The Exchanges

Q like right now I feel like we're basically, and this is partially the cost of the models and the cost of you can compute for them. We're asking every question we can come up with to a really, really expensive PhD. And in the future, I think we'll be a lot smarter about who, what level of intelligence we need for different questions. How do you see that competitively playing?

A I might disagree. I think what will happen is the expensive PhD will become dirt cheap. My bet is, a year from now, it'll cost one-fifth to one-tenth of what it does today, and my advice to all our startups is ignore the cost of compute, because any assumption you make, any dollar you spend on optimizing your software will be worthless within the year, and so forget about it. Rely on competition in the marketplace between thought and Gemini and OpenAI. To, to, to reduce that cost to a point where it doesn't matter. And below a certain level, it doesn't matter. You're paying a certain amount for your iPhone service monthly. If it's 10% or less of that, it doesn't matter. Yeah.

AI assessment note: “I might disagree. I think what will happen is the expensive PhD will become dirt cheap.”

Q Apple, big announcement that we have all been kind of waiting for, the unveiling of Apple intelligence. Uh, that's a great place to start this week. And of course, A big part of that was Apple's partnership with OpenAI, which Vinod was the first venture capital investor in way back when. So, so I got to ask Vinod, what, what did you make of all the news out of Apple?

A Well, first, I haven't caught up with all the news. That was reserved for the weekend. But Apple needed to do something. Siri was starting to go from a reputation to worse, if that was possible. But more importantly, They've demonstrated something that's really important. Computers interface with you directly. You don't have to futz around to get stuff from computers. Computers get you stuff. And I'd done a blog post in this, on this topic a couple of months ago in the information, this idea that computers will learn humans instead of humans needing to learn computers. And I think we were seeing the beginnings of that. That was exciting. Obviously from an open AI perspective, it clearly establishes where open AI is in the competition for the best LLMs. Obviously many people wanted that business. And many things were evaluated, and so it's really rewarding to see where Apple thinks the best AI is, and, uh, and I'm sure they considered where the best AI will be a year or two from now. Uh, so in many ways, validation for OpenAI and a really big milestone in how humans interface with machines.

AI assessment note: “in many ways, validation for OpenAI and a really big milestone”

Q like, right now I feel like we're basically, and this is partially the cost of the models and the cost of you can compute for them. We're asking every question we can come up with to a really, really expensive PhD. And in the future, I think we'll be a lot smarter about who, what level of intelligence we need for different questions. How do you see that competitively playing?

A I might disagree. I think what will happen is the expensive PhD will become dirt cheap. My bet is a year from now, it'll cost one-fifth to one-tenth of what it does today. And my advice to all our startups is ignore the cost of compute. Because any assumption you make, any dollar you spend on optimizing your software will be worthless within the year. And so forget about it. Rely on competition in the marketplace between thought and Gemini and open AI. Uh, to, to, to reduce that cost to a point where it doesn't matter and below a certain level, it doesn't matter. You're paying a certain amount for your iPhone service monthly. If it's a 10% or less of that, it doesn't matter.

AI assessment note: “I might disagree. I think what will happen is the expensive PhD will become dirt cheap.”

Q But Vinod, isn't each large model successively that's coming out costing like an order of magnitude more to train?

A It is costing an order of magnitude more to train, and that's why I don't think open source models will be viable because the training costs. But once you're trained, you want as broad a set of usage for two reasons. One, You want the most revenue from it, and the lowest cost models will get the most revenue, but more importantly, there's huge data generation for training the next generation of the model. So for lots of reasons, you want to maximize usage, and if you're playing the long game, and I think the AI model game is mostly being played sort of on a five-year time horizon, not a one-year time horizon. In that time horizon, costs will drop here today. NVIDIA extracts a fairly good tax from everybody, but every, every model's gonna run on multiple types of GPUs or compute, and they will want the most data generation. So, I'm pretty convinced The next few years, the revenue isn't the important metric. You don't want to, of course, lose so much money you can't afford it, but you're not trying to make tons of money. You're trying to get tons of usage and you're trying to get tons of data from the usage and learn to be a better model. Uh, I do think there's a lot to be gained in intelligence, whether it's reasoning, whether it's probabilistic thinking, whether it's sort of pattern matching, any of these things. And these models have lots of headroom to get better. And I think…

AI assessment note: “It is costing an order of magnitude more to train”

Q But Vinod, isn't each large model successively that's coming out costing like an order of magnitude more to train?

A It is costing an order of magnitude more to train, and that's why I don't think open source models will be viable because of the training costs. But once you're trained, you want as broad a set of usage for two reasons. One, You want the most revenue from it, and the lowest cost models will get the most revenue, but more importantly, there's huge data generation for training the next generation of the model. So for lots of reasons, you want to maximize usage, and if you're playing the long game, and I think the AI model game is mostly being played sort of on a five-year time horizon, not a one-year time horizon. In that time horizon, costs will drop here today. NVIDIA extracts a fairly good tax from everybody, but every, every model is going to run on multiple types of GPUs or compute, and they will want the most data generation. So I'm pretty convinced The next few years, the revenue isn't the important metric. You don't want to, of course, lose so much money you can't afford it, but you're not trying to make tons of money. You're trying to get tons of usage and you're trying to get tons of data from the usage and learn to be a better model. Uh, I do think there's a lot to be gained in intelligence, whether it's reasoning, whether it's probabilistic thinking, whether it's sort of pattern matching, any of these things. And these models have lots of headroom to get better. And I…

AI assessment note: “It is costing an order of magnitude more to train”

Q real criticisms of what's going on. I, I wonder just your views on kind of tech policy a little bit. Like, what do you think both like how the administration's done, but what that needs to look like going forward? Cause there are these huge questions at stake and, and right now the path just seems like a heavy hand, but, but what's your view? You talk to these people.

A I don't think so. I, you know, if you look at the AI executive order, It was very bipartisan and agreed on. I spent a lot of time with the Hill Valley Forum both this year and a year ago. Almost every meeting I've been to, half have been Republicans and other half, you know, either congressmen or senators, and the other have been Democrats. So I do think policy has been reasonable and balanced. I think the Europeans, for example, have ensured they'll never participate in tech again. You know, they keep shooting themselves in the foot. So I, I do think it's reasonable. I don't buy, you know, there's always a degree of regulatory capture. And, and I do think that's somewhat true. But look, Disney was the one who extended copyrights. If you remember the late nineties copyright law, you know, beyond any reasonable measure of social good, Let's just extend it because we can. I would like regulation that says no publication can publish an article where the headline isn't written by the writer. Headlines are written for clickbait. How can it be honest to let somebody who's not dealt with the story write the headline? Every major newspaper.

AI assessment note: “So I do think policy has been reasonable and balanced.”