Go play it. I think it's also a super fun game to go play, but it's probably hard for reasons that are different than your clever take on how to implement a particular algorithm or a piece of the technology stack. It's just got a bunch of complexity that is in many ways divorced from the technology itself. So if you want to go do that, you got to get convinced that that's the game that you want to be playing and where you can add some valuable contribution. I think everywhere else, people should be infinitely pragmatic about how they're going to solve an interesting customer problem. I think we were talking about this a little bit earlier. There's this gigantic capability overhang that I think we have right now with these AI systems where they already, forget about what's going to happen in a year, they're already more powerful than what people are using them for. A lot of the reasons that people are waiting around and not solving problems is that some of the things that you need to do to squeeze the capability out of these systems is just ugly looking plumbing stuff.
Or grunty product building. But tell you what, I'm sure this is true for you as well.
You're in a startup. That's kind of your life. It's more about the grind. 99% of it is just a grind.
It's not clever. It's grind. You can't be dumb, but it really is, any way possible, I'm going to drive a bulldozer through solving this problem.
It's interesting, you know, how one of the particular areas you and I mentioned earlier, you thinking about is how to reconcile the fact that these models are so general purpose. And we've obviously got an ever-expanding context window so you can tailor them to what you want them to do for your particular use case. But ultimately there's still a huge opportunity in terms of long-term memory retrieval. You can call it maybe fine-tuning. But a lot of it is how do you successfully apply these models to big separate context that could be your organization, it could be your function.
Yeah. Where do you think the opportunity there is?
Yeah, I think there's a huge amount of opportunity there. I've been talking about this for a couple of years now. If you just scope down to agents, and you think that the purpose of agents is for humans to be able to delegate increasingly complicated tasks for the agents to go complete autonomously, as much of that complicated task as humanly possible, it's inconceivable that you could do a similar sort of delegation to a human that didn't have a functioning memory.
You have to be able to not necessarily have the entirety of that problem scope in your active memory at any one time, but you have to be able to do retrieval. A biological brain is really good even at imprecise retrieval with mechanisms to drive precision after you've gotten in the ballpark of what you need to retrieve. Agents are going to need to be able to scratch pad their work, remember previous interactions they've had with you, and way more effectively than things like RAG, pull things that they need to solve a problem into context. I think there's a bunch of infrastructure that you need to build there, and even a bunch of application-specific things that you're going to have to build. If you think about biological memory, a bunch of us go to university and we get a bunch of techniques loaded into our brains that help us manage discipline-specific memory. I think you're going to need that in agents. It's going to be task or product specific, and it's not going to just drop out of training a little bit bigger model. Someone's going to have to go do some real work to plumb all of that stuff all the way through.
Yeah, I thought that, I don't know how many of you listened to Sasha on Dwarkesh, I think a week and a half ago, but I thought his articulation of this was super interesting. He said that right now the focus seems to be, can we get these foundation models to out of the box be the world's best knowledge worker for everything, whereas in reality what you really want is a really smart knowledge worker who can go learn about the stuff you want them to do in your company.
Because it doesn't make sense that out of the box somebody would be right at 29 things good at everything, and that's kind of the approach we're taking right now with our whole pre-training, post-training. But really the point of going to university is that you have a bunch of really solid foundations and capabilities that you can then apply in a variety of different contexts. So it's a second thing that we need but we just built half today, I think.
Yeah, I completely agree with that. Amazing. Okay. Maybe to ask the question explicitly, I've already answered this, but I'm still curious. How, with foundation models, we obviously have a number of the large labs and some of the bigger companies converging on what seem to be frontier models that have the sufficient level of feedback loops, flywheels, ability to fund them, and they're also giving access to it to everyone in the world, which is kind of incredible. This has never really happened in technology. Most of the time you build something big and expensive to build and valuable, and then you give it to a small number of people first, the military, and then it makes its way down to companies and consumers. But we have this really awesome thing that is being given to everyone kind of immediately, and that's great for consumers. There's a ton of consumer surplus. But then there's also open-source models now.
Yep, which are actually by some measures probably more than 50 to 60% of calls within applications according to some of the numbers. How is that open source versus closed source in your opinion going to play out over the coming years?
We will see. I'm kind of excited to see progress on both dimensions. The thing that I can't really square with the open source stuff is if we still are in the domain where we get serious returns on model capability from scale, I don't know what mechanism for open source model training is better than a commercial model to drive scale. So it's a little bit different from getting Linux built open source. With Linux, you can have a whole bunch of clever people who want to spend nights and weekends making contributions to the code base, and the whole ecosystem gets better as a function of all of those contributions. But when you download and use the open source model, you're not quite doing the same thing. You certainly aren't initiating a new round of pre-training or post-training on the model that goes back into the ecosystem that makes the weights part of the model better. There are some contributions happening in these open source models around the ecosystem and a bunch of performance optimization work that's really super valuable. But again, I think that scaling thing is the thing that you have to think about. The category error here is thinking that it's got to be either or. I just don't see that at all. We at Microsoft use both. We have them available in our cloud and our own products. We're using a mixture of things. Going back to my earlier point, I like worlds where we're fixated on products and product outcomes, and you just sort of leave it to the engineers to sort out what the infrastructure is going to look like. As long as the product's amazing, getting better, cheaper, faster, and higher quality on all the dimensions that people care about, what do you care what the infrastructure is like? Whether it's a federation of open source models, or your L1 cache is an open source thing running locally and your L2 is a prompt that goes to a big expensive cloud model, whatever solves the problem. Again, great news for us as builders that you have this variety of different options and you're not actually locked in. It's an insane time to build from that dimension.
Yeah, and I look, I love the open source models. I was, I still am, I have infinite amounts of curiosity. Being able to tinker around with stuff and see how it works, I think they're nothing but positive.
Yeah, absolutely. Maybe can we touch a little bit about data? This pertains to open source in particular as well, which is understanding the efficiency curves for data quality versus quantity. Especially going forward, I think you've emphasized that quality is becoming more and more important, and we're seeing this with all the companies striving to create specialized environments for producing this data. But play this out over the next few years. What's your best prediction in terms of where both the big players will continue to get their data, and where will the smaller players be getting proprietary data or unique ways of getting data to train their systems?
Yeah, I think it's pretty clear if you are on or near the frontier that for pre-training at least, we're well past the point where there are enough tokens organically occurring in the world to train bigger and bigger models. It has been the case for quite a while now that the super high capability frontier models that are trained with a ton of compute have a lot of synthetic tokens that they're doing training on. But I also think it's clear to everyone that having experts providing feedback in the post-training part of the process is really critically important. This is a thing I would encourage you all to look at. To the extent that you are doing custom post trains on models yourself, you have to think about where advantage can still exist. It can exist in you understanding your domain and being able to identify experts that you can then use as part of your post-training process to make your particular post-train for your application better. I think that's certainly a thing that we're going to continue to see in a whole bunch of ways. A bunch of this stuff, if you can get to scale, can happen with clever UI design inside of agents. I was on Gemini the other day. Yes, I do use my former employer's product. It's a good model. There's more good stuff coming too, which is the really exciting thing. No one should think that the next announcement any one of us makes is the last great announcement. Stuff's just going to continue to grind and get better. I was using Gemini and I asked it a technical question, and it popped up two completions for the prompt and asked me which one I thought was better. You will have increasingly inventive UI treatments.
That was cool. Yeah, the generative UI stuff as part of Gemini was pretty cool.
Yeah. Yeah. Yeah. But the other thing that's really important when you talk about data, and this may be the most important dimension of data, is you have the data that you're using in your pre and post training part of your infrastructure building, but you also have all the stuff we were talking about a minute ago: how do you hydrate memory? How do you make sure that you've plumbed through all the data sources that you want your models or agents or AI systems to have access to in order to actually solve problems? I think that maybe is the more important data, and it's the place where you're going to have the messier set of constraints. People who have proprietary data right now are on the one hand going to want to be able to use AI systems to get more value out of the proprietary data, but they're going to want to do it in a way that is not leaking a whole bunch of value out of their data.
It's like value preserving for them. Yeah. Exactly.
Yeah. And so that's an interesting set of problems to go solve.
