Microsoft CTO Kevin Scott on how AI can save the web, not destroy it
01:14:38 Share

Support for this show comes from Atio. Atio is an AI-native CRM built for the next era of companies. Its powerful data structure adapts to your business model, syncs in all your contacts in minutes, and enriches your business with actionable data. Atio also allows you to create email sequences, real-time reports, and powerful automations, all to help you build what matters, your company.

Join industry leaders like Flatfile, Replicate, Modal, and more. You can go to addio.com slash decoder and you'll get 15% off your first year. That's A-T-T-I-O dot com slash decoder.

With HubSpot's suite of AI-powered tools, you can get more done way faster. Speed up your lead generation and create attention-grabbing, lead-driving, quota-crushing campaigns in an instant. Get started today at HubSpot.com slash AI.

Support for this show comes from Pure Leaf Iced Tea. When you find yourself in the afternoon slump, you need the right thing to make you bounce back. You need Pure Leaf Iced Tea. It's real brewed tea made in a variety of bold flavors with just the right amount of naturally occurring caffeine. You're left feeling refreshed and revitalized so you can be ready to take on what's next. The next time you need to hit the reset button, grab a Pure Leaf Iced Tea. Time for a tea break? Time for a Pure Leaf.

Hello and welcome to Decoder. I'm Neil Apatow, editor-in-chief of The Verge, and Decoder is my show about big ideas and other problems. Today I'm talking with Kevin Scott, the chief technology officer of Microsoft and one of the company's AI leaders.

This is Kevin's third time on Decoder, and he's one of my favorite guests. He thinks a lot about the relationship between technology and art and culture, and he's unusually willing to dive into the weeds of it all. Kevin joined the show today during Microsoft's Build Conference to talk about the future of search. Microsoft just announced an open-source tool for websites to integrate AI-powered natural language search with just a little bit of effort, in a way that actually lets them run whatever models they want and keep control of their data.

It's neat stuff. I saw some demos before Kevin and I chatted, and the improvements over the bad local search on most sites was obvious. But the goal here isn't just to improve the local search feature on a bunch of disconnected websites. It's to rethink how search even works in a world where AI is more broadly distributed.

Think about it this way. Right now, building a search engine requires you to go and index all of the web pages on the internet, and then keep that index constantly updated, which is an enormous recurring cost. That cost is why there are really only two main search indexes. Google's, obviously, and Microsoft's Bing index, which powers most of the alternative search engines like DuckDuckGo that you might be familiar with.

Those centralized search indexes are also the underpinnings of our current AI search tools, like the search built into ChatGPT and Bing or Google's AI Overviews. But if all those websites suddenly have their own powerful AI-powered natural language search tools,

Well, you might not need that big central index. All you really need is a standard that lets you go ask a bunch of websites if they can answer your questions, which would dramatically bring down the cost of search overall, and maybe let more competitors into the market. So Microsoft's local search project is built on just such a standard. It's called Model Context Protocol, or MCP, and it allows AI agents to go and interact with databases and services in a controlled and predictable way.

not just clicking around the websites, which is what basically all the agent products do right now. MCP was initially developed by Anthropic, and the rest of the industry, including Google, is starting to support it.

There's a long way to go here, but the first step is just enabling MCP on more sites, which is why Microsoft is making MCP-powered local search cheap and easy to implement. It's very cool to think about what the future of MCP-powered agentic search might look like. Maybe there will be more competition in search overall. Maybe websites will get more traffic. Maybe these businesses will be more sustainable. But there are some obvious complexities.

Starting with why anyone would allow a bunch of agents to simply use their services in this way and how anyone intends to make money doing it. I asked Kevin about this and we spent some time really thinking about how the future of the web stays sustainable for all the people actually making the content out there.

I could obviously talk about that forever, but since I had the time with Kevin, I also made sure to ask about Microsoft's unique and often confusing relationship with OpenAI, how he's feeling about AI's capabilities scaling up over time, and whether his thoughts on AI as a creative tool have evolved as lawsuits and controversies play out in the creative community. Kevin is an author himself, after all. He wrote a book about AI back in 2020 with a foreword by a relative unknown at the time named JD Vance.

