Please welcome Meta founder and CEO Mark<br>Zuckerberg and Microsoft chairman and<br>CEO Satia<br>[Music]<br>Nadella. All right, good to see you all<br>again.<br>Uh I hope it's been a good day. A lot of<br>exciting stuff. Really grateful to be<br>here with Zata who really does not need<br>any introduction. and he is sort of the<br>uh the legend who is behind um this the<br>the great transformation of um the<br>greatest technology company of all time.<br>Thank you Mark. Um basically leading um<br>uh you know the you know pushing towards<br>AI and cloud and all these important<br>areas and you you've always been from my<br>perspective a a kind of friend and ally<br>on the open source work that we've done.<br>So, I've really appreciated our<br>partnership on this over time and and<br>the counsel that you've given me on on<br>how we should approach building out the<br>llama ecosystem and the infrastructure<br>around that. So, thank you for being<br>here. Absolutely. It's my pleasure,<br>Mark. And I should say that you know my<br>earliest memory is meeting you um when I<br>was working on Bing in 2008 or N and so<br>on and getting u this massive lecture<br>from you on uh on something that I was<br>wrong about which is even more<br>embarrassing in retrospect. Yeah. I<br>remember this. I'll always remember<br>this. You said you got the web needs<br>people. You need to see people. I'll<br>never forget that. like the the ability<br>to actually have a profile page<br>everywhere you go. And that's a memory I<br>have. Well, I appreciate that you've<br>forgiven me for that.<br>Although the web does need people, so I<br>mean I guess at that level I was<br>correct. But maybe agents now. Yeah.<br>Well, I think maybe both. Thank god.<br>Yeah. Um but anyway, so you know, you've<br>said a number of times that this moment<br>in technology around um the the growth<br>of AI sort of reminds you of of some of<br>the important transformations in the<br>past from going to client server and the<br>beginning of the web and things like<br>that. So I'm I'm curious to Yeah. So for<br>me u you know I grew up um when sort of<br>the client was being born. And I joined<br>Microsoft just after Windows 3 and um<br>and and so I saw the birth of client<br>server then the web mobile cloud and<br>then you could say this is the fourth uh<br>or the fifth depending on how you count<br>and um it's interesting right which is<br>each time there is this transition<br>uh everything of the stack gets<br>relitigated um and you get to sort of u<br>go back to the first principles and<br>start building uh I mean I I I thought<br>like even the the shape of the cloud<br>infrastructure<br>uh for me that I built starting let's<br>say in 2007 8 to what have you the core<br>storage system for training doesn't look<br>like the core storage system you built<br>or this this workload of training right<br>the data parallel synchronous workload<br>is so different let's say Hadoop or what<br>have you so the the fact that you kind<br>of have to rethink I think everything um<br>up and down a tech stack uh with each of<br>these platform shifts is sort of what I<br>think we face from time to time. It sort<br>of grows from what was there. The web<br>was born on Windows but it went far<br>beyond that. That's kind of how I think<br>about this as well.<br>Yeah, that makes sense. I mean, you<br>you've you've made this point a bunch of<br>times around how as things get more<br>efficient,<br>um it it sort of it changes the way it<br>works and people just end up consuming a<br>lot more of the services, right? And um<br>I I guess one of the things that I'm<br>curious about because you guys are are<br>in this great enterprise business and<br>and we don't have as much visibility<br>into this is is sort of how how you're<br>seeing that play out around um around<br>all these AI models, right? You're<br>seeing like generation over generation<br>they're just getting so much more<br>efficient and delivering more<br>intelligence than the last generation.<br>And you know obviously it's all<br>happening super quickly. So I'm not sure<br>kind of how what what you can what you<br>see in that but yeah I mean if you think<br>about it right which is you know we were<br>all you know a few years ago sitting<br>around and saying oh my what's happened<br>to Moors law you know is it over what do<br>we do and here we are in some crazy sort<br>of hyperdrive Moors law and it's always<br>been also the case right which is any<br>one of these tech platform shifts has<br>not been about one scurve it's been<br>multiple scurves that compound right<br>even if could take uh just the fact that<br>the chips are getting better uh you know<br>people like Jensen or Lisa doing<br>tremendous innovation<br>u you know their cycle times have gotten<br>faster so let's say that's Mo's law but<br>on top of