Free with a sign-in.
The best 16 min of 64
Software engineering after code becomes cheap with Satya Nadella
Satya Nadella argues that cheap code generation shifts software engineering toward intent, verification and AI systems firms can control.
The Pragmatic Engineer · Gergely Orosz
- Satya Nadella is CEO of Microsoft.
The best minutes
- 9:55AI coding shifts engineering toward intent and verification4 min
- 15:32Software factories make codebases easier to learn and operatePro pick 5 min
- 46:03Companies need control of their own AI learning loop7 min
Also worth watching
- 20:52Legacy products need staged AI-driven hardening3 min
- 29:19Microsoft's voluntary exits give employees a way to change course7 min
- 6:48Model jumps forced Microsoft to rethink its coding interface3 min
- 36:29Hiring may pick up once AI-native teams are configured2 min
- 53:49Local models can handle sub-agent tasks alongside cloud models3 min
Skip the Linear and WorkOS sponsor block and the closing recap and subscription request.
Full analysis · 6 min read
Real Alpha · Nov 16
Satya Nadella says verification is the hard work after code gets cheap
- Code generation is cheap, Nadella says, while verification has become the harder work.
- His factory examples show workflows, not measured productivity gains or proven cost savings.
- Teams should budget for test design, production feedback, model portability, and runtime controls.
Free with an account
Read the rest of the analysis
6 sections, the 6 other voices it tests Nadella against, and what it means for you. Every episode we cut gets one.
Check your inbox
We sent a link to your address. Click it to finish signing up. Not there in a minute? Look in spam.
or
Signed up before? The same buttons sign you in. The episode, its moments and the transcript stay free.
Cheap code moves effort toward checking
Nadella's central reframing is that generated code is becoming output, while the engineering job shifts around it. At 12:47, he says the harder work is checking what code produces, not producing more code.
He describes taking each failure and building a test case from it. Separately, he says production logs can help teams validate fixes, with tests run before deployment.
That changes what a team should count as throughput. But lines or changes produced matter less if review, test design, and regression checking grow at the same pace.
The episode leaves the cost side unmeasured. Nadella describes new verification knowledge and a different sequence of engineering work, but gives no baseline for checks per task.
On Unsupervised Learning in November 2026, Nebius CTO Danila Shtan said agent reliability may matter more than simply scaling models. That is adjacent to Nadella's point, not evidence for his specific account of code verification.
“what has become hard is verification”
Our read
The shift is credible as a change in where effort goes, not proof that total engineering effort falls. Teams should track checking time and escaped defects alongside changes shipped.
Factories connect intent to production work
Nadella's software factory is not a prompt wrapper. He describes clear intent, a codebase agents can learn, small tasks, tests, and feedback from production.
For example, he cites Azure networking, where an agentic system handles operations tied to repositories and Microsoft's global network. He separately praises production traces returning to teams for validating fixes, without saying the networking system uses that same loop.
So the distinction matters because a factory is a system design, not one branded tool. Its parts need to meet within a real product workflow, including operations where relevant.
That also explains why codebase knowledge matters in his account. Nadella says agents can make onboarding easier by answering questions that otherwise depend on tribal knowledge, a plausible access benefit but not a speed measure.
Zach Lloyd, Warp's CEO, described code-defined factories on Latent Space in November 2026. Nadella's account is broader and more tentative, since he says he is still learning from Microsoft's teams.
“everything is documented”
Our read
The factory idea earns its keep when a team can replay work and trace results. Documentation alone is not a productivity metric, and Nadella gives no before-and-after comparison.
Legacy systems need staged hardening
Nadella's clearest case for adoption comes from legacy work, not a clean demo. He says AI can surface Windows vulnerabilities, but teams still need to harden the existing build path.
Take Windows, whose build system remains in place. Nadella says builds and CI/CD are difficult parts of legacy work, making sequencing and investment part of the engineering task.
He also says coding agents pushed GitHub to rethink its protocol and storage. Yet these are examples of work underway, not measured results or evidence that each migration is complete.
On Moonshots in October 2026, Salim Ismail said agents can operate existing software without redesigning it for AI. Nadella adds the constraint of hardening the systems around that software.
For an operating team, this argues for choosing a bounded job first. Measure whether the change improves the existing pipeline before turning a small success into a broad migration.