So I asked Kevin, how would he feel if someone stood outside the bookstore and simply summarized the book for anyone who seemed interested in buying it? I think you're going to find that back and forth pretty interesting. Okay, Microsoft CTO Kevin Scott, here we go. ♪

Kevin Scott, you are the Chief Technology Officer at Microsoft. Welcome back to Decoder. Thank you so much for having me. If you count our sort of early pilots of Decoder on the VergeCast feed, you are, I believe, our first four-time guest. Wow, really? Yeah, and always one of my favorite conversations, so I'm excited for this one. Microsoft has some news about search in the web, which is just in the strike zone for me on the show because it's

I see that changing so enormously. I want to talk to you about what we've learned now that AI has sort of entered its second era here. And then I want to talk to you about where it's going. Just the usual stuff. Small potatoes. Yeah. Let's start with the news. Microsoft just today, as people are listening to this, announced on stage at the Build Conference a new approach to searching websites locally. Tell us what's going on there. Yeah, look, I would actually frame it a little bit.

bit less is search. And like what I've been spending a bunch of time thinking about is we have this hypothesis and I think it keeps getting borne out that you're going to have

software enabled by all of these powerful new AI models that have been built over the past handful of years. And agents need to be able to do stuff on behalf of users. We're less constrained and like you sort of called it the

Second act, like the way that we talk about it internally as we sort of describe where we are right now is the middle innings. And so we're not reasoning constrained anymore. We're sort of utility constrained, I think, in these agentic ways.

pieces of software that we're trying to build. And- Wait, can you, utility constrained? Yeah. There's a lot of ways to unpack what that means. Yep. Maybe the harshest is, it doesn't quite work yet. Maybe the less harsh is, there's not product market fit yet. Yeah.

What do you think that means? Well, look, I think it kind of depends. Like, if you sort of look at software development, like, there's clearly product-market fit. Like, this stuff has become just an indispensable way that people are building software, like these software development agents that...

we built. And there, you know, I think it's just sort of the early prospector on the frontier of what agentic software can do. Unsurprising, like developers usually build things to make their lives easier before they build things to make everybody else's lives nominally easier.

So we are a little bit further ahead there. Some of the things that we're learning about how we're going to need to make general purpose agents more useful for things other than software development, we've discovered in building these SWE agents. And so one of those things is agents need to be able to access sources of information. They need to be able to take action on behalf of users by themselves.

you know, making changes in state and systems. You have things like book a hotel room or put something on my calendar. And I think the way that you really want all of these things to happen is like, you know, just sort of open up

protocols where you have real interoperability across the whole landscape of agents, where you can have everybody who has a service or everybody who has a piece of content that they want to make agent accessible has a way to sort of say like, Hey, here I am.

And then agents, no matter who's building them, have a way to go connect to that content and services. You know, like the way that we talked about it on stage at Build is the agentic web. So what must exist in this world where we have lots and lots of agents doing things on behalf of users, you know,

that are the moral equivalent to the things that had to emerge when the web was under development 20, 30 years ago. So there's the agentic web. We've been talking a lot about that show. I talked to your colleague, Mustafa Suleiman, about the agentic web and building agents. And I've been calling it the DoorDash problem. I don't have a better name for it. I feel bad for the people at DoorDash because I haven't asked for their permission. But I keep calling it the DoorDash problem, where

Okay, I want to get a sandwich. So I talk to Bing or ChatGPT or whoever, and I say, go get me a sandwich. Alexa Plus, go get me a sandwich. And then it goes out onto the web. And right now, most of the agentic products will literally open a website, try to scan the website, and then click around on the website, and then order me a sandwich. And most of those companies are like, don't do that.

Right? Like their posture is, we don't want you to do that. We're going to block you. And maybe if you're small enough, we'll let you do it. But we need to have business terms that makes it so that you can just take our capability and put it in your product in this way. That problem has to be solved. I'm curious for how you would solve that problem. It sounds like you're operating at kind of just one step of abstraction beyond that, which is assuming we solve the business problem, how can we make the my agent can talk to DoorDash much easier problem?

Because clicking around their website has never seemed like a good solution. Yeah, it is brittle. And look, I think actually solving the business model problem goes hand in hand with solving the technology problem. It's not just about figuring out a technical way to do something. It's about getting all of the incentives in the ecosystem aligned the right way where good things are happening for everyone. Yeah.

you have a business and you want your business to be able to transact with users via their agent, that has to make good business sense in order for you to be willing for that to happen at all. You can't just hack your way around that and expect it to be a durable thing. Even if you can temporarily figure out some kind of

technical magic to get around the, the brittleness of the actual technology. Like you also have to get rid of the brittleness in the business model. So that is the piece that on the web right now, it seems most under threat, right? That the underlying business dynamics of I start a website and,

I put in a bunch of schema that allows search engines to read my website and surface my content across different distribution. I might add an RSS feed, which is standardized distribution that everyone uses and agrees on the standard. There's lots of ways to do this. But I make a website. I open myself up to distribution on different surfaces. What I will get in return for that is

is not necessarily money, almost in every case, not money. What I will get is web visitors to my website and then I will monetize them however I choose to. I'm signing a subscription, I'm putting up display ads, whatever it is. That's broken, right? As more and more of the answers appear directly in particular, the AI-based search products, traffic websites has generally dropped. We see this over and over again.

