00:51:29 · Jordi Hays, John Coogan, Max Kan, Andrew Megalaa, Dylan Patel, Harley Finkelstein
[00:51:29] S0: the show. [00:51:30] S1: Thank you. [00:51:31] S2: Thank you. [00:51:32] S0: On the left, introduce yourself because I think it's the first time. And and then I don't know who this other guy is, but I think he I think he's a big yapper. So I'm looking forward to hearing from him too. [00:51:44] S2: Yeah. Yeah. For sure. Well, first, just thanks so much for having us, guys. My name is Max. I co lead the tokenomics team here at SemiAnalysis. Tokenomics is just the top of SemiAnalysis' full stack coverage. We focus on things like the AI labs, the clouds, the data companies, application AI companies, etcetera. [00:52:03] S0: Amazing. [00:52:05] S5: How's it going, guys? I'm Andrew. I recently joined the Telkonomics team working with Max here. [00:52:09] S0: Amazing. Who's this other guy? [00:52:13] S3: Hey, guys. My name is Dylan. I'm a big loser, so [00:52:16] S4: thanks for having me on. [00:52:18] S0: Awesome. You guys... Yeah. Let's let's get right into it. Why don't you introduce GDPX, and then we can go into some of the takeaways from it.
00:52:18 · Jordi Hays, Max Kan
[00:52:18] S0: Awesome. You guys... Yeah. Let's let's get right into it. Why don't you introduce GDPX, and then we can go into some of the takeaways from it. [00:52:27] S2: Yeah. For sure. So GDPX is, in our opinion, the single best benchmark for measuring frontier model capabilities. We think previous attempts at comprehensive benchmarks, most notably OpenAI's GDPval, are sort of fundamentally misguided because they pick their domains based on percentage of GDP or using things like the SOC or ONET taxonomy. Whereas we believe that sort of the first trillion dollars of AI revenue are going to be extremely concentrated in certain domains, the ones that can sort of move fast enough to actually adopt this new technology and also sort of have high enough of a skill ceiling to continue using frontier intelligence at high ROI. And so we made a benchmark focused exclusively on those domains, which are things like chip design, biology, cybersecurity, RSI, etcetera.
00:53:24 · Jordi Hays, Max Kan, Dylan Patel, Andrew Megalaa
[00:53:24] S0: Okay. So my first question is how how did you think about the gap right now between AI's ability to execute on tasks and do them in an effective way versus roles? Right? Because I think, like, as you look out in... As I look out into the next year, and especially right now with the sort of market jitters that we're having, it feels like we need a capability jump in the same way that we had. Like similar to what we had in let's say Q four of last year, where there was like this momentary correction because everyone was sort of had this idea that chatbots were gonna create trillions and trillions of dollars of value, and then there started to be some skepticism, you remember there was sort of like very momentary like correction and then we got agents and everything ripped again. And it feels like right now, like my question looking at this, you know, at the benchmark is like, are you guys thinking about, obviously there's a lot of room to run for all the models. But how are you thinking about broad capability increases versus just scaling up intelligence in these specific domains? [00:54:36] S2: Yeah. So I actually don't wanna go into this too in-depth because we're currently working on a demand model to sort of project '27 and '28 OpenAI to drop off revenue, which is kind of... Let... It's gonna lay out in full detail the SemiAnalysis. [00:54:51] S3: I do think we can say some things, though. Right? [00:54:53] S2: Yeah. No. I'm not gonna give the overview for sure. Or you can go ahead if [00:54:55] S5: you want to.
