The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch
00:26:07 · Rory O'Driscoll, Jason Lemkin
[00:26:07] S1: I pretty much a 100% agree with you. I mean, the zoom out comment is this. It's a new model. It's different than an LLM. An LLM returns text and is very computationally intensive, and in computing terms, slow and expensive. This just returns true or false or a ranking or a score, right? So there's three answer types, and it's super fast and super cheap. They don't even charge for output tokens because they're so meaninglessly few. So it's just a faster, more precise. That's why it's called... They refer to it in turn on the website as a system one, which is the Daniel Kahneman thinking fast, thinking slow. And this is the thinking fast part, quick, fast answers. But Jason's math is exactly right. But we're talking about this on Monday, the part of meeting is that there's roughly a $100,000,000,000 today being spent on LLM calls between Anthropic OpenAI and the open source models. Maybe that goes if we just agreed that OpenAI is gonna be doing 350, Amazon, 250. So it's probably gonna go to $500,000,000,000 five years from now. 20%. And, Jason, we had exactly the same number. Only 20% of them are relevant to this, where it's... Today, you're using a complex expensive model for something that really needs a much simpler solution. But 20% of a $100,000,000,000 is 20,000,000,000. [00:27:21] S0: Well, I don't think I don't think the dollar is gonna flow that way. I think the tokens will... Don't think the
00:27:21 · Jason Lemkin, Rory O'Driscoll
[00:27:21] S0: Well, I don't think I don't think the dollar is gonna flow that way. I think the tokens will... Don't think the [00:27:24] S1: dollar Let me finish, damn you. The next thing... Because you you got... You sound like the cynics of my group on Monday. We had this... I've already had this discussion once. That's why I can carry it off. So that's 20,000,000,000, but you're exactly right. As one of my partners said, but, dude, you just said the prices are going to down by five x. So that 20,000,000,000 is going to become 4,000,000,000. You're exactly right. It's just like open source. It's gonna take a slug of the total addressable LLM marketplace, attack it with a better, cheaper product at one fifth the price, net result at 20,000,000,000 becomes 4,000,000,000. But if you're a JEV or if you're a TypeSave, you're saying, hey, that 4,000,000,000 becomes mine. And that's the bet. Now, at the same time, going back to your naming concept, which I love, Jaya from Foundation did a nice post at Biggs. This is just a slug of the total 100,000,000,000 spend that was kind of automatically destined to go to OpenAI and Anthropic and is now kind of sluicing off into a cheaper low cost provider. That's what's happening here. You're right. It's not the end of foundation model. The clever shit is still gonna be done using foundation models. It's just a little maiming of 10, 20% of the revenue that's now gonna be done by someone else at one fifth the cost. Now we can talk about competition in a second because I think that's a real issue, but that's what's going on here.
00:27:24 · Rory O'Driscoll, Jason Lemkin
[00:27:24] S1: dollar Let me finish, damn you. The next thing... Because you you got... You sound like the cynics of my group on Monday. We had this... I've already had this discussion once. That's why I can carry it off. So that's 20,000,000,000, but you're exactly right. As one of my partners said, but, dude, you just said the prices are going to down by five x. So that 20,000,000,000 is going to become 4,000,000,000. You're exactly right. It's just like open source. It's gonna take a slug of the total addressable LLM marketplace, attack it with a better, cheaper product at one fifth the price, net result at 20,000,000,000 becomes 4,000,000,000. But if you're a JEV or if you're a TypeSave, you're saying, hey, that 4,000,000,000 becomes mine. And that's the bet. Now, at the same time, going back to your naming concept, which I love, Jaya from Foundation did a nice post at Biggs. This is just a slug of the total 100,000,000,000 spend that was kind of automatically destined to go to OpenAI and Anthropic and is now kind of sluicing off into a cheaper low cost provider. That's what's happening here. You're right. It's not the end of foundation model. The clever shit is still gonna be done using foundation models. It's just a little maiming of 10, 20% of the revenue that's now gonna be done by someone else at one fifth the cost. Now we can talk about competition in a second because I think that's a real issue, but that's what's going on here. [00:28:35] S0: I just think it's a hundredth the cost, but I agree with all