What's going to replace that in the agentic era where we've created new schema for agents to come and talk to my website and receive some answers? What's going to make that worth it? Well, I think one of the things that we're trying to do right now with some of the things that we're announcing and that we're trying to do in an open way is –

you will have technical mechanisms for agents to be able to access people's websites. But like the protocols themselves will allow you to decide what it is you want to make available and how. And so like, you know, if you just look at MCP, which is a thing, like a super awesome thing,

protocol that Anthropic developed that we're doing a whole bunch of work with Anthropic to support. And I know a ton of people in the ecosystem, so OpenAI is working with them. And a bunch of folks have sort of latched on to MCP as the moral equivalent of HTTP for the agentic web. MCP doesn't have an opinion one way or the other about

what a content or a service provider ought to make available via MCP or like what the business model ought to be for that access. And so I think one of the nice things about that is it gives people who have content or services a way to decide what the new business models will look like. So is it that, you know, an MCP endpoint is usable inside of your agent if the user has a subscription to your website?

Is there going to be some kind of new advertising model where you get to give away some stuff for free and like you have used that to drive a bunch of agent activity into your website and maybe there's some advertising that goes on that helps with the distribution and then like there are transactions that are being made where you can sort of price by conversion. I don't know exactly what the business model is going to be, but I do know that the thing that you –

really are going to want to have is agency on the part of the content and service providers so that they get to decide, you know, what they make available and what the business model is for the things that they're doing.

So MCP is Model Context Protocol for the listener. Correct. That is – it's a nascent standard, I would say. Yeah. You show up to my website or something, my service, and I tell you what you can do. It's very much the evolution of robots.txt from what I gather, right? It's more complicated. It's more sophisticated. Yeah. But the idea is very much the same, right? You show up and I tell you what you can do and you can't do.

Can Microsoft just horsepower this into existence? Can Microsoft and OpenAthropic just horsepower into existence? Or is there some upside for everyone else to participate? Well, look, I think... So look, if you're a developer, there's a ton of upside. So this reminds me a lot of how things felt when I was a younger developer when the internet was exploding into existence a few decades back, where...

What I wanted was a set of permissionless mechanisms where I could just go put something up on the web and then I could have other people access it in interesting ways. Yeah. And so, look, I do think that

This is MCP and the thing that we're doing that maybe we'll talk about in a minute on top of MCP, which is this interesting thing called In a Web, are...

you know, a set of open protocols and a bunch of code that helps implement them that let you, without having to seek permission from anyone, decide, you know, what you want to make available on this agentic web so that, like, things that you're doing are accessible by agents. And, you know,

You know, look, I think it's also, you're right when you sort of called it a nascent protocol. Like the thing that's interesting is like just sort of how fast the ramp is right now. Like just how quickly everybody is snapping to this as a way to, you know, make your things agent accessible. And so, you know, the attractive thing, I think, if you are a content or service provider is more and more of the,

user activity gets anchored in agents, that's going to be more and more of where user transactions and user attention gets funneled through. You're just going to want to be agent visible in what you're doing so that you have access to an audience.

you really start to get into this mode where agents are doing things asynchronously for you. So like a lot of what happens right now is, you know, in the current web model,

is everything happens synchronously. So you're sort of sitting there, like I'm staring at a browser right now, like I may have a tool that I wanna go buy on somebody's Shopify storefront, like my attention is focused on this particular task, like I complete the transaction and then I move on to the next thing.

The interesting thing with agents is things are going to start happening asynchronously where you're going to give an agent a task and it's going to go do all of this stuff while your attention is elsewhere. And so that's a super interesting thing where I think there's going to be a bunch of opportunity that doesn't exist right now because in the limit, I only have so much of my attention that I have to go...

spend on websites. If I had a bunch of agents off doing a bunch of research for me and helping me think about my summer vacation, or I've got this crazy project I'm doing in my shop, like building a kiln for all this random pottery stuff I'm doing, if it can help me get a little bit further along so that when I got my attention left to give, I can

go take action immediately, like maybe buy stuff or the attention is just much higher quality. I think that's interesting as hell for

folks who are trying to do business on the web. Yeah. Let's talk about the NL Search project that you're doing and how it connects this larger vision. So I saw a brief demo of it. It was very cool, right? It's a low cost, very easy way to integrate natural language search into a website. One of the demos I saw was TripAdvisor. I was told that the TripAdvisor team looked at this last Wednesday and they were demoing it on a Tuesday to their company leadership. That's cool.

