0:00In a world where code generation has become cheap, what has become hard is verification. In fact, I kind of feel like I got to learn because I've not worked in a team that is mostly now code as output. But the code as output has a particular order of operations. It's not like code as output means everybody's just generating all sorts of random code.
0:18Out of nowhere 2025 clock code. For the first time, I felt that you didn't have an immediate response.
0:24When model jumps happen nonlinearly, the form factor has to change. AI first product building in that sense is unforgiving. In 6 months or so, ide kind of faded away.
0:35I'm asking myself, are there parallels to the open source movement and now the open model?
0:40If you just get past this thing called there's going to be one frontier model to have all of the world's knowledge, right? If you don't buy that, then you have to assume that there should be in some sense as many frontier models as firms in the world, right? Because after all,
0:59What does software engineering mean after code generation becomes very cheap? Today's guest is Sacha Nadella, Microsoft CEO, one of the few big tech CEOs [music] who started his career as a software engineer. Today we cover what the term software factory means for Satcha and how it looks like today inside of Microsoft. Open models, [music] local models, and the idea of unmetered intelligence. the very real anxiety for us devs around AI, Microsoft voluntary exit offers and why the junior hiring slowdown might only be temporary and many more. If you want to understand
1:28what Satcha thinks will take to become a standout software dev in this age of AI and want to understand his thoughtful mental models on how to make sense of all this change, this episode is for you. This episode is presented by Turbopuffer, a ridiculously scalable, fast, and cheap hybrid search engine built on top of object storage used by cuttingedge companies like Entropic, Linear Ramp, and many others. Many companies who switch to Turbopuffer are cutting their search bills by 50% or more and getting much faster results.
1:54Check it out at turbopuffer.com/pragmatic. Satcha, welcome to the podcast.
1:59Thank you so much for having me, G.
2:01Well, it's great to be here because I'm not sure if you know, but I was a Microsofty.
2:05Were you here or where were you? I I was in London and this is the first time I'm here in headquarters. So, you know, like welcome 12 years back. But I want to rewind back there. I was actually at Microsoft when you became CEO in 2014.
2:18And I remember distinctly remember how Microsoft was still very much embracing closed source. And when you came on, you started to embrace open source. There was Visual Studio Code that was released. Microsoft acquired GitHub and went all on Git before we used TFS. Typescript was released as open source.
2:36And I just wanted to rewind back. Why did you people were surprised my my friends outside of Microsoft were surprised even I was surprised. Why did Microsoft go all in on open source?
2:47Yeah. Yeah, I mean at some level sometimes these narratives and perceptions get shaped and it takes a lot to change and what have you because if you look at even I would say commits even into Linux even preme becoming CEO I think Microsoft had significant uh contributions and were committed to it.
3:07But the thing that I think we really made sure the we understood internally and the world understood is Microsoft is committed to open source, right? Whether it is uh bringing some of our workloads to Linux like SQL Server uh or contributing to Linux or or open sourcing projects like Typescript and others. I think we're all part and you know obviously our you know us as stewards of GitHub uh put us at the center of it and we just wanted to make
3:36sure that we were able to earn the trust uh of the open source community by doing right by the community. In fact, it's fascinating. I've always thought about this which is when you look at companies that have used open-source and people who have contributed to open source if there was a true accounting of it at some point I think Microsoft would be at the top or if not amongst the top right in fact if you look at some quote unquote a lot of these other companies that have benefited the most from open
4:05source have hardly contributed it's never and that is I think how you know and like hey you can't complain but that's how the world rolls which is it's on nar narratives versus facts all the time.
4:15Well, it's it's narratives versus facts.
4:17That's true. And and again, I think there's that Windows legacy from the '90s that shifts people's uh you know, like how they think about it. But what I remember made the biggest impact across the developer community was not really open source. It was Visual Studio Code.
4:31Yeah, it was open source, sure, but it was also free. And I was a developer before I used Visual Studio and it was so good. Visual Studio 2003 is still amazing piece of software. and somehow you decided Microsoft to bring Visual Studio for free in an open source version to the broader dev community and boom 10 or 5 years later it's everywhere everyone was using it was that intentional yeah in fact u you know that you know Eric GMA and others who were at
5:00that time in that place really was absolutely it was intentional even building with community in fact uh you know when I look at even product making uh it was the first time I also got exposed to what I think now people talk about you know building in the open.
5:13That was really the first time we all learned what building in the open means, having the community around it and just that feedback cycle. I remember even spending time at that time with Amanda Silver and others learning how that team even the product manager what does it mean to be a PM uh on that team how they every day looked at the community feedback as a way to build versus you know top you know like it's not like a you know I listen to my vice president I listen to my community. So there were
5:41many many cultural aspects of building in the open uh beyond open source that I think were fundamentally transformative. In fact, I would say sometimes I sort of feel like we learned a lot of great lessons there. We should have learned it even more deeply and more broadly. Speaking of of learning more more deeply or more broadly, Visual Studio Code was everywhere.
