00:10:12 · Elad Gil
[00:10:12] S0: And You know, so I wrote a post about this maybe two or four years ago where I tried to predict the foundation model market. And at the time, what I thought would happen was that you'd end up with three players, each of what each would be aligned with a hyperscaler. And it'd be kind of a captive hyperscaler model relationship where the hyperscaler would fund it and then it'd be on the... On that platform. And so I thought it'd be Google for itself with GCP and then OpenAI and Microsoft because there'd be relationship at the time. Then I thought Anthropic maybe partners with Amazon or something. And then I thought NVIDIA would be the primary funder of open source. And if you look at every technology wave, there's usually a big company that funds and monetizes the open source. So during the nineties for Linux, I was IBM, where they spend billions of dollars effectively on Linux development and really monetize it by selling it through services. And I I got it wrong. I got it partially right and partially wrong. I got it right in terms of thinking it'd be in a holopoly market. And the reason if I would be in a holopoly because of the scale of capital needed to do it. So, eventually, you just don't have new entrants really able to compete. But I got it wrong in a few ways. One is Meta and I've been The US based funder, for the time being, China... The Chinese government became the other means... Major source of economic funding for open source models. And then I got it wrong in terms of the models are partnering more aggressively across. Right? And traffic is part of the GCP and Amazon. Microsoft is working with multiple models. It's, you know, Google is working with Anthropic. So I... It's much more of a overlapping world than what I anticipated, and everybody kinda cross and busted. It's actually honestly closer to what happened in China across a few different waves of, like, the big Chinese companies effectively funding the versions of Uber and, you know, other sort of next gen applications. So I I do think the natural structure, assuming one of the companies doesn't truly lift off in terms of model capabilities ahead of everybody else's oligopoly, That seems to be a natural state for now. I do think one interesting thing that's emerged is the importance of the harness, relative to some of these applications like Claude Code.
00:12:09 · Elad Gil
I think a big shift there that I observed was when, Fi two came out for OpenAI. Was less switching than I anticipated because some people argued it was a bit of a better model. But people had already been using cloud in a harness that was really optimized for using cloud, and so fewer people switched than expected would. It could be the model's only incrementally better. It could be a variety of thinking. It's an open question to my mind is how much does the harness matter for stickiness now versus the core model. And if that transition happens, then that has really key implications in terms of how you think about things. And then the other piece of it that's the interesting question is when do the... If certain types of models asymptote and functionality... And if you hit the asymptote and other aspects become more important in terms of switching or not, if you don't hit the asymptote and it continues to be, what's the best model again with this overlay? Is it significantly differently... Different enough that you'll abandon whatever harness you're using? And then the question is, like, what's the enterprise version of the harness?
00:13:06 · Aditya Agarwal
[00:13:06] S1: I don't think anybody in this room could probably confidently predict whether we are kidding that assymptote because it seems as though we've kind of been in, like, a three to four month kind of cycle of, like, someone having a leap for the last almost three and a half years. So I I treat a lot about Claude Code, which means that I have people who are my friends at OpenAI Codex being like, Yo, can you switch and use it? And I'm like, No, I have a bunch of contacts and I know how to do these things. It's hard for me to switch. So I think the hardness thing is real. And it has also surprised me that there hasn't been as much chatter about trying out 5.2 or some of the more recent models. And maybe you're right, maybe it's incrementally better, maybe it would not of feel like I think Opus 4.6 came out pretty close to 5.2, so maybe all of that is contributing to it. But I guess the question in here is, do you do you think that we are running out of model iterations that are significant? I don't personally think so. What did you say today?
00:14:03 · Elad Gil, Aditya Agarwal
[00:14:03] S0: Yeah. No. I think there's tons of room. I think the the question is more what is the closeness between the models, and what is the utility you get through things at ancillary? And the ancillary things include the harness. Includes brand, actually. Yeah. Right? What you could tell your friends you're using. Yeah. Yeah. No. Seriously, I don't know if it matters. So I think I think it's a couple different things. You just get used to having things configured as your way, as your types of prompts, or, you know, everything is kinda set up for you. And I think it's a recent phenomenon. I think it's very new, and that's why I think it's not discussed very much as, like, a thing. [00:14:32] S1: As you kind of, from your vantage point, you know, obviously we've had, we're still continuing to, I would say, innovate along, okay, let's do, you know, longer chains of thought, let's do better RL, let's do longer context, let's do multimodal, let's do omni model, Let's do, like, world's models and so on. Right? So how are you... Like, when you kinda think about the vectors along which these model setting crew over Callaway into forty four month horizon, how do you think about At this stage, you might as well just go work for a foundation model company. [00:15:04] S0: Yeah. We all will be [00:15:05] S1: in the long run. Yeah. We are.
00:15:07 · Elad Gil
[00:15:07] S0: It all gets absorbed. Yeah. Mean, they work on a railroad. Yeah. It's very durable. So I think I think there's a few different types of models. Right? So there's the the core language foundation model, like the LLMs, etcetera. That's mainly what we talk about. And I think there... Two, three years ago, everybody thought, okay. This is gonna go over time. And for things, you need some form of persistent memory. And you need to have these sort of actions, and you need people to tie into these different... This is to integrate with these 30 thin... You know, like, I think it was kinda clear what to do, and it's just a matter of time to go do it. Right? We knew reasoning was coming because we'd have to be seen reasoning in other contexts. And... You know? So I I think I think a lot of it has been reasonably predictable. And, like, I remember talking to Eric Steinberger who runs magic. And three years ago, he was talking about super long contest windows being incredibly important for code because then you could put the whole code base in, and they could work across it. And, you know, so, again, I think a lot of the the smartest people had all these ideas that announced for Tintia. And then similarly, I think there's a very clear roadmap in terms of what's to come. For other areas, I think there's lots of really cool stuff to do. So are... You know, what are the best models for physics simulation? What are the best models for materials? What are the... You know, you kinda go through a whole by circle or area by area, and I think there's super interesting stuff to do. Now that may be separable for economic value. Right? And that's my big question about, some... A subset of the biomodels. I think there's broader things people are doing. But, you know, if you look at the cost of developing a drug, for example, let's say 1,500,000,000 and it kicks fifteen years, tens of millions of that is preclinical work. That's the molecule. And the other 1,450,000,000 is a clinical trial. Right? Now you may be able to impact the success of a clinical trial based on the molecule. You look at Admed and other aspects of developing drug, etcetera. But fundamentally, I I... You know, those companies, in my opinion, many of them will end up just being drug development companies. And the question is, do they have real systemic advantages?
00:17:02 · Elad Gil, Aditya Agarwal
And so those feel like tougher categories and maybe physics or language or image gen or other areas. And so I think part of it is like, what are you doing? How advanced is it? How interesting is it? How how much utility does it happen? Is it overlay is like, what is the economic model that you're dealing with relative to that type of model? [00:17:19] S1: Maybe two questions rolled up into one here, Elad. So I mean, how... I I I think that over the last six weeks, it's probably been more encroachment from the foundation models into a number of, like, different domains. Thought going out for profit, thought for legal, thought for finance, and obviously of all the coding related things. So I guess the two questions here are like, how do you as a startup, if you could start something, how do you think about what is it in the effective blast radius of what these models can do? And then secondly, I guess related to that is what are the durable moats that an application or a product can have? And I think one of the things you said in the past, which actually resonated with me a lot, was that data is maybe overrated as a mode. Everybody talks about we have we have proprietary data, and we can, know, do better, like, context window stuffing. And that feels a little thin to me and an argument. So different questions folded up into the same thing there.