that everything at the fleet<br>the system software uh optimization uh<br>the model architecture optimizations uh<br>the kernel optimizations for inference<br>the app server even the prompt caching<br>how good we've gotten And uh and so you<br>add all of that up for every 6 months,<br>12 months, you have a 10x perhaps<br>improvement, right? And so when you have<br>capability improvement of that rate and<br>the prices drop at that rate,<br>fundamentally consumption goes up. So<br>I'm very optimistic<br>uh that we are at a stage where deep<br>applications can get built. Um, and so<br>these things where you have an<br>orchestrating orchestration layer with<br>these agents with multiple models, I<br>feel like we're at that place because if<br>you think about even the first<br>generation of apps, they were very<br>coupled uh very coupled to one model.<br>But we are finally getting to multimodel<br>applications where I can orchestrate in<br>fact a deterministic<br>workflow an app agent that was built on<br>one model talking to another agent. We<br>even have these protocols are helpful<br>whether it's MCP whether it's A2 to<br>whatever these are all good things if we<br>can standardize a bit and then uh we can<br>build applications that are you know<br>taking advantage of I would say uh the<br>you know these capability building but<br>has flexibility and that's where I think<br>open source absolutely has a massive<br>massive role to play. Yeah. Well, well,<br>I definitely want to make sure we can<br>get into discussing how to use multiple<br>models together. And I think that<br>there's this whole kind of concept of a<br>like distillation factory and the<br>information the infrastructure around<br>that that that you think Microsoft is<br>well positioned to basically provide as<br>there are multiple models. Um maybe<br>we'll come back to that in a minute. But<br>before we do that, you know, Microsoft<br>obviously has been on this interesting<br>journey around open source, right? This<br>is one of the big things that you did,<br>you know, under your leadership early on<br>was embracing it. And, you know, you had<br>the early partnership with OpenAI, but<br>then also were very clear that in<br>addition to working with closed models,<br>you wanted to make sure that Microsoft<br>served open models well. And I'm I'm<br>curious how you think about that and how<br>you think that the open source ecosystem<br>is going to evolve and why that's<br>important to um your customers and how<br>you think about that with all the<br>infrastructure that you're building.<br>Yeah, it's it's it's interesting you ask<br>that because I I I grew up um in fact<br>one of my formative jobs at Microsoft<br>was also making sure uh that we had<br>interoperability with the various<br>flavors of Unix out there between NT and<br>Unix. uh and that taught me one thing<br>which is interoperability is what first<br>of all customers demand and if you do a<br>good job of it um that's good for your<br>business and obviously you're meeting<br>customers where they are and so to us I<br>think that's what has shaped my thinking<br>when it comes to open I mean it's not<br>I'm not dogmatic about closed source or<br>open source both of them are needed in<br>the world uh and in fact I think<br>customers will demand them right even if<br>any one of us has dogma doesn't matter<br>because at the end of the day the world<br>will break that way which is there will<br>be a need for it. So given that I think<br>uh like for example there was SQL server<br>there was my SQL or Postgress uh there's<br>Linux there's Windows in fact there is<br>Linux on Windows in fact my favorite<br>thing to use is WSL on Windows because<br>it just makes it uh easy to take a lot<br>of the dev tools and deploy them on<br>Windows. So overall I think having an<br>posture that allows you to mix and match<br>these two things is super helpful. It<br>also fits with what you just talked<br>about because a a lot of my enterprise<br>customers want to be able to distill in<br>many cases models that they own. It's<br>their IP. Uh so that in that place where<br>an openweight model has a huge<br>structural advantage compared to a<br>closed model. Uh and so I do feel that<br>the world now is better served with<br>great closed source frontier models,<br>great open-source frontier models. And<br>to us as a hyperscaler this is a great<br>thing because after all uh our job is to<br>serve like as if you go to Azure you can<br>get fantastic Postgress you can get<br>great SQL server or you can get Linux or<br>Windows VMs same way we want to have the<br>choice available and great tooling<br>around it. Yeah. So, so what's basically<br>kind of the the pitch<br>or the role that you see Azure playing<br>um for open source but I guess across<br>all of these it doesn't need to be<br>exclusively that