6:00GitHub was doing great. And then with Copilot, I mean I was so surprised. I've been looking back how Microsoft was so early to realizing that AI before chat GPC 3.5 this was I think the 3.0 Over version or or so Copilot was working like magic and there was no real competition for at least a year until 2023 when cursor came out and then cursor came out and I still felt that you guys were super smart about VS code because cursor was based on VS code familiar interface and you know GitHub
6:29copilot and cursor were going has side by side competing with one another keeping up with one another and then out of nowhere 2025 clock code and for myself for people working at startups uh even big tech companies in 6 months everyone just forgot about the ID and I am still so so surprised that Microsoft you guys live and breathe developer DNA is in your your your DNA and it's always been how do you think you miss that this sudden shift would
6:58would come and it for the first time I felt that you didn't have an immediate response yeah no it's a it's a great uh thing to unpack right I think you to your point being early to it uh when we you know this is preGP PT4 right so when we even started there um and even just the code completions were magical that was sort of really what copilot uh you know if you remember we worked on IntelliSense forever and we made good progress and then we said wow
7:27this thing is working and quite frankly there was a lot of skepticism that it would ever work but it started and that was my even my own belief in scaling laws quite frankly was uh fundamentally uh reaffirmed let's say when I started seeing that and I thought we did a good job of trying to build that product. One of the keys of building with uh models is when model jumps happen nonlinearly, the form factor has to change quickly, right? That's the you know what is AI
7:55first product building in that sense is unforgiving because when you you have to be basically build everything around a given model and then be ready for the next model to have some emergent capabilities and rewire yourself for that, right? Don't be in love with what you built. So that's why even when agentic coding or task handling started becoming something that one could do right it's not just code completions but you could assign uh small tasks that's where cursor really you know they took the IDE they forked but they did a good
8:24job on really uh the agent mode and so we then caught up on the agent mode uh because that was sort of something that we could do on top of even the open AAI models but then the thing that hit us was obviously a great work that Ananthropic did uh with their model uh which then truly made the CLI as a form factor uh and then you could even just say just complete full uh handling of you know code be went from input to output effectively right I mean that I
8:54think was the transformation in 6 months or so uh we went primarily that to your point about uh uh ideides kind of faded away and then you really had this new way of thinking about code as input and that's why if I look at uh the new IDE it's more about how do I really input the specs and the markdowns and everything else around the software factory and so that transformation changes interesting enough it's funny that we're talking about it like one of the things thoughts I have is with these
9:24decision engine models finally I'm back again to IDE because you know an interesting way it's kind of like functions right it's intelligent functions that are embedded inside software so I think that there could be a rebirth even of an IDE um as we think about even embedding intelligence in a more token efficient way. So I don't say uh it's a permanent shift over in one way. It's about having multiple form factors now that actually help you with model capability.
9:52I do want to ask you here you mentioned decision models and how you're excited about them. Both of us recently talked with the yogo and in in different different forums. How has your thinking changed on both decision models but also on the limits of LLM specifically for for programmatic tasks for for coding for for building software in a reliable way? Yeah, I think that that's kind of it, right? Which is look, I think the LLMs are fantastic and we can talk about like even how you know we're engineers. Let's cut to the point
10:21where they're not fantastic, right?
10:22Right. uh in the sense of how do you even think about uh the software factory and there's an entire thing that we can go into but the a what I love about decision engines is it's kind of a new primitive right and how Diego talks about it and how our own decision engine and others are talking about it and I like that because you know it's kind of like you know Excel macros right you know what happened to spreadsheets with macros can happen with code with intelligence embedded as a primitive
10:52right And so I don't know you build any system like I think I was looking at an Xbox system that they had built for managing all of the community feedback right it's a classic community feedback comes I make a set of decisions based on it I don't want to really generate LLM code on it I just want to be able to make decisions so it's better to embed this primitive right it has a distribution you can make a call it's sort of a uh it doesn't need to be generated again and again it's token
11:20efficient uh and So I think that this is one of those places where software engineering remains very much alive.
11:27It's changing. One of the things I like about Diego's thing is he wants to make it as boring as databases. In fact, to me, a relational database. I grew up with my my my like the most favorite product uh in the developer chain other than the compilers and the lex and yak and what have you was the database because the relational database is a reliable way to be able to store query uh your information just made all kinds of applications possible and I think
11:55decision engines in the era of intelligence will also do the same. Well, in fact, the databases are the hard and so we talk about what is the essence of a software company. It's the business logic, right? But when you go deeper, it's often in the database. In fact, so much of the business logic used to be coded in store transactions, some of these things. So, I wonder if it's a good one.
12:16It's like it's a like this is the stored procedure of the intelligence era.
12:20Yeah. So, so I wonder if there's an analogy or a hope because I do see that we haven't found the right primitive like we know that LMS can generate code.
12:27we can use LMS to do unreliable stuff but we still we're missing these kind of primitives right yeah yeah so I'm very excited about it and this is I think the beginning of how you know because one of the other challenges you know so let's go to the software factory for a second right if you look at even how co-work is built or even the GitHub copilot runtime is built today and you know I think people have blogged about it and it's fascinating right you know in a world where code generation has become cheap what has become hard is verification right in fact one of the things that I feel like
12:57oh my god I wish you know I want to join one of these Mark Renovich Scott Henselman you know perceptor ships or whatever they call because I kind of feel like I got to learn because I've not worked in a team that is mostly now code as output right but the code is output has a particular order of operations it's not like code is output means everybody's just generating all sorts of random code like the way you got to build the verification stuff how do you take every failure and then build
13:24back the test case and make sure you truly verify the test right I mean there's just so much I would say new type of tribal knowledge people are building up in managing complex software projects in fact it was a nightmare I think even in that rust migration they were having they had to revert back um and so that is sort of all the quote the great advantages and some massive disadvantages so that's why to your point having additional primitives like
13:52this uh decision engine I think are good and you mentioned that as a as a CEO of of like this amazing company, you see all of the things happening, right? You oversee, you get information, you would want to go and work on a team if you could.
14:06Oh, absolutely. In fact, you know, my dream next job would be that, right? if I could go back and just be part in fact in fact one of the things that's interesting you know I was going to send mail to Scott and Mark and say hey they're doing an unbelievable job of creating this new program inside of Microsoft uh called perceptorship I think they even have a relationship with Carnegie Melon or something and that idea that you can have people who can come in even early in career because I care a lot about hey can you have an
14:35early in career person come in join one of these teams go up the learning curve because I think the learning curve I Remember people used to come to Microsoft for an internship. they would go see hey how is you know Cutler's code how do I go look at the Malo implementation so what's the new form of that is what I think we have to really discover and so yes if we you know my hope would be one day I get to do that again I'm now understanding why so many CTO's
15:03uh VP of engineering SVPs do take step downs to work on small teams to learn how to build with with these new tools right now seems all is changing I absolutely I think yeah the organization ational sort of context of it all uh is changing that I think yeah if I um if I think about it I think it's very important in fact you know being a lot more hands-on at a time of great change on many fronts right by the way being close to customers building in the open there's so much to learn on all
15:31aspects of what makes successful products or successful technology uh that yeah there's a plenty of opportunity we've now been throwing around the the term software factory for a little bit so so let's let's actually talk about it. I feel this term is a bit vague but it's coming around. What does it mean for you?
