck out, like, the three key points is clearly like utility that someone will pay for.
- GTGarry Tan
I mean, there's a company in my current pod at YC right now, and they can do accounts receivable. Like they took a 12 person team and it turned into one person working on accounts receivable and 11 other people could work on all the other things in finance. And that's like about as tangible as you can get. Like literally you're taking something that it's like not better than literally being a human passing butter and we're replacing them with butter-passing robots. And there's actually, you know, it's kind of, it's- it's a Rick and Morty joke, but on the other hand it's kind of serious actually. Like how much better is the world that software is doing that thing that, you know, actually is not the most awesome thing that a human could be doing, which is like reconciling like an email and a, you know, bank record to what needed to be paid. Like that's not invigorating like work that, you know, turns on your creativity, right? It's something that is p- sometimes maybe torture for people, you know? And we can give people other things that they can do. And it's- it's- that's the kind of thing that we get to see right now.
- HTHarj Taggar
Uh, we're literally seeing it, right? Like you, Diana looked up some
- 21:41 – 24:14
Increased ARR
- HTHarj Taggar
numbers here.
- DHDiana Hu
I mean the cool stat is, uh, we did mention this in a previous episode on, uh, batch by the numbers. The total revenue that, if we aggregate all the revenue that companies come in when they apply to YC, it was six million. By the end of the batch, about three, four months later, if we add up all the revenue that they grew into, it was 20 million. Which is like huge over just three, four months. And it succeeds, exceeds a lot of the advice that we give that's still ambitious. Like try to grow your company 20% month over month. That's still like an exponential growth. And if you use that benchmark, 20% month over month from six to... from six million over three or four months is only about like 12 million. So like 20 million. We're seeing real revenue that's actually getting accrued, uh, when companies land on a good idea. That value that you're talking about, Gary, their customers recognize it. I mean, customers are discerning, they're paying for it, and there is that kind of value aspect of it.
- GTGarry Tan
Yeah. And then I think the trickiest thing is it can't just be the, you know, f- the first renewal or the second renewal. Like you actually have to survive all the renewals. Like the only way there's enterprise value in a company is, believe it or not, discounted cash flows from the future. (laughs) So that means retention has to be like every customer you get, you better have that customer forever. And that's the only way you can build a business of any value.
- HTHarj Taggar
And that also seems to get to the heart of this, um, are we in a hype cycle argument or not? Because I feel like at YC all of us are on the ground floor seeing examples of companies that are ramping up to like a million dollars of ARR faster than we're used to seeing with companies during a batch, and even in like the six to 12 months post the batch. That revenue ramp is impressive. Gustav's been talking about Layer who just closed like their series A from Benchmark, um, doing like legal AI to automate legal work.
- DHDiana Hu
There's a company that I worked with a year ago that is probably gonna end at 10 millions by the end of this year. This is just 12 months later after they landed on the idea.
- GTGarry Tan
Man, 10X after that you could IPO.
- HTHarj Taggar
And I think this is so different to like Jared and I when we moved to San Francisco in like 2007, right? 'Cause there were, the thing that everyone was chasing, and you always had this preposterous too, like the thing everyone was chasing was like, um-
- JFJared Friedman
Page views. (laughs)
- HTHarj Taggar
Yeah, page views and then like active users and, um, just, I don't know, registered accounts, whatever it was, but it wasn't-
- GTGarry Tan
Vanity metrics is what you're saying.
- HTHarj Taggar
Yeah. (laughs)
- GTGarry Tan
(laughs) Like things that make you feel good, but-
- 24:14 – 28:19
Why a hype cycle might be good for founders
- JFJared Friedman
Well I want to- I want to jump into a point that Harj made earlier, which is like if you define a hype cycle as a situation in which a lot of assets are overvalued, that might be the case right now. Like NVIDIA is the most valuable company in the world right now. I don't know if it should be. Like-Maybe NVIDIA is overvalued and like, TPUs are gonna start-
- GTGarry Tan
Yeah.