“the real world is all about brownfield”
Our read
This is a case for incremental adoption tied to customer work, not a reason to pause product delivery. The episode gives no schedule or cost for Microsoft's hardening efforts.
Companies should keep their learning loop portable
Nadella's enterprise prescription is to separate the harness, context, memory, action space, and models. He says private evaluations should test whether a company can switch models and still improve outcomes it cares about.
But that is also a commercial position. Nadella says Microsoft Foundry offers Microsoft's harness as a managed service, alongside model choices and separately held customer context.
The episode does not show that every company must build these layers itself. A practical portability test is to compare models on a private evaluation while preserving context and task history.
Sagar Vaze made a related but narrower case for open weights on MTS in October 2026. He said firms can adapt models around proprietary expertise. Nadella includes open weights as one option, while also describing multiple closed models and independent context.
And the discipline matters because a vendor's model can change while a firm's own process needs continuity. In Nadella's framing, retaining evaluation and context helps a firm keep learning across model changes.
“you should be able to replace any model”
Our read
Portability is a testable architectural goal, not something the episode proves every company needs to build from scratch. A private evaluation and model-switch test reveal more than a general claim of independence.
Agents need less authority than their capability
Nadella says capability does not guarantee reliability. He wants runtime safeguards and monitoring, with an agent's decision power governed outside the model itself.
That warning matters because a coding agent can affect source, builds, or production workflows. Still, the transcript names insider risk but does not lay out a full control framework.
Roman Ugarte, Grok Bot's product lead, described read-only access before broader delegation on Lightcone in November 2026. That staged permission path is Ugarte's approach, not a procedure Nadella described Microsoft using.
Ethan Mollick similarly described human check-in points for agent tasks on Silicon Valley Girl in November 2026. Those examples offer concrete control choices alongside Nadella's broader warning.
And the separation matters most when agents execute tasks rather than merely suggest code. The more consequential the action, the more a team needs a policy, a record, and an escalation path.
“separate out architecturally authority from intelligence”
- RURoman Ugarte · Lightcone · Nov 14Bots should earn broader access as people learn to trust them
- EMEthan Mollick · Silicon Valley Girl · Nov 3Manage agents with feedback and decision points
Our read
Separate what an agent can infer from what it can change, and make its actions observable. Ugarte's staged-access approach suggests a testable starting point, not a Microsoft policy described here.
Junior hiring depends on team configuration
Nadella says it is too early to read slower junior hiring as a lasting trend. He expects team configuration to come first, with hiring following once new workflows take shape.
His AI-native team model starts with one person and a token budget, then adds a hire as capacity warrants. That is a staffing model, not a hiring forecast.
The voluntary exit offer is framed as giving employees agency through change. Nadella also points to internal mobility and describes demand for infrastructure and AI expertise.
This makes the staffing story conditional. A team that has not decided which work agents own cannot confidently size the next hiring cohort.
He gives no uptake, hiring, or post-reorganization figures. So this interview cannot show whether junior hiring will rebound, or whether existing employees can move into the growing roles.
“we're in the process of a lot of change”
Our read
The hiring outlook remains conditional on how teams are configured. For planning, treat junior roles as a capacity question to revisit, not a settled forecast.
What it means for you
- Measure verification load. Track review time, tests added from failures, and escaped defects on agent-assisted changes. Compare these with completed changes, not raw code volume.
- Map legacy bottlenecks. List the build, source-control, and CI/CD steps slowing safe changes in one existing product. Pick one customer-driven workflow to harden, then measure the result.
- Test model portability. Run a private evaluation across two available models, then check whether context and task history remain usable after switching.
- Set agent permission gates. Start with read-only or otherwise constrained access and name human review points. Expand permissions only after the relevant workflow passes its checks.
What to watch
- Whether Microsoft reports comparative measures for agent-assisted verification or productivity.
- Whether junior hiring changes after teams complete the reconfiguration Nadella described.
Send this to the VP of engineering and finance lead setting next quarter's AI budget.
Watch the minute where Nadella says verification becomes harder than code generation.
Written by Real Alpha from the episode's transcript and the other moments it links, then checked against the transcript. Quotes are the speaker's exact words.