15:48Yeah. So I'm also still kind of coming to grips with it. So I'm pattern matching um in what is teams of people in a software team today do right. So you know I think I'd point your viewers to uh the the blog that Bass wrote. I mean in fact there are two nice blogs right the co-work team and the GitHub copilot uh team. Both of them have these blogs out. one is what one what and one individual developer was able to do that was on the GitHub
16:16side which is maintain an entire runtime um and re you know I think they also migrated I forget whether it's from Python I guess to rust um it's always to rust [laughter] to rust uh but nevertheless they kind of that's a beautiful case study of what the productivity of a single developer and what does it mean to do that maintenance and so on then there is the team I think it's nine or 12 or whatever number of engineers it's kind of crazy that that's the total size of that team um and the software factory and they've also done a nice blog but there's so
16:46many things there I love I mean this is where you know uh if if I could just go join that team and really learn it firsthand versus just reading blogs or talking to bass and team but for example the fundamental transformation as I understand it is right one is code has gone from being input to output great then you kind of have to say okay that means specification and intent right which was always important [clears throat] design docs were important and so alignment
17:16align and so that that's kind of yeah very important now the other fascinating thing I'm also finding is and I talked to one uh sort of early in career uh person who went in getting into a code base now is actually much easier because the reason when you're managing a software factory everything is documented um and so and then you can you can ask it questions patient mentor onboarding buddy.
17:42Exactly. And so you can have you your onboarding onto a codebase is so much more easier than you scrging around for tribal knowledge that's in people's heads if you will. Right. So the other thing is I'm learning a lot about this verification stuff right which is the way you want to do agentic coding is like these little tasks that you give it make sure that it's truly done. How do you have the critique? Right? There's so much to be said about having a separate model family. In fact, that's the other
18:11thing that I I I've always worried about. It's like, hey, if there's one model family and these model families are going to have all these alignment issues, how do you use a critique to ensure that one model family's challenge doesn't become your challenge? And so anyway, I I kind of look at say there's an entire new way to build uh that is emerging and it's new software engineering and it requires software engineers. It's just a level of interaction maybe that's why I like the even the analogy of you know when you know uh you went from writing your own
18:41assembly to saying oh wow I have a compiler um and suddenly you felt like oh wow I can now go from intent to code and I don't need to know assembly there's a little bit of that it's a level of abstraction but even how you work in a team has changed and how you construct a product has changed the other one that I'm really excited about is um the the traces from production coming back. I love that too, right?
19:07Which is, you know, we're talking observability, but you have more more opportunities.
19:11Yeah. I mean, when people say like, you know, RSI, that's kind of my RSI, right?
19:15Which is it's sort of the ability to take my production logs and then to be able to verify the fix. In fact, that verifiability loop is great that you can then deploy the fix and deploy it all the way through that it actually passed all the tests. Uh, that actually is very helpful. And so I think there's a lot of learning uh that these teams are having and it's spreading across Microsoft. I mean uh it's just good to see how this is now becoming the way teams are
19:44structuring. The other one that I love um is our networking Azure networking team, right? In fact, my that used to be my favorite story where this team said, "Hey, we you know Microsoft manages something like 500 fiber operators all over the world for our global network." And you know, always it's a physical asset something or the other goes wrong.
20:06You know, the fancy word for managing all that is DevOps, but the reality is it's emails and you know, all sorts of ad hoc processes. And so they went and built this entire Egyptic system that actually does the ops and that ops is connected even uh to their repos. And so now it's not just the repo on its own or the codebase but it's even the operations that are around it. And so you can start seeing how the nature of what technical teams do is build these
20:36factories of ops plus uh software that do the job versus doing the job. But still they're doing a job which is slightly different.
20:45I I do wonder because an AI labs uh it's it's pretty common and and they're very open about sharing how software engineers there their job is becoming to allow the agent and the loop to do a better job. So, so what you just said, but just sticking to practical level like at a place like Microsoft, you're you will have teams that can experiment and they're building new factories or or their platform teams and maybe they have extra cycles and they're doing all these things, but you also have product teams who actually have a product. you have real customers and this is where I see a
21:15lot of companies struggle with who do have a real business and they're seeing these aspirational things from AI labs or maybe even they're listening to you right now and they're thinking hang on like I can't just drop everything and build a factory because what am I going to do with with the thing how how are you thinking about both this balance and especially inside of Microsoft what what do you tell engineers like how is your job changing especially when you're working on a product that is a good product it's a profitable product customers depend on it yeah it's a great point I mean like first of all you know We don't get
21:43permission to do things that are not adding value to customers. Like let's start there to your point about a little bit of and we made our share of mistakes of you know going and doing random AI things just because they were you know you know cool to do. You kind of ultimately are going to I think everyone's done that.
22:00Everyone's done that and I've done my fair share of it and learned from it hopefully and we'll get better. But that said, if you sort of say, well, let's take even let's face it, right, which is today when I look at the CVE in Windows, right? Um because you now have AI models that are finding vulnerabilities. Yep. And so we now need to go fix them all.
22:19And so therefore, there's a new pipeline. So in fact, if I look at even the windows from crash traces all the way to fixes, they're building out essentially a new type of software factory except this is a brown field, right? which is the build system of Windows still is still the build system of Windows. In fact, if you think about in all these brownfield projects, the most difficult part is builds and CI/CD.
22:44Um and then of course you got to really you know even you know if I look at the amount of change that underneath even GitHub right which is if you think about it right you know we've had I mean suddenly all these coding agents showed up to work at GitHub and the scale has gone through the roof superhuman level.