00:28:35 · Jason Lemkin, Rory O'Driscoll, Harry Stebbings
[00:28:35] S0: I just think it's a hundredth the cost, but I agree with all [00:28:37] S1: of You're right. A 100... Which which... In which case... Yeah. [00:28:40] S2: We're gonna see, like, the unbundling of ChatGPT, and we're have users go, oh, well, this one [00:28:44] S0: would be No. Good [00:28:45] S1: The the the problem is not not ChatGPT, the app, but the OpenAI API and the Entropic API. You're right. This is a developer product. You as an end user, I as an end user, don't use it. You can go on and try and use it, but it's full this morning, I tried. But it's really, I mean, if you listen to the launch announcement, it's very focused on developers. And the idea is, and this is an interesting thing, almost comes back to Muse, that I think you might see, I could be wrong on this, that the needs of humans, especially consumers for AI, are gonna continue to diverge from the needs from software developers for AI. And this is a core software developer thing. You'll never need it, Harry. But someone who's building a software app might realize that a significant portion of the core intelligence they want is system one intelligence. Just tell me if this is an a or a b, you know, animal, mineral, or vegetable. Is eggs better than y? Give me a quick answer. I don't want to blather and have you tell me that's a great question, Rory, like a sycophantic LLM. Just give me the damn answer. So it's not for you, the user, but it's unbundling at the developer level where it's more likely to happen. [00:29:50] S2: You're right. I require far more intellectual answers.
00:29:50 · Harry Stebbings, Rory O'Driscoll
[00:29:50] S2: You're right. I require far more intellectual answers. [00:29:53] S1: No. You just require a bundled product. Wait. Wait. Actually, Harry, I think what you really like is the sycophantic part of the LLM when they tell you, you're so smart, Harry.
00:30:01 · Jason Lemkin
[00:30:01] S0: But to answer your question, what I did learn from using JEV all weekend long, failed on Saturday, figured out the prompts and the use case, got it to work on Sunday. So cheap it doesn't, you don't even to measure it 10 times faster is the use case, but only for these, for these, for these limited use cases. What I did learn, may almost sound tangential is man, using a harness to pick a model and getting it right is effing exhausting. This is going to get even harder and harder and harder. When should I use JEV? When should I not use it? What about... You have to run so many evals, so many tests because there's so many things that I do with JEV. You know what doesn't work in JEV? Between Harry, between Rory and Jason, who's the better person to join twenty VC as a partner? It can't answer. You'll find it's gonna get that wrong. Okay. So between that and should Rory and Jason meet for coffee, it's gonna get that one right. It's not gonna guess. It's actually going to know, hey, we're both in the Bay Area. We've known each other. We should meet for coffee. You have to QA and test every single use case to get those benefits. It's exhausting. And then I got switched over on Replit to Autorouter where switches between Astra, Fable, the open source ones and everything. Now I don't even know which one it's using. And then I saw some performance degradation. So then I had to switch it all back to Astra. Do you trust your harness? Can you QA 10,000 uses of JEV? This isn't like a bad thing, but this is like the renaissance of DevOps or something. Everyone's gonna need this massive team to optimize it. And the needs are going up with JEV and friends. Like your team's gonna have to get more. It's not just benchmarks and evals. Like we're gonna be running these 20 fourseven across countless permutations and it's good, but it's for me it's it's too much. Like I can't do it anymore. I can't pick these models anymore. I'm tapping out. These harnesses are only so like Databricks is it's easy to save money with the harness. Hey, we route 80% to open weights models and now we're doing 10% like this. It sounds great to the CFO, but I found it doesn't work for me. And a lot of folks were saying they're using JEV as an instant router. It's like, JEV, JEV, you make the decision. Here's what we're doing. You pick the model in milliseconds. Great.