Y Combinator Startup Podcast
00:10:31 · Vivian Shen, Paul Graham
[00:10:31] S0: And part of that is because he says, you know, it's really easy to raise money these days. AI lets you work on many things in parallel. You could spin out a bunch of agents and work on a bunch of things. Do you agree with him, or would you have a rebuttal? [00:10:45] S1: I've never really been clear what a lean startup is. I mean, we already had a kind of startup that we funded in YC, and maybe it has a lot of overlap with what was described in this book, The Lean Startup, but I haven't actually read it. And so, I mean, why would I, right? Tolkien didn't read all those other fantasy books people wrote. And so I don't exactly know what a lean startup is. But if you tell me what qualities you might think might be obsolete, maybe I could answer that. [00:11:13] S0: Let's say starting with just a little bit of money on a very focused area. [00:11:18] S1: You mean it might be dead starting with a little bit of I don't think so. I think there's a lot of things you can do for cheap. Mean, tokens are very expensive at the moment. But that's just because there's a shortage of GPUs. Token prices will certainly, inference prices at any given level of inference go down dramatically, like 30x a year or something like that. So even token prices are somewhat misleading because you're getting higher quality tokens over time. So I think, no, technology always gets cheaper. I think you can still start a startup on not much money. [00:11:53] S0: Even if it's a rocket startup or any of
00:11:53 · Vivian Shen, Paul Graham
[00:11:53] S0: Even if it's a rocket startup or any of [00:11:56] S1: the It's harder to start a rocket startup on no money. But even a rocket startup, can start on not much money because what you do is you adjust for how much money you've got. Like you can't build an actual rocket. What you can build is you can build a design for a rocket and show that to and you can simulate it, right, or show it to experts. If it's convincing enough, you can go to investors and get a bit of money to do the next stage. You're always able to raise another round, right? So you just do what you can on the money you've got. As long as you can get to some kind of milestone, then you can convince investors to give you more. [00:12:31] S0: Yeah. I think Philip from Starcloud was saying that they basically just wrote a white paper and booked a launch, and they raised the money for Starcloud. [00:12:39] S1: Yeah. Nothing as convincing as booking a launch, right? It's got to happen if you booked a launch. But he's a huge famous expert in the stuff he works on. Right? So he comes with automatic cred. It's easier for him to do that than some kid who just graduated from college. [00:12:54] S0: So let's transition to talk a little bit about AI more broadly. I'm curious what surprised you about AI.
00:12:54 · Vivian Shen, Paul Graham
[00:12:54] S0: So let's transition to talk a little bit about AI more broadly. I'm curious what surprised you about AI. [00:13:01] S1: Well, I was studying AI back in the 1980s, a very different kind of AI that would never have worked. I mean, was a joke. So I've been thinking about this a lot. And I had a lot. I spent a lot of time thinking about what AI would eventually be like. And the way we thought it was all going to play out was you start out with like a perfect fly. Right? You know what? It doesn't do much. It's like a fly. Like a fly's brain. Right? It can only do what a fly does, but it does it as well as a fly. Right? And then you work your way up from flies to mice, you know, to cats and eventually you get monkeys and then humans. Right? But at every point, you're you're perfect. You're doing what they do well. And instead what we got was basically a full on human but full of shit. Like the first AIs, the first versions of chat GPT were like an undergrad trying to bullshit his way through a paper. Something plausible sounding. Full human stuff, like not even this fly stuff. Full human, but like terrible. Right? And so you start... And so instead of starting with perfect and then working your way up to human, you start with human and then work your way towards perfect. So it's sort of the opposite quality ended up being the one that got optimized. That, nobody expected that. Nobody expected when the first plausible AI turned up, it would be like a bullshitting undergraduate. Right? [00:14:26] S0: Just the slot machine.
00:14:26 · Vivian Shen, Paul Graham
[00:14:26] S0: Just the slot machine. [00:14:28] S1: Yeah, and still has a lot of that flavor, doesn't it? [00:14:31] S0: Yes. [00:14:32] S1: I hear all these stories about how AI is solving famous open problems in math and I can't even get it to answer questions about when restaurants are open. [00:14:40] S0: That's what's called the jagged frontierness. [00:14:42] S1: Yeah. They must know things I don't know. Or they have research versions. Maybe they can find out when restaurants are open. [00:14:49] S0: The other thing, are famously very good at taking complex topics and making them sound very simple. So what is the official PG definition of AGI?
00:14:49 · Vivian Shen, Paul Graham
[00:14:49] S0: The other thing, are famously very good at taking complex topics and making them sound very simple. So what is the official PG definition of AGI? [00:15:01] S1: Oh. Well, honestly, I mean, could you do better than the Turing test? I think Turing was pretty good, right? One of the one of the ways I knew we were getting close to AGI was I actually had to go and look up the definition of the Turing test. That in itself is a piece of evidence. Like, what was the Turing test again? Are we there already? I think what's happened now is we thought back in the day that AI would be sort of like this finish line and AGI would be this finish line, we'd cross it, it would be very clear. But when you're standing on the finish line, you realize it has width. From back in the 1980s, it looks like it's just a line going across the horizon. But here you realize, okay, some bits of AI are way across the finish line. Some bits are like maybe in the middle of it. Who knew the line had dimension? The finish line is actually the sort of smear. And I think the best answer you can give is like, we're on the smear. It turns out to be a smear and we're on it. [00:15:57] S0: And that's why you can't get your restaurant times. [00:15:59] S1: Yeah. Because they're in a bad part of the smear. Whereas the Riemann hypothesis, maybe we're already there.
00:15:59 · Paul Graham, Vivian Shen
[00:15:59] S1: Yeah. Because they're in a bad part of the smear. Whereas the Riemann hypothesis, maybe we're already there. [00:16:04] S0: Okay. Changing back to talk about startups. So the other thing you've said is that the best predictor of success for a startup is the pace that they ship new stuff. Yeah. With AI, does that change? Do you think there's a new metric? [00:16:18] S1: No. No. There's... I mean, we have all these powerful AI tools, there's still a lot of startups in this batch that are not shipping fast enough. And so, obviously, there's variation in shipping speed. It's not just the rate at which you can produce things. You have to think of these ideas first. Right? Everyone always asks me, like, with all this AI, is anything still the same? And the answer is so far, almost everything is exactly the same. The only weird the only weird new thing is the companies have these giant AI bills. It used to be start ups' big bill... Big cost was salaries. That was all that mattered, salaries. Right? Everything else was cheap by comparison because it was just laptops and people. And now suddenly you got to pay for GPUs. Giant, giant. I mean, tens of thousands of dollars a day in tokens. [00:17:05] S0: I'm also curious, how does being in a batch like YC give founders an advantage?