Minute by minute
A Pro write-up, open to you here. Try Pro free for 7 days
A Pro write-up. The minute by minute is in Real Alpha Pro; the episode, its moments and the transcript stay free.
Read it with Pro, 7 days free- 2:12Microsoft built with open-source communities
Orosz notes Microsoft's shift from a closed-source reputation, and Nadella says the company has contributed to Linux, moved workloads such as SQL Server to Linux and open-sourced TypeScript. He says Microsoft aimed to earn the open-source community's trust by doing right by it. Orosz also points to Visual Studio Code as a free, open-source product that became widely used.
- 6:48AI coding keeps changing form
Nadella says Copilot began with code completions before GPT-4, Cursor advanced agent mode, and Anthropic's model helped make the CLI a new coding form factor. He says AI products must adapt when models gain new capabilities. He also sees room for decision engines: embedded, token-efficient intelligence for repeated decisions, like spreadsheet macros.
- 12:26Verification is now the hard part
Nadella says that as code generation gets cheaper, verification becomes harder. Teams need to check each task, turn failures into test cases and verify that fixes work; he says this requires new knowledge for managing complex software projects.
- 15:32Software factories change team workflows
Nadella describes a shift from code as input to code as output, with specifications and intent taking on greater importance. He says documented codebases let new team members ask an AI assistant questions, while agents handle small tasks that engineers verify. He also points to production traces feeding back into tests and fixes, and to an Azure networking team's agentic system linking operations with its repositories.
- 21:31AI must work in existing products
Nadella says teams still need to add customer value, including by using AI to find and fix Windows vulnerabilities. He describes adapting brownfield systems as work on existing build systems, source control and CI/CD, with the order of changes mattering. Unlike adopting a framework such as React, he says AI can change the runtime, tools and stack themselves.
- 26:53Engineering expertise still matters
Nadella says software engineering expertise is not devalued; engineers need to apply it through a changed toolchain and workflow. He sees both AI-native graduates and experienced engineers with deep expertise and curiosity flourishing. He says Microsoft should help employees build skills that the company and the wider job market value.
- 29:19Exit offers give employees agency
Nadella says team sizes and workflows will change, even as Microsoft tries to create new opportunities internally. He frames the voluntary exit offers as a way to give employees agency to find roles that fit their long-term opportunities. He also acknowledges the anxiety created by predictions about AI and jobs.
- 36:29Junior hiring may follow team redesign
Nadella says hiring trends need to be watched over time as companies reconfigure teams and build new software factories. He suggests the slowdown in hiring new graduates may reflect the time needed to make those teams ready for new hires. He describes an AI-native team as starting with a person and a token budget, then adding people as the work grows.
- 41:57Open models support firm-specific knowledge
Nadella expects closed models to compete alongside open models, which he says give firms flexibility. He argues that firms need to control their own knowledge and can use open-weight models, context layers or multiple closed models. He says usage traces from closed models could help train a smaller open model for a firm.
- 46:13Companies need replaceable AI layers
Nadella says firms should keep their harness, context and memory, action space and model choice separate. He describes Microsoft making the Copilot harness available through Foundry as a managed service, where customers can use multiple models and their own evaluations. He says replacing any model while still improving on a private evaluation is a test of whether a firm controls its AI system.
- 51:11Agents need a trust architecture
Nadella says increasing model capability does not guarantee reliability, so companies should separate intelligence from authority and treat agents as insider risks. He calls for runtime safeguards, monitoring and observability, along with auditable, logged reasoning.
- 53:39Local models can complement cloud
Nadella says a router can assign some tasks or subagents to local models while using cloud models for others. He argues this hybrid approach makes architectural and economic sense, and sees edge devices as another source of tokens alongside cloud computing.
- 57:28Microsoft plans a multi-model lineup
Nadella says Microsoft will continue using OpenAI models while developing its own MAI model family, including MAI-2 Base. He says the goal is to offer choice, with agents able to select among model families for different tasks rather than relying on one.
- 59:25Hands-on learning beats the hype
Nadella advises people at every career stage to focus on understanding the change and learning through hands-on work with teams that are building. He says he would change his own ratio of talking to doing.
Published Nov 16, 2026.
Transcript · 52 min read
The Pragmatic Engineer, hosted by Gergely Orosz, with Satya Nadella.
Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.
What does software engineering mean after code generation becomes very cheap? Today's guest is Satya Nadella, Microsoft's CEO, one of the few big tech CEOs who started his career as a software engineer. Today we cover what the term software factory means for Satya and how it looks like today inside of Microsoft. Open with models, 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 what Satya 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 cutting-edge companies like Entropic, LinearRamp, and many others. Many companies who switch to TurboPuffer are cutting their search bills by 50% or more and getting much faster results. Check it out at turbopuffer.com slash Pragmatic. Satya, welcome to the podcast. Thank you so much for having me, Gurgay. Well, it's great
to be here because I'm not sure if you know, but I was a Microsofty. Were you here or where were you? I was in London and this is the first time I'm here in headquarters. So, you know, like 12 years back. But I want to rewind back there. I was actually at Microsoft when you became CEO in 2014. And 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. And I just wanted to rewind back. Why did you, people were surprised. My friends outside of Microsoft were surprised. Even I was surprised. Why did Microsoft go all in on open source? 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 looked at even, I would say, commits even into Linux, even pre me becoming CEO, I think Microsoft had significant contributions and were committed to it. But the thing that I think we really made sure that we understood internally and the world understood is Microsoft is committed to open source, right?
Whether it is bringing some of our workloads to Linux, like SQL Server, or contributing to Linux, or open sourcing projects like TypeScript and others. I think we're all part, and obviously us as stewards of GitHub put us at the center of it. And we just wanted to make sure that we were able to earn the trust 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?
Read the whole transcript
In fact, if you look at some, quote, unquote, a lot of these other companies that have benefited the most from open source have hardly contributed. It's never. And that is, I think, how, you know, like, hey, you can't complain, but that's how the world rolls, which is it's on narratives versus facts all the time. Well, it's narratives versus facts. That's true. And again, I think there's that Windows legacy from the 90s that shits people's, 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.
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 an amazing piece of software. And somehow you decided, Microsoft, to bring Visual Studio for free in an open source version to the broader community. And boom, 10 or 5 years later, it's everywhere. Everyone was using it. was that intentional yeah in fact uh you know that you know eric Amar and others who were at that time in that place really was absolutely intentional, even building with community. In fact, when I look at even product making, it was the first time I also got exposed to what I think now people talk about, building in the open. That 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 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, you know, I listened to my vice president. I listened to my community. So there were many, many cultural aspects of building in the open, 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 learning more deeply or more broadly, Visual Studio Code was everywhere. GitHub 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 ChatGPT 3.5, this was, I think, the 3.0 version 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 copilot and cursor were going has a 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 even big tech companies in six months everyone just forgot about the id and i am still so surprised that microsoft you guys Live and breathe, developer DNA is in your DNA and it's always been.
“when model jumps happen non-linearly, the form factor has to change quickly”
“in a world where code generation has become cheap, what has become hard is verification, right?”
happening, right? You oversee, you get information. You would want to go and work on a team if you could.
Oh, 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 called Perceptorship. I think they even have a relationship with Carnegie Mellon 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 early 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 Cutler's code? How do I go look at the Malik 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 CTOs, VP of engineering, SVPs do take step downs to work on small teams to learn how to build with these new tools right now. It seems a lot is changing. Absolutely. I think, yeah, The organizational sort of context of it all is changing that I think, yeah, 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 aspects of what makes successful products or successful technology that, yeah, there's plenty of opportunity.
“The way you want to do agentic coding is these little tasks that you give it. Make sure that it's truly done.”
I do wonder because in AI labs, it's pretty common and they're very open about sharing how software engineers, their job is becoming to allow the agent and the loop to do a better job. So what you just said.
“the real world is all about brownfield, and the brownfield requires stepwise hardening of systems, rethinking of the core primitives”
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 it. So maybe it's not all that different to how we're adopting it. But what is different about this? Yeah, it's a good one. Sometimes 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 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 joined the workforce at the cusp of when we went from You know, in fact, in college, I was mostly working on a VAX, right? You know, with a VT, you know, terminal. This is the mainframe. Yeah, or a mini, right? VMS systems, right? So, 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 two technologies were new frameworks and new runtimes. It was not like the concept count radically changed. Even the mobile thing was sort of a miniaturized PC with new sort of 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. Every single part of... It's a little more of a change in how you think about, I think.