At a low cost, it runs with all the models. You can run it with DeepSeq. You can run it with 4.0 Mini. Great. That's what gives you the MCP capability.

Right? So you've run this tool with your website. Yep. You've exposed a benefit to your users. Yep. Here's some natural language search as expressed in a chatbot or a custom interface if you want to build one. But then now you've added this MCP schema to your website that lets a Microsoft co-pilot agent show up and interact with your website in some structured way that you can control. Yep. All of that is very cool. And I understand how the incentives line up.

Isn't that just a bunch of APIs by a different name? There's just a part of me that makes that very reductive and very small. Yeah. Look, I think that's actually not a bad thing. It's actually a super good thing that it's really a simple set of protocols that are enabling a bunch of super rich behavior that, again, going back to the fundamental premise, if you want agents to actually be very, very useful where they can

live up to their name. So an agent ought to be a piece of software to which you can delegate increasingly complicated tasks over time. In order for those tasks to become increasingly complicated, they have to be able to do work. And like the best way for them to be able to do work is like you need

ubiquity of content and services available. You need everyone's sort of incentives, both hurdles to adoption and, you know, like business model and, you know, just sort of economics of everything to make sense where you get fairly broad adoption. So simplicity is a like a definite feature. We need to take a quick break. We'll be right back.

Support for Decoder comes from Vanta. No matter how big your company is, protecting customer data is essential. And once that data is protected, you have to prove it. That can eat up a ton of valuable time, and it's not something you ever want to cut corners on.

That's where Vanta comes in. Vanta helps companies get SOC 2, ISO 27001, and HIPAA compliant. And it automates up to 90% of the process, saving you time and money and a lot of potential headaches. If you're a startup founder or security professional, you know what I'm talking about. Collecting evidence and doing ongoing controls monitoring used to be incredibly time consuming.

Vanta changes all of that. Now, a ton of tasks that used to be done manually can be automated, and it gives you live visibility into your progress. Compliance is Vanta's middle name, or it would be if Vanta had a middle name, which it doesn't because it's a company. But maybe it should, though. Go to vanta.com slash decoder to meet with a Vanta expert about your business needs. That's vanta.com slash decoder.

Support for Decoder comes from LinkedIn. There are those people who seem to know everyone in the world. They have someone for game tickets. They have someone for nice suits. They have eight people for IT. They know actuaries and ballerinas and the best fishmongers. Unfortunately, if you don't know everyone in the world, recruiting talented people quickly can be a challenge.

That's where LinkedIn comes in. LinkedIn wants you to feel confident that you're hiring the best. Based on LinkedIn's data, 72% of small and medium-sized businesses using LinkedIn say that LinkedIn helps them find high-quality candidates. LinkedIn says they make it easy to post your job for free, share it with your network, and get qualified candidates that you can manage all in one place. And LinkedIn's new features can help you write job descriptions,

and then quickly get your job in front of the right people with deep candidate insights. Find out why more than two and a half million small businesses use LinkedIn for hiring today. Find your next great hire on LinkedIn. Post your job for free at linkedin.com slash partner. That's linkedin.com slash partner to post your job for free. Terms and conditions apply.

Support for Decoder comes from agency. With AIs, LLMs, and new systems being developed every day, there's a lot to be excited about in technology these days. But there's a fundamental gap in how all these tools can complement each other. While single agents can handle specific tasks, we have no standardized infrastructure for these agents to discover, communicate with, and work alongside each other. That's where agency comes in. Spelled, Decoder.

A-G-N-T-C-Y. Agency is an open source collective building the internet of agents, a global collaboration layer where AI agents can work together. Agency wants to connect systems across vendors and frameworks, solving the biggest problems of discovery, interoperability, and scalability for enterprises. With contributors like Cisco, Crew AI, Langchain, and MongoDB, Agency wants to break down the silos and build the future of interoperable AI.

We're back with Microsoft CTO Kevin Scott. Before the break, we were talking about the company's big ambitions to help kickstart the so-called agentic web, which will require a lot of different parties to rally around early standards like MCP or Model Context Protocol.

If that happens, the entire structure of the web could change. So now I really wanted to stress test some of those ideas with the hurdles and constraints facing the industry today. And in particular, one big player on the web today, Google.

Let me ask you about the elephant in the room when it comes to the web and building new capabilities on the web. It's obviously Google. Right now, the web is organized around Google's priorities, its needs, its traffic whims. There's an entire class of SEO consultants that wave sparklers at the sky to collect traffic for you. It's great. We love it. That's obviously changing, right? As Google keeps more of the traffic for itself or thinks differently about training data, all this stuff is changing.