so for developers who<br>are getting started like what's where<br>are the areas that you're trying to<br>differentiate and be the best. So the<br>first thing is it's not like an AI<br>workload just has an AI accelerator and<br>a model that you know at inference time<br>right because the reality is in fact if<br>you look at any underneath any AI<br>workload there's storage uh right there<br>is other compute than an AI accelerator<br>and there is lot of dependency on the<br>network and so on. So the core<br>infrastructure so for us in Azure we<br>want to build compute storage network<br>plus AI accelerators at as even<br>infrastructure as a service that's<br>worldclass for someone who wants to be<br>able to build uh the next generation of<br>agents. Then above that we're also<br>building uh with foundry uh an app<br>server like for example every platform<br>shift of ours uh there's been an app<br>server how do you package up all the<br>services uh you know search or memory or<br>um uh safety all of these services that<br>are needed for someone or eval uh these<br>are all things that every developer is<br>trying to go do so if you wrap them all<br>in uh frameworks for them tools for that<br>uh that's I think the core and then the<br>other one is we we're very focused on<br>GitHub copilot uh as the tooling as<br>well. Uh we're excited about sort of the<br>progress that's making. So the<br>combination of great tools, great app<br>server and great infrastructure<br>uh for us is what I think is needed to<br>accelerate application development.<br>Yeah. So so how um you mentioned agents<br>and and increasing productivity and<br>that's obviously a huge theme for for<br>the whole ecosystem and community. I'm<br>I'm curious how are you seeing that play<br>out inside Microsoft and then also how<br>are you seeing that um what are some of<br>the most interesting examples that<br>you're seeing with develop yeah I mean I<br>think the the the thing that obviously<br>has been most helpful for us to see is<br>what's happened with software<br>development right I mean there are a<br>couple of things if you look at even the<br>evolution of GitHub copilot you started<br>with code completions then you said<br>let's add chat so that that means your<br>you don't need to go to Reddit or stack<br>overflow and you could stay in the flow.<br>So that is good. Then uh the agentic<br>workflow so you could just go assign a<br>task. Um so those three things if I look<br>at even any one of us using it you're<br>using all three at all the time right so<br>it's not like one substitutes the other<br>and now you have a proto agent even uh<br>and so you can literally go highle<br>prompt or you can just get a PR assigned<br>to a suite agent. So all four of those<br>and the productivity gains the biggest<br>lesson learned there Mark is you got to<br>integrate all of that with your current<br>repo and your current developer workflow<br>right you I mean it's one thing to build<br>a new green field app but none of us get<br>to work on complete green field all the<br>time right so uh you are working in a<br>large code base with a large sort of co<br>you know set of workflows so you got to<br>integrate the tool chain and that's the<br>systems work uh that I think any<br>engineering team has to do and that's<br>when you see the productivity. The same<br>applies quite frankly for the rest of<br>knowledge work as well. So when in our<br>case with copilot deployment uh for<br>knowledge work um you know if you take<br>sales like you know one of the things I<br>always describe is the uh let's say I'm<br>getting ready for a customer meeting the<br>workflow of how I get ready for an<br>enterprise customer meeting has not<br>changed since 1992 when I joined<br>Microsoft right basically someone will<br>write a report that'll come in email or<br>will be shared in a document and I'll<br>read it before the night before now I<br>just go to researcher and copilot and I<br>get the thing which is a combination of<br>what's on the web, what's internal and<br>even in my CRM all done in real time,<br>right? But that's a change in there's no<br>need for somebody to prepare anything<br>just because it's all available on tap.<br>So it requires you to change work<br>artifact and workflow and that's where<br>that's a lot of change. Um and it<br>happens slowly at first and then all of<br>a sudden right I saw that with the PCs,<br>right? I I you think about how the world<br>did forecasting before email and Excel<br>like inter off you know the faxes went<br>around. Yeah. I guess you never lived<br>that world. I I was in middle school.