Then let's not forget, a new model could drop Whether it is a new frontier model or a new thing like decision engine, and it could change again. This is where I think we also have to be ready for the regime to completely be transformed 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 are at our best individually or as companies or as teams in managing change, I think is always important. But also being open-minded that this time is different is also good. Yeah. And also specifically for the biggest change for us software engineers and 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 earlier, it was hard to reconcile because we've never seen it with any technology change. It always added a new tool. How do you see people inside of Microsoft or outside who talk with reconcile this, who are really good engineers and they spend decades getting very good at coding, which is hard to do, right? Yeah, no, I think that that's right. So this is where what at least I am observing, and as I talk to even people who are both being great software engineers previously and now are prolific even in the new mode of production, is the same skills are still required.
It's just, again, that what you think about, the order of operations, the tool chain 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 about software engineering no longer being important. And I think that we're in that phase, right, 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's AI native who comes out. Like, in fact, it is funny, last GitHub universe, I went down and I met a bunch of college kids. And that was the first batch of college graduates who grew up in the AI era, right? Because they went to college when GitHub co-pilot, right? So they didn't know an IDE world pre-GitHub co-pilot. It was like stunning to me that that was sort of what's happening. So there is going to be people who are going to come at this or are 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 curiosity and is sort of learning the new ways to work are absolutely flourishing in it. And so one of the things 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. And so therefore, we want to make sure at Microsoft, you can both value the skill set you have and build a new skill set that the marketplace will value. I'm really liking how you're looking
“if you're in a team where you feel stuck professionally, you can vote with your feet.”
“They start with a token budget and a human, right?”
our season sponsor. Until recently, I was thinking of Linear as a tool that helps plan and organize the work, but it's much more than that. A lot of teams use Linear to ship changes from IDEA to merge in GitHub, all without leaving the tool. I sat down with Dima, 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 unsent draft comments, Linear warns you about these comments. But it didn't tell you where those comments were. A customer
complained and Dima decided to fix it. But for the fix, Dima did not open his ID. Instead, he opened up linear and described the issue to the agent. The agent built a fix, created 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. The top of the review has a summary and a diagram of the change and the pull 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 Dima is finished with all of this he just self-approves the code and it's merged into GitHub all 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. Check it out and get a six-month trial at linear.app.pragmatic. If 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, role-based access. None of it is what you set out to build, but the deal waits on it. WorkOS is the auth infrastructure that covers all of that so nobody on your team has to debug a customer's SAML setup. The newer version of the problem is agents.
Once an agent is acting inside of a customer's systems, the enterprise wants to So, what is it allowed to do? That's what WorkOS Airlock is for. It's intent-based access control. Most 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 calls, an AI judge evaluates a policy at runtime. You write these policies in plain English, and Cloud can pull the ones you already had out of Notion or Jira and import them for you. Your agents, whether that's Claude or Devin, connect through a single access point, and your MCP server sits 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 slash airlock to learn more. Now, we know when everyone is building their software factories or changing how they work, a big trend now 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.
A lot of them are spinning up their inference or companies like Revolut are buying their inference, they're buying their B200s and just launching the models on it. They all have their own reasons. But from your vantage point, Microsoft is also in the open model serving business with Foundry, for example how big is this trend and is this inevitable i'm asking myself are there parallels to the open source movement and now the open model yeah i think that there are for sure the parallel which is you know you know there's sort of going to be close models and there'll be competition amongst them and there will be sort of an alternative to obviously closed would be open Open gives you flexibility.
In fact, one of the ways I come at it is with just this, what does a frontier ecosystem look like, right? It can't be, if you just get past this thing called, there's going to be one frontier model to sort of have all of the world's knowledge, right? If you don't buy that, right, 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, if you go back to actually Hayek, his description of what an economy is, it's all about knowledge being distributed, right? In fact, his entire thing about pricing 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 essentially got both human capital and token capital. So if every firm is going to have both token capital and human capital, that token capital can be an open-weight model into which you've embedded your knowledge. That's one. It can also be in the context layer. You could even be using multiple closed models. In fact, collecting even the traces of your usage of those closed models. And then using those traces to maybe train a small model and an open-way model for yourself and keeping, you know, a combination.