22:59Yeah. And so therefore we have to fundamentally change the entire the git protocol the storage system. We are working on every part of this to make sure that we are living in the brown field as opposed to I mean the green field stuff we'll always do but the real world is all about brownfield and the brownfield requires stepwise you know hardening of systems rethinking of the the core primitives right all the way from source code control to build to
23:27CI/CD that works in a brownfield project and that takes time that takes investment that takes uh thought and the sequence of oper operations or the order of operations is very very critical. So one of the things that I want our teams is is not about well one nice thing I like about the transformation is this is not about oh here is some new shiny object and here is the team that's doing that and everybody else lives in the uh old world like everybody across the company gets to use the new tool chain
23:54and the new ways of working and tackle the challenge in front of them on any product whether it's Windows or SQL server or a cloudnative service what have you and that I think is what we're seeing To be fair, like this doesn't sound all that different to what we did in the past because when there was a new technology, let's say React, which was a new front-end technology, everyone wanted to use it. And what we did is no one got permission to just stop and migrate to React. But when you built a new product or you added a new feature, you start to add it in, you start to mix
24:24it. So maybe it's not all that different to how we're adopting AI. But but what is different about this? Is it?
24:29It's a good one. Like, you know, sometimes I I you know, I kind of say, "Yeah, I've seen this before and what have you." But I think the one thing that I do feel that's different this time around is that the the physics of how we do things right that order of operations the level of abstraction maybe in fact one of the fascinating things I think about is it is like when people moved right I I I joined the workforce at the cusp of when we went
24:58from you know in in fact in college I was mostly working on a vax right you know with a VT you know terminal uh this is a mainframe yeah or a mini uh mini right VMS systems right so and then and then I joined actually for a couple of years at sun on a workstation and then I came to Microsoft and so and then you know the world kind of was stable right you mostly wrote code in an IDE throughout my professional career it sort of didn't really change I mean new technologies came but all those new technologies were
25:27new frameworks and new runtimes it was not like con the concept count radically changed Even the mobile thing was sort of a miniaturized PC with new sort of uh user you know user interface right and new frameworks this is a place where everything what is the runtime what is the framework what is the IDE what's the stack you're right all every single par it's a little more of a change in how you think about I think then let's
25:55not forget a new model could drop right whether it is a new uh frontier model or a new thing like decision engine and it could change again, right? This is where I think we also have to be ready for the regime to completely be transformed uh at any moment if you will. And so therefore, I think these are two things at least I think about that thinking of it learning from the past when we were at our best individually or as companies or as teams in managing change I think
26:24is always important but also being open-minded that this time is different is also good.
26:29Yeah. And and also specifically for for the biggest change for us software engineers and you know Microsoft has probably majority of the workforce is software engineers or people close to the code is we used to spend most of our time writing code and most of our effort was learning at college at previous jobs to be good at this and this part suddenly disappeared which I know for a lot of people including me for for earlier it it was hard to reconcile cuz we've never seen it with any technology
26:58change it always added a new a new tool How do you see people inside of Microsoft or or outside who you talk with reconcile this who who are are really good engineers and they've spent decades getting very good at coding which is hard to do right yeah no I think that that's right I think so this is where what at least I am observing and as I talk to even people uh who are you know both been great software engineers uh previously and now are
27:26prolific even in the new mode of production is the same skills are still required.
27:32It's just again that the what you think about the order of operations, the tool chain um has changed. And so that's one of the other reasons where I don't think it's about devaluing the expertise. It's about really taking the expertise and making sure that you can exercise that with new leverage. Right? That is where the nature of how you work has to change as opposed to thinking that this is
28:00about uh software engineering no longer being important and I think that we're in that phase right uh where we are trying to figure out how to value the new thing and in valuing the new thing you are suddenly going to even start valuing the old thing if it makes sense right which is when I look at the people there's obviously somebody who is AI native who comes out of like in fact it is funny last last GitHub universe uh I went down and I met a bunch of uh college kids and this that was the first
28:30batch of college graduates who grew up in the AI era right because they went to college when GitHub copilot co-pilot right they so they didn't know an IDE world pre GitHub copilot it was like stunning to me that that was sort of what's happening so there's going to be people who are going to come at this who are quot going to be native to this tool chain but I also think that people when I see at Microsoft anyone who has great depth and expertise in any part has
28:59curiosity and is sort of learning the new ways to work are absolutely flourishing in it. And so one of the things I I'm hoping to ensure is we create that opportunity for folks because at the end of the day, right, you know, it's about economic return on your professional careers that matters.
29:17And so therefore, we want to make sure at Microsoft you can both value the skill set you have and build the new skill set that the marketplace will value.
29:25I'm really liking how you're looking at this and also how you talked about how junior new new grads, early career people are contributing. they can get up up to speed. But one thing that's been going around in dev circles, and I I'll just lay it out for you, is people are speculating, why is it that that Microsoft is making the voluntary retirement offers? Like, what does this mean? There's this fear that could this be a prelude to potentially a layoff because of AI making teams more
29:54efficient? And I just wanted to ask you dead on the head, there's never been anything like this at Microsoft, which is what I actually paid most attention to. and also the fact that it is voluntary. How did your your thinking and the company's thinking come together on this one? It's a it's a very important question and you know one of the things that we discussed a lot and we felt was let's acknowledge two things right one is that or maybe three things one is that let's acknowledge that there is going to be plenty of opportunity
30:22going forward the shape of that opportunity may be different than what is the opportunity right if you let's even narrowly think about just software engineering as we were discussing so let's start there then the next thing we should say is okay Great. Let's create more of that new opportunity inside uh the company, right? That's where everything that we're talking about, whether you're working on Windows or whether you're working on AI training, quite frankly, you get to really equip yourself with
30:51new skills that are going to be valuable inside of Microsoft and in the marketplace, right?
30:56But the shape of what this new thing looks like is going to be different than what the shape of what it was before, right? team team sizes team workflows.
31:07So we need to take it's not that we don't want the 200,000 people at Microsoft flourishing all of them and growing even but the the 200,000 people at Microsoft and growing will have to be configured in different ways right when I look at even at inside of Microsoft some of even let's say even the economic return right if you look at the wages for a systems engineer right or an infrastructure person is through the roof same thing if you are a researcher in training right so some of the expert
31:37ertise is highly competitive, highly valued, right? And so we want to give more agency. Quite frankly, in fact, this voluntary thing was mostly about let's give everybody more agency to get themselves slaughtered into places uh that make sense for their economic opportunity long term. That's I think the key thing. You know, I've been CEO for a long time. I was a senior exec before that and what have you. But at the end of the day though I did grow up
32:07as from an IC to this right. So I think about lot of the times when I was an IC or a lead right and maybe uh a manager of uh leads that's sort of the lens through which I think about how should we create opportunity inside the company. I don't want to be stuck in a team that's not transforming even inside of Microsoft. Let's face it, right? It's not like I can sit here and say every part of Microsoft is sort of really got the memo and they're all transforming at
32:37the eap I want to give them you know if you if you're in a team where you feel stuck professionally uh you can vote with your feet I don't want you to because quite frankly there are many places you can go inside of Microsoft itself. So that is another opportunity we want to create. So in that context, we are trying to give more agency to people. Uh realizing that this is a time of tumult, right? I I wish I could just sit here and say it'll all be the same.