- JFJared Friedman
... working and like, you know, like all- all- all- all those things could change. There's a bunch of startups that raised a billion-dollar-plus valuations. In hindsight, the- like, some of those deals will probably look really silly. The good news for our business is we don't care. (laughs)
- GTGarry Tan
Yeah.
- JFJared Friedman
Like, you know, if we were public market investors, you'd be very worried about, you know, whether NVIDIA was overvalued or not. But given, like, where we play, it literally doesn't matter. (laughs)
- GTGarry Tan
Yeah. I mean, I would argue for founders, it's good, actually.
- JFJared Friedman
And as founders, it- it- yeah, it's actually good and even if some things are overvalued, yeah.
- GTGarry Tan
Like, I mean, it's actually sort of free money for the whole ecosystem. Like, NVIDIA being overvalued equals if they needed to raise money, they could raise it very cheaply. It's more capital that allows them to drive the future sort of much faster. And then everyone in the rest of the ecosystem basically benefits from that. And this is where there is a huge difference between where we and the founders are and, like, public market companies are. Like, public market companies have to deliver quarter after quarter. Like, if they miss an earnings report, the stock tanks, employees start getting itchy. It's a massive distraction. But when like, you know, as YC we're investing in people when they're just ideas. Like, we don't expect the things- we don't expect to know if something's working or not for like 10-plus years, right? Which is the same for the founders and viewed on like a 10-year horizon, doesn't really matter if things are overvalued today. You just care about directionally, is this going to be worth more in 10 years than it is today? Well, I think the businesses are actually quite a bit more healthy or can be. Li- like, not always. Like, you know, I think maybe that's one of the more, uh, dangerous things about, uh, whether it was ProfessorCoins that raised hundreds of millions of dollars and like never launched, or, you know, frankly, some of the AI darlings today that might have already been blessed by the, you know, world's biggest VC funds, but they're looking at a balance sheet that has $100 million or $200 million or $500 million and, uh, absolutely zero revenue. Like, how do you actually- you- it's sort of like looking straight up at Mount Everest and saying like, "How am I gonna actually climb this?" And it's a weird contrast to a bunch of the YC companies that we see where, you know, yeah, they might have only raised a million to $2 million out of demo day, but they're starting to hit their milestones of five or $10 million. And that actually because they were profitable from day one or relatively quickly, the bank account has only gone up for them. (laughs) And they look down and say, "Well, I raised the seed round, I don't have any board. Uh, you know, I don't have to sell any more of my company, so why don't I just build the company the way I want?" And, you know, I think maybe ten years ago, there were not that many examples of it. In the YC world, you know, there was sort of Weebly that only raised a seed round and then sold for, uh, you know, a massive sum to Square later. Uh, and they could do it because they were profitable in our era of social software. Even today, there's- Zapier is probably one of the most dominant pure software companies that only ever raised a seed round and, you know, they're driving hundreds of millions of dollars in revenue. It's, uh, sort of like this insane moment for them.
- JFJared Friedman
And they're making a huge bet on AI.
- GTGarry Tan
Yeah. And so it's such an interesting moment. Like, do you really need the mega round? In fact, the mega round might be like such a giant lodestone around your neck, you'll never actually overcome that.
- DHDiana Hu
And the fascinating thing is with all these companies that we kinda mentioned, they are hitting profitability or breakeven and they don't need to raise and they have this- this revenue that is coming in and growing exponentially and they don't have to. So perhaps you don't hear as much about them because they don't go public and try to do the mega rounds because they're really playing the long game. And
- 28:19 – 34:43
Early signs things are working
- DHDiana Hu
if I think about the categories of things that we're starting to see now that we have a bit more data, I think in every single area for AR- for- for AI, we are seeing early signs that things are working. Even for some things that people think are over-hyped, like, uh, think of a generative AI to generate images. I think it's hard to imagine, oh, is this just for toy or just for art or just for fun? But there's a company, uh, Photoroom, that actually is making real revenue, um, and I think they recently announced they're valued at half a billion dollars and they're generating images for e-commerce. So when you go in a vertical specific like them, they figure out something and they got generative AI to be very profitable because it's very hard for brands to take pictures of- images of products and product placement and all that is very hard and they've done it very cheaply. You don't need a team of, uh, photographers, right? Or editors. That's like one example I could think of. The other category, uh, we talked a- a lot about this as well, is, uh, kind of this, uh, AI agent workflow, right? Where you're replacing a team of ops you mentioned.