And I look at even Copilot. In fact, whether it's GitHub Copilot or Copilot, they're all designed as these multi-model systems, 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 any one model because otherwise you have a complete lock-in. In fact, one of the things I think about is 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 do you own versus what is owned by the model? because the model will basically collapse into it, everything it sees. And so I think that this dynamic is getting better understood. 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, everybody needs to sort of have a real thing, architectural approach to figuring out how do they manage to control their own learning and knowledge and have firm differentiation. And in that context, open-weight models is one strategy that I think has legs.
You know, this is very interesting because I talk with a lot of mid-size tech company startups, and almost all of them are building the things that you just drop, their own harness, their own context layer, their own context graph, their own inference. And I thought initially, I'm like, well, this kind of makes sense because engineers love to build. And if you want to get AI engineering experience, the least risky thing is to build something internal. However, I never thought 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
“you should be able to replace any model and still hill climb on an eval that's your private eval.”
have control over your learning loop. As an engineer, 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 system 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, five to 10 years or something. But right now, this is the job. It's correct. And in fact, in fact, one of the things we have our conference Ignite 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 are 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 patterns of success become much clearer. Now, as an engineer, one thing that 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
“then you can invoke for some sub-agents a bunch of local models while you're using even the cloud models together.”
edge as also token factories in fact i kind of think of it as the unmetered intelligence yeah because you you don't need to worry about how much you're using it that's right i 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 assume that 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. There's the infrastructure there. And I'm like, huh, could there be a future where like other closed models? So 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. You know, to your point about the code flash models, getting them even all the way to a laptop is core priority for us. And we are excited. And that's kind of when I look at the code tab inside a copilot, which is even independent of GitHub copilot. Now, 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? First of all, through the partnership with OpenAI, we have the OpenAI models, and we are going to continue to use those OpenAI models. On top of that, what we are saying is, okay, let's sort of innovate in a new model family ground up, that truly... Ground up meaning you build it from zero.
First of all, I think we disclosed even the new MAI-2 base, which is the pre-training. And in fact, it's one of the best disclosures of a pre-training recipe and regime. So for those who are your listeners, go look at it. And people tell me from the outside that it's the most detailed descriptions of pre-training, because there are very few places where you can go to and get it. And it's completely clean room, so to speak, implementation of 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 on the frontier where you're both flop sufficient 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-agents can be using multiple model families. That's, I think, the way. This idea that, oh, I'm wedded to one model family may not be true if there are choices. Agents will be 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. I really like how you're setting out choice because as a developer, I love choice. It also helps leads 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 change maybe similar to this, maybe not as big, but not as quickly. You've been through multiple chain 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 an amazing software engineer in this new age of AI in a year or two?
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? It's not just about the headlines, and it's not just about... It's definitely not the headlines. You know, I love X and, you know, it's great, but it's just not about the hype cycle and that constant need to talk. But it's ironic in a podcast to say that. It's do more, learn more, and hands-on. 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 fresh college grad the more you can take the time to learn to do, be in teams that are doing, the better it is versus I think anything else. Because at times, at some level, I sometimes think about it. And when am I at my best, it's when I'm in the school mode. So let's get back to school. Satya, thank you so much. Thank you so much.
It's such a pleasure. Well, this was an awesome experience, and we really covered a ton of ground with Satya. 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 do 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. And I assume that it's because engineers just like to build with novel technology.
But Satya pointed out that it's actually suicidal strategy for a company to not own their AI infra and their AI intellectual property. In fact, he thinks that every successful company will build out this own 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. Finally, software engineering roles and expectations are also changing. Satya 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 years ago. Satya expects to see a lot of reorgs inside of Microsoft for teams to 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 is probably not just true for Microsoft, but it will be true for a lot of other tech companies. In the end, I walked away being more energized from this conversation.
Yes, the industry is changing, but what I understood is how Satya sees that a lot of software engineering work is waiting to happen with these new AI-related primitives like decision models and some of the primitives that have not been invented. And us engineers 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 concepts like evals, inference engineering, loop engineering, and more. If you've enjoyed this podcast, please do subscribe on our favorite podcast platform and 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 one.