33:04I cannot, right? And you know, and and you know, this is one of those places where I I want to be as transparent about the change. I want to make sure we do I I do the hard work of creating all the opportunity at the company and then make sure that our workforce is able to really make their professional growth come true here.
33:23I mean I think you're saying this very nicely and thoughtfully but how I would translate it is you know the world has changed big time and we saw for example with clot code has been such a fast change and the teams who fall asleep at the wheel if you will we you cannot afford it. Uh so you need to either go with this realize that the the world has changed my work as a software engineer will change my tools will change and either I'm on board with this I figure out how to be enthusiastic about this and see all the opportunities all the engineering challenges with building the
33:51factories with the new workflows or I decide well I'm not going to do that and most places you're kind of stuck but you're actually giving an option for people to decide if you would like to check out from there you can right now and there's no hard feelings I think so and I also feel Like look I think let's sort of if you come down to it and I care a lot about this right which is we have created enormous anxiety enormous anxiety in people sort of feeling that they don't
34:20have control of their own destiny because the way we have talked about this AI revolution is yeah you know been people say hey this maybe you know the singularity is here and therefore maybe we are building there will be one company or you know there'll be no jobs. All that sort of leads to anyone who even wants to change be sort of anxious.
34:46And so I kind of feel that this is where look, we're a company of 200,000 people. You know, I know that there's going to be a lot of change in what each of us do. Like think about it, right? One of the fastest growing jobs at Microsoft is FDE.
35:02Yeah. Across the industry right now.
35:04Across the industry. Because in some sense if you think about it right you have an account rep very important then right next to the account rep is sort of our uh customer uh you know the technical support folks right and the customer um it's I think CSU is what we call a customer success uh and then we have even the technical support people and then you have uh the FD team forward the engineers all of these things require a complete like they all should be going to the
35:33same preceptor ship that any software engineer is going if you will right because you can all like I know even LinkedIn I think they made this change inside where they call them full stack builders right which they they combine roles so all that change I think there's opportunity but you have to sort of walk the hard walk of really making the structural changes getting people slaughtered giving them opportunity giving them uh the wage support knowing
36:01knowing how to evaluate people with their productivity and their excellence in their new jobs in their new slots.
36:09This is the hard work any incumbent you know so that's why you know the celebration of the new which always should be there right that's how competitive marketplaces work uh look you know we are a 51-year-old company we're going to sort of be comfortable in our skin we want to serve our customers well we want to create fantastic economic opportunity for the employees at this company doing great work given our mission to empower people and organizations on the other side of this but I I just want to get this on the record so when
36:38we talked out hiring new grads. So you you are still hiring new grads, junior people. You're investing in them. You're seeing that because one of the worries in a lot of other companies and startups is and they've slowed down or even stopped hiring because they're thinking well with AI the most the people who can use it the most efficiently as as we previously mentioned the ones who have expertise and they can leverage that and they're very hesitant to hire any new career people.
37:01That's a great point. So, so one of the things that we're trying to do is in this change, this is where I think a little bit of the trends have to be watched carefully over sort of slightly longitudinally versus just you know one year because um we're in the process of a lot of change right so first job is to make sure internally all of the folks are getting reorged to places that make sense right step one I mean like even inside of a team right you just want to
37:30get the configuration Right? Then it makes sense to oh what is the industry higher or a new college h higher you add to this right and so a little bit of what's happening I think is the slowdown may be a slowdown to get those quote unquote new factories of software engineering built such that new inputs not just tokens humans can come in and then in this new configuration give you high leverage right I mean let's face it
37:59when what is an AI native company they start with a token budget and a human, right? Then they add that person maxes their token budget and and then says let's add the next human and then that human gets a token budget and then their factory gets the budget. Yes, that is the constructionist way and it's not like okay those companies if they succeed they will also be big one day and they will and some already are are and you every company is going to start looking like that and every project team
38:28or product team will look like that and so that process of reconfiguring new discovery uh and configuration it'll take time and that's where I think you know trying to draw too many lines as long-term trends I think it's too early. Before we get back to the conversation, I'd like to talk about Linear, our season sponsor.
38:49Until recently, I was thinking of Linear as a tool that helps plan and organize the work, but it's much more than that.
38:55A lot of teams use Linear to ship changes from idea to merge in GitHub, all without leaving the tool. I sat down with Dimma, an engineer on the linear team, to see how devs inside of Linear use linear to build all of Linear. So, here's a real issue. When you submit a review with unsend comments, Linear warns you about these comments. but it didn't tell you where those comments were. A customer complained and Dimma decided to fix it. But for the fix, Dimma did not open his ID. Instead, he opened up Linear and described the issue to the agent. The agent built a fix,
39:24created a PR, and attached storybook screenshots for easy verification. You can see the fix with the draft comments linked on this screenshot. In the PR, you can then open the preview with one click from linear to play around with the change and see that it worked. And here's my favorite part, guided reviews, which is work in progress. Guided reviews is a new code review experience where the important stuff just pops out.
39:46The top of the review has a summary and a diagram of the change and the poll request is divided into chapters that walk you through what has changed. These blocks highlight the important changes in a way that's easier to understand and review. Oh, and when Dimma is finished with all of this, he just self- approves the code and it's merged into GitHub.
40:03All without leaving linear. If you take away one thing from this, Linear is no longer just a place for planning tickets and organizing the work. It's where all the work happens. Creating the issue, kicking off the agent, checking out the preview, and doing the review and merging to production. Linear is the product development system for teams and AI agents. And I'm loving this evolution. Linear is the product development system for teams and for AI agents, and I'm loving this evolution.