- GTGarry Tan
Yep, like-
- DHDiana Hu
Greenlight.
- GTGarry Tan
Greenlight's a great example. Um, PermitFlow is another company that's doing essentially an AI agent that uses AI to just fill out construction permit flows. And again, if we- if we're trying to put like the devil's advocate, argue the other side about these things, it would be- I think there's two attacks you make. One, you say, "Oh, these things are like automating, but they haven't fully got the human out of the loop yet. And so like, you know, to unlock these valuations and to be worth this much amount of money, they really have to get humans out of the loop." Um, and the second would be that just enterprises are never gonna trust these things. And so you're not gonna sign up like six-figure or million-dollar contracts with big Fortune 500 companies relying entirely on AI.
- JFJared Friedman
Oh, I thought the second one that you were gonna say, Harj, is that like they're gonna be commoditized. That they're just like GPT wrappers and like all the- all the real value is going to be cued at the foundation models and there's gonna be like 100 PermitFlows, and so how will they capture any of the actual value?
- GTGarry Tan
Uh, on that one though, I just feel- I feel like that- that indeed was the big criticism. That was the main attack like a year ago. And I- I- maybe I'm just- this is based 'cause I so strongly believe it's just been debunked at this point.
- JFJared Friedman
(laughs)
- GTGarry Tan
But just everything we talked about, like multiple models, open source, like being PermitFlow being the winner in your space is, as far as I can tell, all about...
- HTHarj Taggar
... figuring out how to sell, getting the details of the UI correct, like, owning, like, all the little details to get the product perfect.
- JFJared Friedman
I agree with you, Harj. I was just, uh-
- HTHarj Taggar
Yeah, that's fair. (laughs)
- JFJared Friedman
I was making other people's argument. Yeah, it seems, if anything, that the opposite is likely to be true, that the value's going to create with the permit flows.
- DHDiana Hu
There's even, like, even to the, it's not just a wrapper. Even at the technical layer, I know they do a lot of work to really fine-tuning to, uh, domain-specific. There's a lot of power to private data in these spaces with permits, with banking data that cannot be done and replicated, and there's actually a lot of finesse there that they actually... it's not just a wrapper. There's, like, a lot of really thought-out work.
- HTHarj Taggar
That is a great point. I mean, there's a Mark Zuckerberg quote I think he made, he, he said this maybe a week or two ago about how even if all the model development progress froze today, there would still be five years worth of innovation to go on just, like, the application level, um, or the application layer of building point solutions on top of the models.
- DHDiana Hu
And even, uh, on the side of, uh, co-pilots, right? Even with, uh, GitHub Co-pilot, the most famous example, is apparently the fastest growing product for GitHub's revenue. It accounted for apparently 40% of the growth recently and has been re- rumored, reported that they're making in the hundreds of millions of revenue. That's still to be verified but around that much, which is significant over just, when was it released? A couple years ago, right?
- HTHarj Taggar
Another anecdote I could throw out on the point of this needs to get humans completely out of the loop to really be as valuable as everyone says it is anecdotally what I've heard from some startups is, uh, they're building it into the workflow so that the AI, like, does the work but then you can sort of spot, you have a UI to spot-check or just, like, have a human do the review process. But increasingly, like, the customers are not even using the feature anymore. They're just, like-
- DHDiana Hu
I heard that too. I heard that too. And I know this other company that's basically replacing a lot of the call centers, and they are processing hundreds of thousands of calls. And same thing, they're like all these BPOs, offshore f- call centers in the Philippines or Mexico, they fired that whole center and this big enterprise is using this startup for, like, 20 times, 100 times cheaper and faster, and same thing.