The trade here is make your website more agentic. And then MCP as a protocol will allow you to build some new business models on it. The problem as I see it is that the traffic to the web is in precipitous decline as Google referrals go into decline. How do you fix that problem so that everyone is incentivized to keep building on the web? Yeah, I don't know. Like, honestly, one of the ways that I could answer it is like I...

I have, uh, just so that I can see for myself, uh, I've set up a Shopify storefront. Uh, so like I have a tiny little business that I operate on the side just so I can feel what it's like to be a web-based business owner. And what, how does that feel? Uh, well, look, it's, it's sort of interesting. Like I, uh,

It's very interesting. The dance that you have to do to get traffic driven to your business and the amount of energy that you have to spend tending that traffic stream. Funny enough, most of my referrals that are coming into my storefront are not from

are not from Google. Like most of it comes in through social media and social media advertising, which is really different than what I was expecting. So I don't know that I've gotten any interesting referrals and certainly no conversions coming in through organic search. What about Bing? Like no organic search.

Um, so like, you know, this isn't me trying to, you know, it's good to know that you don't have a thumb on the scale. That's, that's really what I was asking. Yeah, no, I, I certainly don't. Uh, in fact, I, I think most people at Microsoft are hearing for the first time that like Kevin Scott has a Shopify storefront. Uh, so yeah, definitely do not have any kind of thumb on the scale. The thing that I would love as a website operator is I would love to have a way to

not to spend so much of my time worried about traffic referrals, period. I would love to spend more of my time

building up an authentic relationship with people who might be interested in my products and services and have a way to curate that relationship. I would love for there to be a thing like InnoWeb where people could buy my products and learn more about the services that I have to offer and then have a way where they become

loyal customers that I actually have a relationship with the same way that I would have if I had a physical storefront and I had foot traffic coming into my store. Social media at least feels to me a little bit more like what I just described, where I've done a very deliberate job trying to curate a social media audience to

just have followers who might be interested in the things that I'm doing as a maker, for instance. Well, what's interesting about that, I've asked a lot of people over the past several years about

why would anybody start a website? And the frame for me is when we started The Verge, the only thing we were ever going to start was a website. We're a bunch of people. We want to talk about technology. In 2011, we were going to start a website. We weren't even going to start a YouTube channel. That came later that people started doing YouTube channels at scale. At the time that we started, it was you're going to start a big website. Now I think about, okay, if I had 11 friends who wanted to start a technology product with me, we would start a TikTok website.

There's no chance we would be like, we have to set up a giant website and have all these dependencies. We would start a YouTube channel.

And I, I asked people, why would anyone start a website now? And the answer almost universally is to do e-commerce, right? It's to do transactions outside of platform rules or platform taxes. It's to, to send people somewhere else to validate that you are a commercial entity of some kind and then do a transaction. And that is the point of the web. The other point of the web, as far as I can tell, is that it has become the dominant application platform on desktop.

And whether that's expressed through, I don't know, Electron or whether it's expressed through the actual web itself in a browser, it's the application layer. And so I get why you'd want to say, okay, we're going to do agents that are going to go traverse the open application layer that exists and use those tools. I still am lost at like, well, if I just want to communicate to people, I'm going to go to some closed platform.

And then we just kind of enter a place where even the AI tools just have less information to work with.

Because everybody talking about what to buy might be on TikTok, and then all the stuff to buy is on the web. And that's the loop that I can't quite close. I think this is one of the things that could potentially happen with things like MCP and InnoWeb. If the way that people want to do research or they want to transact business is via their agents, and that's where the intent is,

lives and where the user desire originates, then you're going to want to have some mechanism where you can go connect to that. So even if, let's say, you and your 11 friends in 2025 are going to go start a TikTok channel to talk about tech.

If one of the things that you're doing there is doing a bunch of reviews about tech products or tech sites and you want to reach an audience and the audience is sitting inside of Copilot or ChatGPT or something like that, you're going to want those agentic pieces of software to have some way where they can reach into...

your media channel so that you're sort of exposing that audience to what you're putting out there. And like, you know, web might be a good way to do that where, you know, maybe you're not offering up everything, but you're offering, you know, kind of what, what's happened with search, like teasers, like snippets, like things where it's like, okay,

you ask your agent to like, hey, I'm trying to buy like a new phone. Here's kind of what I think I want. Can you go find me some sources of information about this? And if you have from your TikTok channel a way to let your agent know what your content is, like maybe that's the referral traffic back into TikTok from the agents. Like, hey, go watch this video. Like, this is super interesting. Yeah, I'm really curious to see if the

big platforms enable themselves to be searched or acted on by agents in the way that they somewhat had to allow themselves to be searched by the big search engines, right? There wasn't a choice. I think maybe the biggest platform that recused itself from search for a minute was Facebook.

But Instagram is still searchable, right? Like there was a trade where you wanted to be exposed and to be found on these tools. So everyone sort of opened up. The dynamics of how you will open up to agents, I think, for a variety of reasons, many of which make sense, are just not clear. Like why would we do this when we could build our own agent is still...

We're still in those early stages. To put it in search terms, are you going to do vertical search or big horizontal search? And horizontal search sort of totally won. Yeah, and it's just sort of hard to say what exactly might happen here. I think it will largely get decided by users. So one of the things that's going to happen is users will just decide what it is they're going to tolerate. So if using an agent to help you sort out your...

life and what you're doing becomes such a huge preference that people have, things that aren't connected to the agent will kind of turn invisible for folks. Like you'll just sort of think, oh, well, you know, X isn't reachable by my agent. Like, you know, maybe X is kind of broken and I'll find another way to do that. And,

I think the thing that you want is the market kind of figures out what it wants is you want as many open protocols as humanly possible so that people can make those late binding decisions about like, okay, like this is what the users have chosen. Like they're expressing their preferences, like at least like have this thing be open so that I can opt into it. Yeah. When the preference is clear. Yeah. I'm dying to know how this plays out. I can see a bunch of,

sites like TripAdvisor and others that would really want this kind of distribution. Obviously, we need to build the front end tools, the aggregators that say, okay, here's your agent that's going to go out. Do you think, having looked at the agents that have been demoed so far or announced and then not shipped or announced and shipped to five people, do you think that initiating MCP across the web like this is a necessary condition for agentic systems to operate? Because none of them work so far.

Yeah, I think something like this is really kind of necessary. I mean, I remember back when I was working in the early days of mobile on advertising, and the reason that I worked on mobile advertising is I wanted to figure out a way to help

who were building mobile apps and services to figure out distribution and to figure out how they monetize themselves. And before things like AdMob came into existence, the only way that you could get distribution is you went and cut a BD deal with a mobile phone company. And they would decide whether or not they were going to give you placement on... At the time, it was WML and it was placement on their deck. And it's just kind of a barter

barbaric arrangement. It made perfect sense back then, but if you look at what happened with the technology and fast forward through time, it's like, yeah, why would anyone choose that? And so I think there's a little bit of that dynamic right now where people are absolutely finding utility in these agents, even as constrained as they are right now. And so in places like software development where you need to have

a much narrower range of things that the agents can actuate, like where you build some completeness there, like, oh my God, the adoption is great and people love what these things are doing and there's a ton of competition and it's just really transforming how software development is working. So what I think we're going to see, and this is me, Kevin Scott, the optimist, is if you had...

a real complete agentic web where MCP was the lingua franca of this agentic web, like HTTP was. Everybody can just stand up an HTTP server and start serving HTML.

and they get to decide what the HTML payload is, you're going to see this really interesting organic unfolding of what's possible. I don't know what the moral equivalent of the Amazon.com or the early winners in the web are where when you get enough of that plumbing hooked up that

things are going to be super useful. But yeah, I mean, like, I think that some of this protocol stuff has to happen before you get to full utility. It's why MCP is interesting. You know, we kind of think about it in a web is kind of like the HTML layer for, yeah, so it is a thing that,

That lets you not have to go do a really tremendous amount of low-level work in order to be able to sort of plumb your stuff up to the agentic web. Yeah. You know, there's – the parallel that comes to mind here is –

Apple's attempt to build an agentic Siri, which is built on a thing in their operating system called App Intents, where iOS apps will expose themselves to Siri in some way and let Siri take action inside those apps. There are rough parallels here. Obviously, MCP is more open standard. It's more nascent. App Intents has run into the same business model problems. Like if you're an app developer on iOS...

Why would you let Siri use the app and not the user so you can upsell them or sell them in an app subscription? Whatever you're going to do. That's one parallel. The other parallel is Alexa Plus, which I joked earlier has launched to some people. No one knows who they are. Google has some agentic ideas. There's computer use from Anthropic, OpenAI. None of it's worked yet. Have you seen anything that says this is definitely going to work?

To answer your question very specifically, outside of software engineering and outside of demos, I haven't concretely seen the thing that works. And I'll just sort of ground it even more. So if I look at my everyday life and I'm sort of looking at how I use these things,

Outside of software development, there's just not an awful lot I can go – that I'm choosing, Kevin Scott is choosing to go delegate a bunch of stuff to this agent to do on my behalf.

But I can kind of smell it with MCP. And I do think it actually has to be open. I think it's kind of hard to do this in a vertically integrated way. The other question I want to ask, I'm really putting the media training to use here. I'm going to ask you about Google. Google had an opportunity to achieve some of the success it did because Microsoft was in the throes of antitrust pressure.