40:28Check it out and get a six-month trial at linear.app/pramatic.
40:32If you've ever worked on B2B software, you know how this goes. A customer likes a product, then their security team sends a list. SSO, skin provisioning, rolebased access. None of it is what you set out to build, but the deal waits on it. Work OS is the infrastructure that covers all of that. So nobody on your team has to debug a customer's sample setup. The newer version of the problem is agents. Once an agent is acting inside of a customer systems, the enterprise wants to know what it's allowed to do. That's what works airlock
41:02is for. It's intentbased access control.
41:05Most access control only asks who you are. Airlock also asks what the agent is trying to do at the moment it tries to do it. Clear-cut rules like no agent deletes cloud resources run as plain code. For the fuzzier air calls, an AI judge evaluates a policy at runtime. You write these policies in plain English and Claude can pull the ones you already had out of notion or JR and import them for you. Your agents, whether that's Claude or Devon, connect through a single access point and your MTP servers
41:32sit behind it. Airlock is a new layer on top of your existing permissions. So an agent can never get more access than the person who invoked it has. When something risky comes up, airlock can escalate to human for approval, say over Slack. The idea is to write the rules once and spend human attention only on the calls that need it. And every request and decision is logged. Visit workwise.com/airlock to learn more. Now we know when everyone is building their their software factors or changing how they work. A big trend
42:01now that we go down to a little bit of the how this is done. big trend across the industry that I'm seeing and I'm sure you're seeing as well. A lot of tech companies are moving to open models as well. The a lot of them are are spinning up their inference or companies like Revolute are buying their inference. They're buying their B200s and just launching the models on it.
42:19They all have their own reasons. But from your vantage point, you know, Microsoft is also in the open model serving business with with Foundry, for example. How big is this trend and is this in inevitable? I'm asking myself are there parallels to the open source movement and now the open model movement?
42:35Yeah, I think that there are for sure the parallel which is you know you know there's sort of going to be closed models and there'll be competition amongst them and there will be sort of an alternative to obviously closed would be open. uh open gives you flexibility.
42:50In fact, I think you know, one of the ways I come at it is with just this what is a frontier ecosystem look like, right? It's can't be if you just get past this thing called uh there's going to be one frontier uh model to sort of have all of the world's knowledge, right? If you don't buy that, right, uh then you have to assume that there should be in some sense as many frontier models as firms in the world, right?
43:15Because after all if you go back to actually Hayek uh the his description of what an economy is it's all about knowledge being distributed right in fact his you know his entire thing about pricing uh is a beautiful sort of explanation of hey it happens because of knowledge being tacit and distributed throughout the economy and every firm needs to be in control of it and if so then I would even say every firm is
43:43essentially got both human capital and token capital. So if every firm is going to have both token capital and human capital, the token capital can be an openw weight model into which you know you've embedded your knowledge. That's one. It could also be in the context layer. You could even be using multiple closed models. in fact collecting even the traces uh of your usage of those closed models and then using those traces to maybe train an a small model and an open weight model for yourself
44:13and keeping you know combination when I look at even copilot in fact whether it's github copilot or co-pilot they're all designed as these multimodel system right in fact I would say maybe even by the next time you and I talk everybody would say it's not about even open or closed everyone has a model system in Fact, it starts with even a decision engine which says which model to use for what task and then I have my context and memory and my action space separate from
44:41any one model because otherwise you have a complete lock in. In fact, one of the things I I I think about it's like, wow, if you really got your harness, your model, your memory and all coupled, how would you ever get out, right? In fact, what are you doing and what what do you own versus what is owned by the model?
44:59Uh because no, the model will basically collapse into it everything it sees. Um and so I think that this dynamic is getting well better understood. It's you know most people again think oh this is just some competition amongst tech people. Remember there's no limit to intelligence. Every bank, every healthcare company, every insurance company, every analyst out, everybody needs to sort of have a real thing architectural approach to figuring out
45:28how do they manage to control their own learning and knowledge and have firm differentiation. And in that context, open rate models is one strategy that I think has legs. You know, this is very interesting because I I talk with a lot of midsize tech companies, startups, and almost all of them are building the things that you just dropped, their own harness, their own context layer, their own own context graph, their own inference. And I thought initially I'm like, well, this kind of makes sense
45:56because engineers love to build and if you want to get AI engineering experience, the the least risky thing is to build something internal. However, I never thought of of this part where it's actually these engineers building for the right reasons, which is the strategic reason or not. They're probably doing their company a service.
46:11And you're saying that most companies should probably consider owning this system, having replaceable parts, primitives because that's correct. So, in fact, what you're observing is correct, which is it's fascinating, right? In fact, I would say if you're a bank, you should check how software companies that are trying to manage in this ecosystem to become frontier uh are doing it. they're not seeding control. So in some sense that's what the bank should be doing. So what does the bank in there they will have
46:40leverage. So for example what we have done you know what we describe as this IQ layer is we have taken for example the harness that we have underneath copilot and github c-ilot made it a firstass thing that you can now go to foundry as a managed service and you have the harness. So in there, let's say you're building a supply chain system, you can start with our harness just as essentially an open or you can bring PI or any other open- source harness as well in there.
47:06Uh but you can just have it like as a managed service, right? So battle tested uh harness from Microsoft which we ourselves are using is just available there and then you can have all the foundry models, right? So you have a harness, you can use multiple models as endpoints, you can have your context layer that is completely separate, right? your memory and context uh is yours then so I think now we have a model system uh oh by the way the action space so with all the plugins and skills
47:33and connectors right all separate from any given model being only having the visibility to it so I think a good sort of classic middle tier platform is emerging even for model first development where you're not just captured and then you can eval everything right you can say hey I have an eval and my private RIV evals will matter more, right? All these public benchmarks, which by the way is great, but you need to have your own private eval for what is the outcome that
48:02matters to you and then look at in your RLE the real performance for across all these models and continuously optimize.