- HTHarj Taggar
And to your point, like, even if it's not perfect right now and there is not, like, another 10X frontier model coming out anytime soon, all of the work, like the fine-tuning, gathering the data, like, the private data repos, like, it's very plausible that alone will just squeeze out and get to, like, the right level of quality where big enterprises will just start, like, spending millions and tens of millions of dollars on all of these solutions.
- DHDiana Hu
That's actually the fourth category that I think is also working out. Uh, while yes, there's a lot of tooling right now to build LM, to do evaluations, fine-tuning, et cetera, et cetera, there are a couple of our company, a number of them that actually are having success with doing fine-tuning specifically, this tooling for enterprises that are working with their private data. If you want to skip the specific application, even the tooling is valuable.
- JFJared Friedman
I also think there's just, like, a lot more places to apply LMs too that people haven't even thought of. Like, Diane and I have a company in our group that just two weeks ago went to this conference o- of, like, particular industry conference and they had this, like, eureka moment where they realized that there's an entire industry that is a perfect application for LMs that l- literally nobody who's a technologist even knows about or knows exists. And so, like, there's no one who's even trying and they're just, like, uh, you know, like, off, off to the races with this, like, billion-dollar opportunity in an obscure corner of the world that's a perfect fit fo- for this technology. So as, as you're saying, Harj, I think there's, like, years worth of just unlocking of the current technology.
- DHDiana Hu
I think, Garett, you had a very good framework to kind of
- 34:43 – 37:18
The fog of war
- DHDiana Hu
think about this in terms of that Warren Buffett popularity and weighing machine. Could you kind of tell us about that?
- GTGarry Tan
Yeah. I, I, I think ultimately that's what's playing out here. I mean, it's played out in, uh, every hype cycle, boom-bust cycle in not just technology but sort of in the world. There's sort of, uh, this mania that begins, like, that the world is going to change, uh, and then it's so hard to actually understand what's going on that, um, a lot of it is in sort of hearsay or in the media and, "This is what X says," like, you know, on Product Hunt there's this. Like, there's sort of a fog of war around what's going on and then that, what that results in is a popularity contest. It's basically, uh, in the short term, all businesses are sort of voting, uh, subject to the voting machine, right? Like, it's not what is actually happening yet. Like, it's happening so quickly that, uh, humans are sense-making machines and it takes time for us to make sense. And so in the short term, we are tricked by very fast-talking hucksters. We are tricked by, uh, fancy credentials or, "You know, I worked at such and such company." They're sort of the Twinkles of the world, like the fast-talking Stanford dropout that, like, scams very smart investors out of millions of dollars, right? Like, you get a lot of that. And that's just, like, the vagaries and the madness of the voting machine, right? Like, it's not like people are trying to put money into things that are scams. It's more that, um, just by purely social effects we cannot make sense of the world fast enough and, you know, basically the mob gets it wrong. (laughs) And then on the flip side, like, in the long run though, like, ultimately the value of every company is, uh, discounted cash flows from the future. You know, you need your customers to actually have a problem be solved, people pay them and then, uh, you know, your customers stick with you forever. That's why when you look at a Google or a Meta or, you know, any of the Magnificent Seven, those companies are the most valuable companies in the world because people feel that they're... you know, those companies may well make money forever, right? And there's safety in that. That's sort of, like, what people, you know, believe now. And then the public markets ultimately are sort of, uh, you know, they themselves are both totally crazy voting machines as well (laughs) but they eventually resemble weighing machines. Like, ultimately you actually have to make money, ultimately you actually have to have customers and at that point you need to have made something of actual heft and weight that actually works.
- 37:18 – 37:35
Outro
- GTGarry Tan
Well, that's all we have time for today. We'll see you guys next time.
Episode duration: 37:35
Install uListen for AI-powered chat & search across the full episode — Get Full Transcript
Transcript of episode _-5xJQ4U8g0
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome