Well, moving on. Microsoft unveiled six new AI models today designed for tasks like coding and cyber security, with the pitch of lower costs. But it's also building a general purpose model to challenge OpenAI and Anthropic, just as OpenAI releases two new models of its own. Microsoft AI CEO Mustafa Suleyman is here. Mustafa, welcome.
Great to be with you, Dara. How's it going?
Great. I'm so glad that you could make it here in person. How has Ignite gone? And I mean, maybe explain the rationale behind working towards a general purpose model, because for people who may not be as familiar, right, Microsoft kind of started by taking the best models, especially from OpenAI, and serving them on your platform. So why now go and build your own?
So what we hear from developers every day, and particularly CIOs and CFOs, is that the token cost is the thing that is going through the roof at the moment. And so everybody is trying to get the best performance you can get with using, as you know, as little as possible in terms of token spend. And so we've designed a family of models for this particular launch that basically gives you 98% of the performance of GPT-6.1, so like pretty much the top model other than the one that got launched today, at 40% of the cost. And that's exactly what developers want every day. It does routine tasks incredibly well. It's incredibly strong on all the benchmarks. We're deploying it across GitHub Copilot. It'll be available in Foundry soon, soon to come to Copilot. And so that's really the focus that we have at this point. But you're right, we're also building our own general purpose, you know, full scale frontier models, and they're also coming along really, really well. We announced just a couple weeks ago our pre-trained models are now state-of-the-art.
Are they meant to compete with the very best models from Anthropic?
They will. Absolutely. I mean, certainly in the new year we're hoping that they'll get to the absolute frontier. Today we've been focused really mostly on the coding models and also on cyber models. Our cyber models are incredible. They basically give you the same performance as GPT-6 Astra at 2% of the cost. So the goal is basically to have these cyber security agents run continuously in the background, monitoring your systems, identifying vulnerabilities, patching them in real time.
And that's what Microsoft and you have really been focused on, are these specialized models. But I want to go back again to this sort of general purpose big frontier model that you want to reach. How big of a factor is compute in that, right, when everyone is looking at sort of compute shortages and bottlenecks, even in getting data centers up and running? Microsoft obviously has its own data centers. Is that an advantage that Microsoft has? How much compute is it going to take to get to the frontier?
We have frontier scale compute today for training and of course for our inference. So compute isn't a bottleneck for us at the moment. We just started our super intelligence team 12 months ago. We renegotiated the contract with OpenAI and that freed us up to go and pursue the absolute frontier. And obviously along the way we want to make sure that we deliver great impact for our products, and that's what we've been doing.
Can you withhold compute from your competitors as an edge?
That's not how we think about it. I mean, OpenAI is an important partner for us. We have a huge history with them. They've been incredible. We're still huge fans. And to be clear, we also have Anthropic models in the mix. We have Mistral models. Every model that you can think of is available in Foundry.
And certainly you benefit in a way because of Microsoft's stake in OpenAI when they can, you know, push the frontier as well and bring in more revenue. But I wonder, so is this an era where Microsoft is going to use more compute for itself and less available for not just competitors but any of your other customers?
I mean, the way to think about it is that compute is the most important thing happening right now. We have gigawatts of compute across the Microsoft fleet, so we don't necessarily feel compute constrained at this moment. Who knows how things might change next year. I mean, we are expecting to see a massive uptick in demand. I mean, things are already going through the roof. We have to revise our demand plans every quarter at this point. So I think that you're just seeing the tip of the iceberg in terms of adoption. People really are starting to understand how these long-running agents sit in the background, monitoring, supporting, synthesizing, basically doing autonomous work for you behind the scenes. And so I think we expect that trajectory to really, you know, inflect up next year.
Especially when you bring agents into the mix, right? I know that Microsoft has its own with Copilot. How quickly is that going to climb?
Yeah, I mean, very fast, I think. I mean, agents really are the kind of new, like, tool, the new workhorse in the background for everybody's enterprises. We've heard it over and over again from, you know, our partners is that they actually want these long-running agents that can constantly support them. That's why we put so much effort into making sure our cyber models can operate at 2% of the cost. I mean, if you can do it, you know, 50, 60, 100x cheaper, as some of our versions do, I mean, it means that you can then completely change the type of work that you ask your model to do.
Now, when I think about more efficient and nearly as capable models, I think about Chinese open-weight models. Is this, do you think that they will continue to gain as much adoption, or do you think that that's a market that's ripe for an American company like Microsoft that's able to do something similar, efficient, nearly as capable model?
I mean, we think it's good to have an open ecosystem. Like, we want choice. We want lots of options, and we support that across Microsoft. So our customers know that the promise that we make them is that not only will we give them MAI models at the best cost, we'll also give them all the options if they choose a different option. You know, at the same time, I think these long-running agents really are going to be the future. And so that means that a company needs to really own and control the agents they build. Like, no one wants to pour all of their IQ, the knowledge, the data, the workflows into somebody else's model. So we have a new program called MAI licensed weights where we basically give you an MAI frontier model that is pretty much, you know, 98% of the absolute frontier, and you can pour all your own data into that model. You get to control it in your own Azure environment. You have complete control over who gets access to it, how you use it. And I think that's going to be a really, really big part of this next wave.
Controlled within Azure, though, so it has to stay within Azure.
Yeah, not any open-weight models. We have no plan to do open-weight. No. I mean, we think that the frontier is going so fast. These things are so powerful that we have to have, I mean, we're seeing it in the last few days, last few weeks with all the cyber attacks and so on. I mean, we have to have guardrails around these things.
Now, it's funny because you said this, your latest models are 96% as capable as the OpenAI models before the ones that were released today. Today, it's moving so fast, right? And the ones that we're getting today, I've heard from people that they're not even as capable as the ones we haven't seen, because this year, you mentioned cyber security, they're holding them for alignment and so on. So how do you, how do you even know what you need to catch up to?
I think that the challenge is most people don't even quite realize how to use frontier intelligence in their day-to-day work, right? So, I mean, you do see like people jump to the latest model and then they probably like switch back to the prior generation. As you start to figure out what the difference is between the absolute very best on a day and then the one below and the one below, and I think so just selecting between these small and big models is actually the craft here, and we're just trying to put new points on that frontier curve to give people choice.
What do you get if you're at the frontier? Is it still this pursuit of AGI? How does RSI, recursive self-improvement, AI that does AI, how does that fit into this? Is that still the prize?
Look, I think that it's pretty obvious that we're starting to see models improve models. You know, in our own development, we use models across the fleet all the time to, you know, check for bugs, to produce training data, to create evaluations, but increasingly they're fully involved in the process of improving their own models. And so this is definitely a dangerous capability. I mean, I'm really, really nervous about it. I feel like the last three months have been a total watershed moment. And I think that we do need to be extremely careful, and I don't think that folks in the industry are being alarmist. I think this is very serious, and so I think that requires trust and also it requires a bit of caution, you know, proceeding step by step when we can first prove that they can be safe before we put them into production. I think that's a very rational strategy given how powerful these things are.
How does that look in practice if Microsoft is now trying to reach the frontier? Is there something you would hold back if you thought that it could become dangerous?
I mean, for sure, the obligation is on us as a provider of models and a developer of our own models to make sure we have the guardrails, they can be contained. We've aligned their behaviors. We've addressed all the security flaws, and that might mean that we have to go slower to go faster.