<br>Yeah. I know it's but there was a world<br>where people sent around faxes and<br>people went and did uh inter office<br>memos and then somebody said hey I'll<br>send a spreadsheet in an email and<br>people enter numbers and that changed<br>how people did forecasting and that's I<br>think what we are at the very beginning<br>of and you see it in customer service<br>you see it I think in marketing<br>collateral creation content creation so<br>that's where we are and you're seeing<br>tangible progress and tangible<br>productivity gains yeah interesting um<br>in terms of the coding and and how it<br>improves that do you have a sense of how<br>much of the code like what percent of<br>the code that's being written inside<br>Microsoft at this point is written by AI<br>as opposed to by by the engineers yeah<br>so there's two sort of things we're<br>tracking one is the accept rates itself<br>right that's sort of whatever 30 40 it's<br>going up monotonically uh and it depends<br>like one of the big challenges we had<br>for a long time is we are a lot of our<br>code is still C++<br>um C and C# is pretty good but C++ were<br>it was not that great. Python is<br>fantastic. So we've now gotten get<br>better at that. So as language support<br>has increased the code completions have<br>gotten good. The place where the agentic<br>code still it's very it's sort of nent<br>for new green field it's very very high.<br>uh but as I said it's nothing is green<br>real uh in many cases and so therefore I<br>would say maybe at this point the PR oh<br>by the way code reviews are very high so<br>in fact the agents we have for reviewing<br>code uh that that usage is increased and<br>so I'd say maybe 20 30% of the code that<br>is inside of our repos today in some of<br>our projects are probably all uh written<br>by software what about you guys um I<br>actually don't have the number off the<br>top of my head but I And it's um I I I<br>think you know we<br>I think a lot of the stats that people<br>say are still effectively of this like<br>autocomplete variety. Yeah. But we have<br>a bunch of teams that are working on um<br>on basically doing feed ranking<br>experiments and ads ranking and like<br>very contained domains where you can<br>study the history of all the changes<br>that have been made and like and and and<br>make a change. And that I think is like<br>is kind of an interesting area for for<br>us to work in. But the big one that<br>we're focused on is um building an AI<br>and a machine learning engineer to<br>advance the llama development itself.<br>Right? Because I mean our our bet is<br>sort of that in the next<br>year probably you know I don't know<br>maybe half the development is going to<br>be done by AI as as opposed to people<br>and then that will just kind of increase<br>from there. So I was just curious if you<br>were if you were seeing something<br>different. Yeah. I mean to<br>me the the SW agent is the sort of the<br>first attempt. So the question for us is<br>in the next year can we get um like<br>let's take a kernel optimization right<br>will we get to sort of something like<br>that that happens I think it's more<br>likely whether it comes up with a novel<br>model architecture change probably not.<br>So the question is which task? Yeah.<br>Yeah. know optimizations, security<br>improvements, that type of stuff. I<br>think seems like it's it's pretty high<br>opportunity.<br>Um yeah, no I we're also trying to solve<br>a different problem on it because I mean<br>you guys serve like a lot of developers<br>and engineers that's like your core<br>business. Um whereas for us we're<br>thinking about this more as a thing to<br>improve our internal development and<br>then improve the llama models which<br>other people can use but it's not<br>something that we do the endto-end<br>workflow on in the way that you do. So<br>it's always just interesting to hear how<br>you're thinking about that. Yeah. And<br>the other thing for us is yeah to your<br>point our core business in fact you know<br>Bill started the company as a tools<br>company. And so to us uh the interesting<br>thing I'll think about now is maybe the<br>way we we should reconceptualize our<br>tools is and infrastructure quite<br>frankly are the tools and the<br>infrastructure for the agents to use<br>because even the three agent needs a<br>bunch of tools and what shape should<br>they be uh what should their<br>infrastructure what should their<br>sandboxes be so a lot of what we're<br>going to do uh is essentially evolve<br>even what does the GitHub repo construct<br>even look for the sui agent Yeah. No,<br>it's that that's a it's a very<br>interesting concept and I mean I I tend<br>to think that like like every engineer<br>is effectively going to end up being<br>more of like a tech lead in the future<br>that has sort of their own little army<br>of of of engineering agents that they<br>work with. But yeah. Um so with that<br>with that all in mind, I mean I guess<br>I'm um I mean there are a few directions<br>to go in. I mean I'm curious like how I<br>mean you mentioned your personal<br>workflow for for using AI. I'm curious<br>how that's changed. Um, I'm I'm I'm also<br>kind of curious because you were talking<br>about how uh how Microsoft got started<br>on this and the legacy there, but I mean<br>I guess there's always this question if<br>you were getting started as a developer<br>today building something, how would you<br>think about which tools you'd be using?