00:54:56 · Dylan Patel
[00:54:56] S3: Yeah. I was just gonna say, like, I think it's very... Like, people people tend to think that AI will diffuse very broadly and and that everyone will use a ton of AI and that's gonna drive AI spend, but we have a very different view of this. Right? We don't actually think that that's the case. We think that there's a lot of challenges to AI being adopted this way. We think that the general populace will not be able to effectively use a ton of AI in their tasks and given the fact that AI at a given level of capabilities falls anywhere from a 100 x to a thousand x in cost a year because of the insane progress that, these these firms are having in research, data, compute, etcetera. It it it means that, you know, the the places where you actually can absorb all of this usage of AI is where, you know, sort of people people can take extremely large amounts of intelligence and get ROI from them quickly. Right? Replacing white collar work is something that will happen, right, but that's not something that's gonna drive the first trillion dollars of AI spend. Those are not industries that are going to diffuse AI capabilities quickly. It's gonna take them quite some time to, you know, reorder how their OpEx is spent. Right? Because companies just have a lot of inertia with regards to, you know, people's jobs, you know, changing how their compensation works, changing how many people work on what. There are areas where we think that intelligence is very, very important, though. These include areas like chip design, cybersecurity, financial services, life sciences, material science, you know, software engineering. And in that, you know, there's sort of two... You know, there's... You can absorb unlimited intelligence, essentially. And there's two reasons for that. One is that it's just a domain where humans aren't anywhere close to the frontier, and there's just so much to discover. Think material science. Think chip design. Think anything in the scientific realm. Right? We figure out new stuff. We deploy it. We make everyone lives better. There's the other world which is adversarial. Right? Think specifically finance and cybersecurity.
00:56:59 · Dylan Patel, Jordi Hays
It's human versus human, and we're constantly just gonna be leveling up, either trying to arm each other, you know, like quants do and, you know, make make markets, or it's gonna be, you know, constantly trying to hack and then defense, hack and defense. And so there's this, like, constant demand for intelligence because when you have more intelligence than the other parties in the market, you win. [00:57:23] S0: Yeah. So so specifically though on on, let's say, like, material science, like, if I throw... If if I spend a billion dollars on tokens, and I come up with a bunch of a bunch of, like, basically, like, ideas, there's potentially, like, a duration. I I might have, like, a duration problem where I have hundreds of, like, sort of promising ideas, but then that creates those ideas mean that I actually have to put spend behind each idea to sort of run them down and see if they're viable. And there still could be this massive multi year lag. And so I'm wondering, my question is what is the willingness to spend for some of these companies that obviously have massive balance sheets, they could either be heavily funded startups, or they could be large legacy players in these domains. But I'm struggling to see how you can get a $100,000,000,000 of spend in a domain like this, just because very quickly you're not in a situation where you're like, we have a shortage of ideas. It's like we have a shortage of capital that we can put and and point at these different experiments. Does that make sense?
00:58:36 · Dylan Patel
[00:58:36] S3: Yeah. So one one thing that I think we can sort of point to is there's there's two domains, right, when you think about science. Right? There's discovery and there's engineering. And everyone always talks about the discovery side of things, but actually the bulk of the spend is on the engineering side of things, and you have to get really good at the engineering before you can discover something and bring it to market. And so when you look at, for example, Periodic Labs, they are a company that's specifically focused in this realm. Most of their work is on engineering. Right? Increasing the engineering velocity of these materials. Hey. If I make something, you know, this sort of deposition technique 10% more efficient, is that gonna be... What is the benefit to humanity? And, actually, it's not about inventing a new material. They're, of course, going for that as well. But if I if I if I improve the, you know, the... Let's just say, you know, in semiconductors, the deposition of something, and so then your chip has less resistance, or you can shrink it a little bit further, That's so much economic value that it's immediately delivered. Or, hey. I'm a battery company, and I'm optimizing my, you know, you know, my anode or my cathode or whatever it is, and I've made it have 5% more capacity. That is so much economic value delivered to society. And it's and it's a bunch of small engineering. And the big problem here is when we look at software engineering, everyone's, like, used to pretty advanced software engineering techniques. But if you go look at the broad space of any of these companies in batteries and semiconductors, many... In in in in automobiles and aerospace, all of their tooling sucks, and their research velocity is slow. Their engineering velocity is slow. And so, you know, using AI to create simulations to help, judge all of the data that's coming out. You know, an Intel wafer, you know, they've made this comment. Intel, when they make a wafer, they have about a 100 terabytes of data that gets generated, and they don't do anything with the vast majority of it. Right? There's there's small loops that get closed. Right?