48:10In fact, when I look at even, you know, even with our customers, they start with a frontier model, then they say, well, great. I've got all those traces. Now, I'm going to flip to, hey, this is a deterministic workflow. I mean, it's not like I don't need the next big model drop. I want to now cost optimize. I'll drop, you know, an MAI flash model, drop it in there, use the traces to fine-tune it, and there you have it. So therefore, that's all stuff you can do in Foundry.
48:35So that type of work, I think, is what's going to increasingly be true. We're we're saying that's you're you're already seeing that a lot of companies have their reinforcement learning environment as well on top of the evals so that they can they can stay in control control pricing and also I guess not leak their business data to some of the AI companies that might want to train their models on it.
48:56That's right. So if you think about it right so the way it's a simple analogy I have at least in my head is you know people you know in the past whether you know in an insurance company in a bank in a healthcare company said hey I'm going to get a relational database great I'm going to deploy a relational database on top of it I may even buy some software that runs on that relational database that is sort of comes with its own business logic but then I'm going to extend it with my own business logic that I write right and
49:24then I control it and never at any point is my data in the database going to somebody else. It's not like when the database vendor says, "Hey, I'm no longer going to give you a database update." You still can migrate out.
49:38That's the question you have to how are you going to migrate your memory out if your memory is in the embeddings of a given model when that model company says, "You know what? Tough luck. I can't do a next update to you." Plus, I guess what we talked about when we have new primitives come, how are you going to integrate them? Great point because because if you're too reliant on on a vendor or a series of vendors, you'll have to wait for them to maybe do it.
50:03And I guess it's also easier to build right now like right if you have competent engineers, we're good at this.
50:09Yeah. So, I think that there are two enterprise architectural decisions that I think interestingly enough are I'm hoping that people are going to start talking, right? because and in fact to your point all you have to do is look at a software team today in a software in industry they're all figuring this out to stay independent I think this has to become more of the reality in the commercial space uh and there are two are how do you avoid this reverse information paradox which is don't leak your value you as a firm need
50:38to be in charge and control of your own learning loop what's the architectural approach how are you making sure your context is separate from the models and you're using multiple models mod uh that's and your harness is separate from the context and the models because that I think is the only way you can assure yourself that you can replace any model.
50:57In fact, my asset test for that is you should be able to replace any model and still hill climb on an eval that's your private eval.
51:06If you can't do that, then you have to ask yourself how are you a sovereign firm, right? Like that means you do not have control of your own output. Yeah.
51:18Uh so that's kind of the just sitting on taking time bomb.
51:21Yeah. I mean you're done. I mean like is my my my point. And so the then the next thing is you now need a trust architecture right which is you kind of have to like in a world where you now know scaling laws work except with scaling comes new capability but not alignment. In fact, no mechanistic understanding that hopefully someday there's a science breakthrough, but right now we're getting more and more capable and less and less reliable like which is the most dangerous place to be in. And so that means you have to
51:50separate out architecturally authority from intelligence. If your intelligence has authority and it's unreliable, again, you're taken, right? You have a massive insider risk problem. And so you've got to treat these agents as insider risk. That means you got to have a trust system. And there are a lot of things right even this chain of thought man like this the amount of like this chain of thought has to be not in new release it has to be in in in human language it needs to be auditable it
52:19needs to be logged such that it's not tampered it's not just in you know pre-training or in uh you know RL that things happen right which is in test time compute things could happen because they could reward hack and so we have some serious amount of safeguards you have to build um at at runtime as I describe it which is the trust architecture that ensures that you can treat models and and agents as inside a risk and monitor them and have
52:48observability and what have you and then ensure that you have control over your learning loop as as an engineer all all I'm hearing and all I'm seeing is there's so much new type of work that needs to be done so many novel systems need to be built which we don't have a blueprint for I mean and probably the best way to do is well build it also talk with peers who are building it at the same time and I'm sure we'll have blueprints and certifications in I don't know 5 to 10 years or something but right now this is a job it's correct and in fact in fact one of the things we have our conference
53:17Ignite coming up next week and I'm excited about it because that's kind of what I want to make sure we do a good job is communicating that hey this is all doable like there are blueprints there is a lot of innovation happening in how companies and so I think that we're reaching that stage where it's not just about the shiny object and the novelty effect and I think people have to sit back and say hey how am I creating enterprise value in my firm and once people sit and talk about that in a little more depth then I think these
53:45patterns of success become much clearer now as as an engineer one things that I I feel very lacking is all of these AI models they run in the cloud on someone else's server or a super expensive B200 that I cannot afford or my company will not buy it for me or if if if you are I mean you're lucky but would it not be nice if it just ran on my laptop That's a great and local AI until now it's been difficult very hacky actually funny Microsoft is is the one that's innovating there's a new Surface Ultra
54:14laptop that Scott Hanselman gave me a demo I now want to get my hands on because it it has a chip it has a local coding model running on it where do you think local models fit into the the AI loop and as a developer I feel is desperately missing but am I kind of alone you talk with way more people it's it's a great point right so one of the things is we're very excited about making sure that the the PC as the form factor um you know we innovate in uh in
54:44the age of AI and so to me the way I've been thinking about it and how we've been executing on it is I've always felt like distributed computing will always remain distributed so why not use all the power and all the tokens that are out there versus just the stuff in the cloud so that's sort of the inspiration even for ultra but but there is lots of innovation in fact from you know whether it's Nvidia or AMD or Qualcomm so it's fantastic Intel there's good silicon work happening and systems work and this
55:14is in fact it's already true what is it like last year's frontier can run on your laptop and last quarters uh Frontier can run on your desktop I mean like that's kind of like the capability so it's like how pieces used to evolve right exactly so it's great so now that you have that then one of the things that I think about again goes back to my decision engine and routers right so when I think about hydra fusion um I think that's the term we use I forget now which in github we have a you know a
55:41way to route right I love the github uh because all nowadays in fact in copilot the last time I chose a model copilot fusion yeah copilot fusion so basically we want to be able to have this router pick the models right for the task and so that means in fact I see it in ultra sometime I mean it you know my suddenly I see the fan go right because my local model is being used because especially in an agent swarm uh where you have multiple sub aents then you can invoke for some
56:10sub aents a bunch of local models while you're using even the cloud models together u and so yeah this makes sense architecturally it makes sense economically it makes sense and so we're in the beginning of what I'll call this hybrid intelligence era cloud will continue to grow there's going to be infinite amount of sort of demand for these tokens but that means you can also start thinking about the, you know, the form factors at the edge as also token factories. In fact, I kind of think of it as the unmetered intelligence.