Microsoft bundling Internet Explorer, putting pressure on Netscape that led to whatever amount of legal trouble Google was able to swoop in. It created the application layer. It was able to put Chrome on Windows. Everybody has a story. That really created the antitrust pressure on Microsoft, really created the opportunities for Google to succeed. Here we are, it's many years later, it's decades later. There's a lot of antitrust pressure on Google, particularly in regards to how it controls the web, both in the advertising layer and the search layer.

there's the suggestion that the government make Google divest Chrome. This is a lot of antitrust distraction. And here I am talking to a Microsoft executive about new ideas for the web and new standards for the web. Do you see the opportunity as the same, that there's space because Google is being distracted? Uh,

Look, I think the opportunity is the moment that we're in. The technology itself is just ready for some of this stuff to happen. I mean, this demo that you saw that Guha gave you.

You couldn't have done it two or three years ago just because the technology wasn't ready. It would have been just impossibly difficult to do. And there was no way that TripAdvisor could see a thing on a Tuesday and be demoing it on their own data on a Wednesday. That is entirely a function about technology maturity. So I don't know.

what's going to happen with anything that the government is doing with any of the other tech companies. But I think part of what's happening right now is you just have...

a new set of technologies that are capable of a new set of things. And you've got a bunch of big tech companies and small entrepreneurs who see the possibility in the thing. And like you, like the thing that I want is I want to see as much energy in this ecosystem becoming more mature as humanly possible. And again, you know, like I, my, my pattern matching is, you know, goes back to,

The last time that I was a happy young developer, which was like when the Internet was emerging and just that sensation that you can have when.

A bunch of hard things just became easy and a bunch of the protocols are open. Yeah. And you don't have to ask anyone's permission to go try something wild. Like that's when interesting things happen. But let me just put a little bit of pressure on this. Then I want to actually broaden out and just talk about AI generally. You know, if you had showed up and said, okay, here's a new standard for accessing the web and structuring websites from Microsoft. Yep. Two years ago.

Everyone said, great, we're going to wait for Google's riff on this or for Google to adopt the standard. Google's under a lot of pressure. A lot of trust has been erased from Google. You now do have the opportunity for OpenAI and Anthropic and Microsoft to show up with a new standard and to feel that there might be real adoption here and that Google can't show up with its own standard tomorrow and take the wind out of your sails.

That has to be true for you, right? You feel that. Because the comparison I would make is, I don't know, in the late 90s or early 2000s, someone would announce a new standard and Microsoft would show up with like a proprietary Windows riff on that standard and the other thing would go away. And that was part of the problem. Do you see that reflection right now? Yeah, I don't know. Like I'm not trying to be, like sometimes I am trying to be evasive. Like I'm not trying to be evasive here. Like I...

Sometimes it feels to me as an engineer, like certain things are technically inevitable. Like I've had a bunch of conversations about, you know, with folks about MCP, like inside of Microsoft, where it's like, oh, you know, like, you know, this isn't exactly what we would have chosen. And I'm like, yeah, but it kind of doesn't matter. Like sometimes...

There is a real problem in an ecosystem where the simplest solution that everybody can choose to adopt is the winner because we all win because ubiquity is the thing that really matters. And it feels like we've got a bunch of those opportunities right now. And so the thing that I think...

That's really beneficial is some of them have become such simple things where you actually don't need a multi trillion dollar company to go do an enormous amount of work in order to create the conditions for adoption to happen very quickly. In some sense, you know, with with MCP and NL Web,

you actually don't need a big tech company pushing for it. Like we're just sort of a voice out here saying like, hey, here's this interesting thing. It's open, like, you know, go do with it what you will. And like, that's all I can do. Like I, you know, I don't have the ability, you know, in terms of open protocols to tell anybody to do anything. Like we will shine a light on it and hope good things will happen. We need to take another quick break. We'll be right back.

That was so sus.

Nobody wants this. For your Emmy consideration in all categories, including Outstanding Comedy Series. Support for the show comes from Mercury. What if banking did more? Because to you, it's more than an invoice. It's your hard work becoming revenue. It's more than a wire. It's payroll for your team.

It's more than a deposit. It's landing your fundraise. The truth is, banking can do more. Mercury brings all the ways you use money into a single product that feels extraordinary to use. Visit mercury.com to join over 200,000 entrepreneurs who use Mercury to do more for their business. Mercury, banking that does more. Support for the show comes from Mercury.

What if banking did more? Because to you, it's more than an invoice. It's your hard work becoming revenue. It's more than a wire. It's payroll for your team. It's more than a deposit. It's landing your fundraise. The truth is, banking can do more.