<br>and and um yeah I think that one of it<br>the one of the biggest sort of I'll call<br>it I don't know dreams pursuits<br>questions that Bill sort of inculcated<br>in all of us was what's the difference<br>between um uh he used to sort of talk<br>about it more like what's the difference<br>between a document and an application<br>and a<br>website right now if you use meta<br>chatgpt copilot what have you. It's<br>unclear to me what's the difference<br>between a chat session Mhm. Uh and and<br>then I go to pages in our case like<br>literally even coming down you know I<br>was like reading up everything about<br>Llama 4 all the models like literally I<br>was just doing a bunch of chat sessions<br>adding it to a document effectively in<br>pages persisting it uh and then you can<br>go give it I mean since you have code<br>completion you can go you know make it<br>an app or what have you and so this idea<br>that you can start with a highlevel<br>intent and end up with what is an<br>artifact that is a living artifact that<br>you would have called in the past an<br>application is going to have profound<br>implications I think on workflows and I<br>think we are at the beginning of that uh<br>and that's what my dream is so if I sort<br>of say as a builder of infrastructure<br>and tools and as a user of it these<br>artificial category boundaries not<br>artificial or these category boundaries<br>that were created uh mostly because of<br>limitations of how our software worked<br>perhaps Perhaps you transcend um in fact<br>the other thing we used to always think<br>about is why word excel powerpoint<br>different why isn't it one thing and<br>we've tried multiple attempts of it uh<br>but now you can conceive of it right<br>which is you can start in word and you<br>can sort of visualize things like excel<br>and present it and they can all be<br>persisted as one data structure or what<br>have you and so to me that<br>malleability that was not as robust<br>before is now there. Yeah. Interesting.<br>Makes sense. So, one of the things in<br>our conversations over the years that<br>that has kind of stuck with me. I I feel<br>like you have a<br>very kind of I it's a very like<br>reasonable way of looking at the the the<br>the way that technology trends unfold<br>and there's all this hype around AI and<br>I I feel like you've been able to kind<br>of see through that and make very<br>rational investments at each step along<br>the way. And one of the points that<br>you've made is that okay there's all<br>this hype but like really at the end of<br>the day if this is going to lead to<br>massive increases in productivity that<br>needs to be reflected in major increases<br>in GDP and that this is going to take<br>like some like multiple years many years<br>to to kind of play out and I'm curious<br>how you think what's your current<br>outlook on on sort of what we should be<br>looking for to understand the progress<br>that this is making and and how we would<br>and and kind of like where you expect<br>that to be over like a three five, seven<br>year period. To me, I think that that's<br>right. I mean, because to us, I would<br>say it's a pretty existential priority.<br>Quite frankly, the world needs sort of a<br>new factor of production and input that<br>allows us to deal with a lot of the<br>challenges uh we have. Um and uh and the<br>best way to think about it is hey, what<br>would it take let's say for the<br>developed world to grow at 10%.<br>um right which may have been some of the<br>peak numbers during uh let's say the<br>industrial revolution or what have you.<br>Um and for that to happen then you have<br>to sort of have productivity gains in<br>every function right in healthcare in uh<br>in retail uh in broad knowledge work in<br>any industry and for that to happen that<br>I think AI has the promise but you now<br>have to sort of really have it deliver<br>the real change in productivity and that<br>requires software and also management<br>change right because in some sense<br>people have to work with it differently.<br>uh you know people always quote what<br>happened with electricity right it was<br>there for 50 years before people figured<br>out that hey we got to really change the<br>factories to really use electricity<br>differently right and that is the the<br>famous Ford case study and so to me<br>we're in somewhere in between I hope we<br>don't take 50 years um but I do feel<br>that by just thinking of this as<br>whatever the horseless carriage uh is<br>also not going to be uh the way we're<br>going to get to the other side So it's<br>not just tech. Tech has got a progress.