Now please welcome to the stage Martin Casado, partner at a16z, for a conversation with Eric Newcomer, 'The Safest Stuff We've Ever Made.' Hey everybody. Um, hello Martin, is this working? Can you hear me? Yeah, okay, good. Yeah, hi. All right, in your prep call you called it the safest stuff we ever made. I love that. You know, there's a... I feel like AI is ruled by fear-mongering, whether it's sort of the existential risks or nearer-term things. Yeah. Why the safest stuff we've ever made?
Okay, so listen, that title is clearly edgy, right? It's clearly a complicated topic, but I'll make a couple of points. The first one is, um, like we got compute wrong for a really long time. Um, I mean, so much so that like we didn't understand what we were building and we killed people, right? There's a very famous paper called the Therac-25. If you haven't read it, I'd recommend you read it. It basically built this medical device, it required thread concurrency. Thread concurrency is hard. It overdosed patients and then they died, right? Like this is like a very well documented study, both clinically, you know, the internet... soon after the internet showed up, we had a new type of attack, right? We'd have viruses, we'd have worms. We actually changed the posture of the United States. We had a new doctrine. So we went from kind of mutually assured destruction to something called information asymmetry, where, you know, if you're using it a lot, you're more vulnerable than a company that's not, right? And so we've seen these epochs of new technologies that are very disruptive, and at those times you could point to like very clear marginal risk changes that resulted in very clear outcomes. But if you look at the discourse around AI, you're like, well, it drives better than humans, it saves people's lives, and like we actually haven't identified a marginal risk profile. And so instead, the discussions around risk are these kind of platonic hypotheticals of like non-real systems. And so maybe it is the safest, maybe it's not the safest, but for sure the dialogue is the most ridiculous. How's that?
I mean, you guys fought aggressively against the California legislation. Now with the new administration coming into the White House, do you think, I don't know, those fights are a thing of the past? Or how much, you know, I guess in Europe and California, are you still very worried about regulation slowing down AI? And how much time are you going to spend on the regulatory piece?
Um, I mean, I can't predict what's going to happen. I mean, clearly we do need a shift in policy, so I'm not like anti-policy, right? I mean, the reason that I was against, you know, SB 1047 is not because it was anti-policy, it was a paradigmatic shift in how we've treated tech policy in the last 40 years. So I think if you're going to change your paradigm for policy, you should probably have a good notion of marginal risk, you probably should have a good model, right? And if you don't have that, maybe we should draw on the learnings of 40 years, right? So listen, I can't predict how like, you know, the new presidency is going to change policy. I'm certain we should still both do policy. I think that's a positive thing. Um, you know, and I'm certain that this will continue to happen in California and in Europe, and just hopefully we'll have a bit more sanity around the conversation. Like listen, I actually think like the Biden executive order started us down this kind of weird path of like making up policy that wasn't all that sensible, that everybody kind of decried as like this great idea, which really wasn't. And then afterwards they're like, oh yeah, we got it wrong. But you always get these things wrong. And we're like, actually we don't always get them wrong. Like often we know what we're doing, we're just like, because it's AI, we're being a little sloppy. So let's just bring a little bit of reason to the conversation.
All right, let's talk about the technology. I feel like you're someone really good at zooming out. It feels like, I don't know, last year when Cerebral Valley started, it was like, oh my God, so much money's going into models. Yeah. You know, there's been an inflection... I have theories, I don't want to... how would you describe this moment, or what's sort of the earned zeitgeist of this moment in AI to you?
So I think it's very hard to reason about AI in general, and everybody tries to do it, and you just can't reduce it to a few aphorisms. So maybe I'll just tell you how I think about it. I think there's kind of two markets that are worth looking at, and they're very, very different. So the one that most people talk about is these large language models. Um, and like listen, they're huge, they're expensive, they're great, but it's also kind of a fake market because they're massively subsidized, right? So everybody is like, oh, the people who invest in them get revenue back, so they have an incentive to do it. Yeah, I mean, listen, like if I am Google and I get a 20% uplift on a massive business, that's like amazing for me. That would be horrible for a startup, right? And so a lot of people look at these large language models and try and extrapolate that somehow to the startup ecosystem, which is just a nonsensical thing to do, for not only because like the type of uplift you need for a larger company is different, but just the level of investment required. And these are backed by companies with hundreds of billions of dollars on the balance sheet, right? And so it's very interesting to track. Some of it is profitable, some of it is not profitable, but you can't like draw a conclusion on the entire industry from like what's going on with these large... That's one bucket. That's one bucket. The second bucket are, there's a lot of smaller models that are like in companies that are profitable and as fast as we've ever seen when it comes to growth, right? And these, like think about all the diffusion models, whether it's image or video or speech or music. Um, you're seeing a lot of interesting stuff on the coding side of things. So these models are much smaller, the training time to do them is much less. Like there are many viable businesses today, but for whatever reason, like people conflate these two types of things, and so they haven't, I think, appropriately segmented the market. So I'm very, very bullish on AI as a space. I just think that we need to be a little more reasoned in how we think about it.
Is it good to be a wrapper business today? Or I feel like last year that was the worst thing you could say about anybody. But then as the foundation models reached more parity, there were these smaller models, it was cheaper, it was maybe, oh, you need to be more... you have sales relationships, understand your customer, and so you could be a wrapper. Or how do you think about that tag?
Yeah, I just feel like that... it's almost a leading question if you use the word wrapper. It's like, that's pejorative, therefore it's bad. Yeah, well, no, it's not. I mean, it's... no, it's not. I just don't think any company would consider them a wrapper, even if you called it a wrapper, right? So like my mental model for this is super simple, which is we kind of see how like the capabilities of these models are scaling. So if you look at a company and the model continues to scale there and like it removes the differentiation, probably not the company you want to build. Now that said, just because you're building an application on top of a model, often, listen, I'm an enterprise guy, often the enterprise companies over time, like the value is the long tail of customer stuff, understanding distribution, you know, understanding the sales motion, understanding the features, and like that's not something I think a lot of these model companies are going to do. So yes, it is technically a wrapper on it, but if you actually look at the value of the company, the value of the company is the vertical understanding. Does that make sense? So if I come to you and I'm like, the value of my company is I am 10% better than like, you know, GPT-4, then you'll be like, well, okay, like GPT-5 is going to be probably better than that. But if my value is like, I take GPT-4 and I literally solve like this long tail, grimy enterprise problem, that's good. Yeah. So I think that, and I would encourage everybody here, the mental model should be just keep kind of inching the capabilities of the models. If that erodes the primary use case, then probably not a good company.