56:39Yeah. Because you you don't need to worry about how much you're using it.
56:42I guess this is where it's now starting to come together for me because the uh MAI code uh what the the Flash 1.1 model, I'll be honest, I wasn't too impressed. I don't think anyone was too impressed. But the one detail I was impressed with when it comes to local models running on any laptop, I always assumed it must be open models. However, the MAI code can be downloaded on a Windows machine even though it's a closed model and it can be run there.
57:06There's the infrastructure there. And I'm like, huh, could there be a future where like other closed models? So I I could choose on my laptop like pick and choose like either I download them and maybe it's free or I pay very little.
57:17you know to your point about the the code flash models getting them even all the way uh to a laptop is core priority for us and we are excited and that's kind when I look at the code tab inside of copilot right which is uh even you know independent of github copilot now code writing a document or writing an application you can easily use these models on your PC and that's exciting but just to kind of step back and say what are we doing with even MAI models right first of all through the partnership with open AI we have the
57:46Open AAAI models and and we're going to continue to use those OpenAI models, right? Uh on on top of that, what we are saying is okay, let's sort of innovate in a new model family ground up that ground up meaning you build.
58:00Yeah, first of all, we have a you know, I think we disclosed even the new MI2 base which is the pre-training and in fact it's one of the best disclosures of a pre-training recipe and regime. Uh so for those who are uh you know your listeners should go look at it, right?
58:13It's probably and people tell me from the outside that it's the most detailed descriptions of pre-training, right?
58:20Because there are very few places where you can go to and get it and it's completely, you know, clean room so to speak, uh, implementation of a a pre-training run. So therefore, we're continuing to do that. But then the place where we are able to sort of innovate is put I'll call it new points on the frontier curve where using our products, right? Whether it's cyber, whether it is knowledge work, whether it's coding, how do you put new points
58:47on the frontier where you're both flops efficient on training and you are efficient on tokens in deployment and once you are in a world where you're handing high level intents to many sub agents, the sub aents can be using multiple model families, right? That's I think the way like this idea that oh I'm like wedded to one model family may not be true if there are choices agents will
59:13be smart in picking all of the models that allow them to get the task done in the most efficient way and that I think is our goal. Our goal is to give more choice as opposed to just one model family.
59:26I I really like how like you're you're setting out choice because as a developer I love choice. It also helps least to competition, better prices and so on. And as closing, this changes very fast. Everyone I talked to who've been in the industry, they said they've seen, you know, like change maybe similar to this, maybe not as big, but not not as quickly. You've been through multiple change cycles. When you started, it was a very different time. How would you approach this change as a software professional when your goal is I would like to be a amazing software engineer
59:56in this new age of AI in a year or two?
59:58you know, it's sort of the stuff we talked about. Quite frankly, even for me, I think just making sure that one gets grounded in what is this change about, right? Uh it's not just about the headlines and it's not just about the headlines, you know, I love X and you know, it's great, but it's just not about uh the hype cycle. Yeah.
1:00:21And that constant need um uh to talk.
1:00:27But it's ironic in a podcast to say that. But it's do do more uh learn more and and hands-on experience experience it firsthand. That's you know quite frankly that's kind of what I would even say for me if I could change the ratio of my you know talk to do I would do it. This is the time. If one piece of advice, whether you're mid-career or even a CEO like me or or a
1:00:53fresh uh college grad, the more you can take the time to sort of learn to do be in teams that are doing uh the better it is versus um I think anything else because at times like sometime at some level I sometimes think about it and when am I at my best? It's when I'm in the school mode. Uh so let's get back to school.
1:01:18Satia, thank you so much.
1:01:19Thank you so much. It's such a pleasure.
1:01:21Well, this was an awesome experience and we really [music] covered a ton of ground with Satia. A recurring theme that kept coming back. What kind of software engineering work becomes more important when coding becomes cheap with AI? We talked about how the things both before and after coding that is planning and validation both become more important and this will be activities as us engineers will spend more time on. I also observed a lot of tech companies building their own coding harness, context layer, and AI infra. [music]
1:01:49And I assume that it's because engineers just like to build with novel technology. But Sachio pointed out that it's actually suicidal strategy for a company to not own their AI infra and their [music] AI intellectual property.
1:02:01In fact, he thinks that every successful company will build out this own [music] AI infrastructure. And those relying on a single vendor or AI lab are making a massive strategic mistake. Talk about us devs enjoying building new stuff and it's paying off. I can see us having a lot of opportunities to build this new type of AI infra and these internal software factories. [music] Finally, software engineering roles and expectations are also changing. Sasha has shared that the one of the fastest growing jobs inside of Microsoft is forward deployed engineer. This is a position that has not existed a few
1:02:31years ago. Sachi expects to see a lot of reorg inside of Microsoft where [music] teams become smaller and roles change and the voluntary exit program is an opportunity for people who leave who don't want to go through what will probably be a really intensive period of change much more than before. I would assume that this kind of change with reorgs and role changes all of these being more intense than before [music] is probably not just true for Microsoft but it will be true for a lot of other tech companies. In the end [music] I walked away being more energized from this conversation. Yes, the industry is
1:03:01changing. But what I understood is how Satia sees that a lot of software engineer work is waiting to happen with these new AI related primitives like decision models and some of the primitives that have been not been invented and users will need to build these new systems with these new primitives deciding what goes where, how to change it, what makes sense, how to deploy it and just keep learning from it. Do check out the show notes below for related deep dives on AI engineering concept like evals, inference engineering, loop engineering and more.
1:03:28If you've enjoyed this podcast, please do subscribe on our favorite podcast platform and [music] on YouTube. A special thank you if you also leave a rating on the show or share with colleagues or friends. Thanks and see you in the next