<br>Uh you got to put that into systems that<br>actually deliver the new work work<br>artifact and workflow. Yeah. Well, we're<br>all investing as if it's not going to<br>take 50 years. So I hope it doesn't take<br>50 years. Um but all right. So, so we've<br>kind of been, you know, we're doing more<br>of the technical questions up front and<br>then getting to the big picture stuff,<br>but I realized that I we forgot to dive<br>into the distillation factory thing and<br>how you basically combine all the<br>different um all of the the different AI<br>models that are getting built out for<br>for open source and like what the<br>infrastructure is that you think is<br>going to be necessary to build that out.<br>And I mean this is something that you've<br>talked about and I think have like a<br>Yeah. To me I think that that that to me<br>is I think one of the biggest roles of<br>open source right which is to be able to<br>take let's say um your some of your even<br>even inside of the llama family taking a<br>large model and then to be able uh to<br>distill it into a smaller model that has<br>even that same model shape uh I think is<br>a big use case. Um and so to be able to<br>then build the tooling for it as a<br>service um and make the barrier to that<br>like I mean to your point standing up<br>some of these large models as a bunch of<br>infrastructure not everybody needs to do<br>it but if you do it as a cloud and then<br>pull the tools around it and the<br>outcomes a distilled model in our case<br>let's just say there was let's say for<br>every tenant of Microsoft 365 if they<br>could have a distilled tasksp specific<br>model that they can create as an agent<br>or a workflow that then can be invoked<br>from within copilot. That to me is a<br>breakthrough scenario and people are<br>already doing a lot of that and we want<br>to make that a lot easier. And so when I<br>say distillation factory uh that's the<br>the many to many type of or one to many<br>relationship I want between one large<br>model uh these distilled models that<br>then get composed uh with a lot of other<br>workflows inside of something like a<br>product like GitHub copilot or copilot<br>because now they all support with MCP<br>servers and so on the invocation of<br>these other agents. Yeah. No, it's it's<br>it's I've always been fascinated by<br>this. I mean, I think the distillation<br>is one of the most powerful parts of<br>open source. And I I think just because<br>of our kind of respective parts of what<br>we do here in terms of, you know, we're<br>training the initial llama models, but<br>we don't build out most of the developer<br>infrastructure ourselves. I think having<br>companies like yours that that basically<br>are going to build out this complex<br>infrastructure. We have models like um<br>like the behemoth one that we're working<br>on, which I I I think it's really<br>unclear how you would use it except to<br>distill it into more reasonable forms.<br>Yeah. So I mean even for us to use it<br>ourselves we had to build a bunch of<br>infrastructure internally even just to<br>be able to like post-train it and you<br>know there's no way we're going to run<br>infra Maverick is you said is distilled<br>all yeah a bunch of Maverick I mean the<br>way that we basically got the<br>performance on Maverick to be at the<br>level it is as it is right it's<br>basically it's multimodal right so it's<br>leading multimodal on the text<br>performance it's basically you know up<br>there with the other leading um text<br>models but it's smaller right so I mean<br>deepseeek is a bigger model than it But<br>on text it's it's basically comparable<br>and then on images and and kind of<br>multimodal it exists and the others you<br>know don't. So um so that is yeah I mean<br>a lot of how we basically got that is<br>you know we we have the pre-train of the<br>behemoth is is done and we're we're<br>working on the post training but even<br>just kind of getting it I mean the<br>distillation it's just like it's magic.<br>you basically can make it so that you<br>can get 90% or 95% of the intelligence<br>of something that is 20 times larger in<br>a in a form factor that is so much<br>cheaper and more efficient to use. So<br>then the question is okay how do you<br>make it so that that's available to<br>people who are not as um you know who<br>aren't able to build up their own<br>infrastructure aren't as technically<br>sophisticated to go do that because<br>right now I think that there's a<br>relatively small number of labs in the<br>world that could do that kind of um<br>either distillation or even operate<br>models of that scale and I think by the<br>time that that like this vision that you<br>have is built out and like it's it's<br>accessible for most developers around<br>the world to be able to not only distill<br>from a single model but hopefully over<br>time be able to mix and match and take<br>different aspects of intelligence from<br>different models where they're stronger.<br>Um that just seems like one of the<br>coolest things that I think is going to<br>get built. Yeah. No, I think that that<br>that's correct. And so therefore a<br>little bit about like what's the um if<br>you have multiple models you're<br>distilling from and then what's the eval<br>around this distilled model that you can<br>then qualify. I think that's where uh we<br>can do a lot of our tooling work, our<br>infrastructure work, reduce the barriers<br>for people to be able to have that<br>flexibility. Uh and the good news here<br>is um it's sort of already started.<br>People have ex there's existence proof<br>of all this. It's just a question of can<br>you reduce the barrier to building all<br>of it. And the other thing is the speed<br>with which people can move. One of the<br>challenges to date has been uh I do<br>something with one model, I fine-tuned<br>it, a new sample drops, I need to move<br>fast to the new sample. So that's the<br>other thing that we have got to get good<br>at uh because you can't be saddled with<br>what you did. Uh because the world is<br>moving too fast. Yeah. Yeah. And also I<br>mean developers just need things in<br>different shapes. I mean the llama for<br>shape of 17 billion parameters per<br>expert was designed because the atomic<br>unit that we have at meta is an H100<br>right so we want to be able to run like<br>these things very efficiently on that um<br>it's it's one of the things where I mean<br>you look at some of the other models<br>that have come out some of the other<br>open source models it's like they're<br>good intelligence but sort of awkward to<br>inference because of the scale and maybe<br>they're targeting different kinds of<br>infrastructure um but that's but we're<br>basically we're we built this for kind<br>of server production, but a lot of the<br>open source community wants even smaller<br>models. So to be able to, you know,<br>build, you know, the most popular Llama<br>3 model was the 8B and we we'll we'll<br>work on a smaller version. I talked<br>about that earlier. The internal we<br>refer to as little llama, but we'll see<br>what what we actually ship um at as but<br>um but being able to basically take<br>whatever intelligence you have from<br>bigger models and distill them into<br>whatever form factor you want to be able<br>to run on your laptop, on your phone, on<br>whatever whatever the the thing is, I<br>think is just I I don't know. I mean, to<br>me, this is like one of the most<br>important things. Yeah. And also I think<br>you guys are obviously working on this<br>and um and I think that this is good to<br>see which is um if we can get to these<br>hybrid models um you know whether it's<br>you know dense plus thinking models that<br>combined and then you're able to get the<br>latency uh that you want or the thinking<br>time you want uh and it's flexible uh<br>that I think is where we will all want<br>to end up. Yeah. All right. Well, maybe<br>a good note to end on is just, you know,<br>when you look at everything that's going<br>on, I mean, I'm curious what is what<br>gives you like the most optimism or what<br>are you most excited about over the next<br>period for what developers are going to<br>be doing over the next few years? Yeah,<br>I mean, look, I mean, you know, I'm I<br>always take my inspiration from that<br>whatever that Bob Dylan line, right,<br>which is either you're busy being born<br>or you're busy dying. It's get better to<br>be busy being born. And especially at a<br>time like this um I think what gives me<br>optimism is even with all the various<br>constraints it turns out software in<br>this new form of AI is still the most<br>malleable resource we have to use to go<br>solve these hard problems. Um, and so<br>that's what gives me the optimism and<br>and I also sort of say that's the call<br>to action I think for the people in the<br>room and for all of us is to be able to<br>sort of take the opportunity to lean<br>into this uh but then also build<br>solutions. the when I look at the<br>whether it's an IT backlog in a company<br>or the unsolved problems in the real<br>world both of them need something new uh<br>and in order to work that and that's<br>where I think the greatest uh benefit of<br>all of this tech is and it'll only come<br>down to you know developers in<br>particular uh being able to go at it<br>fearlessly. Awesome. All right. Well,<br>thank you Satia and thank you all for<br>coming out. This has been an exciting<br>day and I'm very excited about what we<br>are all building. So,<br>[Music]