Mercury brings all the ways you use money into a single product that feels extraordinary to use. Visit mercury.com to join over 200,000 entrepreneurs who use Mercury to do more for their business. Mercury, banking that does more. We're back with Microsoft CTO Kevin Scott. Before the break, we were discussing some of the practical implications of the agentic web and where it might run into competing pressures from platforms like Google Search.

But I wanted to make sure I asked some bigger questions about AI, in particular Microsoft's OpenAI partnership, which Kevin was instrumental in building. I also wanted to know how Kevin's thoughts on copyright and compensation in the creative community have evolved since we last talked about these issues a year and a half ago. Let's talk about the AI industry kind of broadly. You described it as middle innings. I'm describing it as sort of act two. This technology exists. Everyone has used it. We've all played with a chatbot.

Some reporters have had the chatbot ask them to leave their wives. I will never stop making this joke. Broadly, what did you get right in your initial bets and what did you get wrong? What surprised you? Well, look, I think we accurately spotted the scaling law trends with the reasoning power of foundation models. I think we've been more right than wrong with our conviction that those models

trends will continue to play out. I think we still have a tremendous amount of progress to be made, like increasing the reasoning power of models, but they kind of... I don't want to make light of how difficult it will actually be to continue the scaling, but they feel like a set of reasonably solvable problems if you've got the right resources and focus.

I think the thing that's sort of hard right now is I feel like we've got this capability overhang with models where the models are actually capable of a lot more than what they're being used for. And so I...

Even inside of Microsoft, I maybe overestimated how quickly people were going to just lean all the way into the platform capabilities of the basic AI models. So I think we're a little bit behind right now on product. And by we, I don't mean Microsoft. I mean everybody other than Microsoft.

you know, this rapid progress we're seeing in software development tools. So I think there are things in healthcare that could be a lot better than they are right now. Like, you know, a bunch of things are inhibited by some basic plumbing stuff, which is, you know, the topic of this particular conversation. But like a lot of it is just, you know, we, we just sort of need a lot more companies to get created and a lot more products to get made just to use what's already possible with these models. Yeah.

And a little bit of it, too, is I have this conversation over and over and over again. I was at a gathering with developers late last week. There's this interesting conservatism that is especially unhelpful with exponentially improving development.

And it doesn't even look like conservatism if you didn't have the exponentially improving platform. It's like somebody will sort of look at a thing and say, ah, you know, like this is a little bit too expensive for me to use for this particular thing that I'm, or particular problem I'm trying to solve. Or it's like marginally useful right now. So, you know, it's gets things right about 30% of the time, you know, but like that's only marginally useful. And,

And, you know, and then it's like, OK, like I'm just going to pause and wait. And that may be the right thing to do. But like the wait part of the pause and wait is too long in many cases right now.

Where the next time people go in and sample to see if it's gotten cheaper or more capable, it's already sort of raced past where it needed to be. And then you're just too late trying to get your thing to market. And so I think that's the thing that I'm seeing over and over again, where we collectively are making mistakes, where our pattern matching isn't as good as it could be. Right. You're saying that you should envision the products even if they're not 100% great yet. Yeah. Yeah.

It's interesting you mention that because you were one of the architects of the OpenAI relationship at Microsoft. A couple of years ago, you were on the show. I asked you about that relationship and where it came from. And you described that relationship in terms of platforms. Yep. Right? That Microsoft was a platform company. Obviously, Azure is a massive platform. And you said OpenAI is aligned with you on the platform vision. And we wanted to structure a partnership together so we could go build the platform together.

Yep. Things have changed in two years, I would say. The companies have pulled apart a little bit, maybe a lot. I watched the Senate hearing on AI the other day. I noticed Brad Smith from Microsoft and Sam Altman from OpenAI. Yep. Opposite ends of the table. OpenAI has become much more of a consumer company, right? They're obviously trying to go make big consumer products, not platform products. Anthropic is much more of a platform company, I would say, than OpenAI is today. Yeah.

How do you see that relationship now? Is it fully decoupled? Are you working together still? Are you still trying to build a platform? Yeah, look, I still spend a huge amount of my time on open AI things. And like there's a huge amount of technical stuff, you know, just sort of me as an engineer, like we, you know, we're building big computing systems together, like they're gigantic Azure customer, like their workload is like,

really, you know,

non-trivial part of our platform, particularly in terms of AI compute. Yeah, I mean, so we're working with them all the time trying to make sure that we are building things for them that they need and still just a ton of work that we're doing together across the board. Just from everything from how do we optimize the infrastructure that we're building to how do we take...

these models that we're training and get them optimized so that they can actually be platform components all the way to, like we still operate a joint deployment safety board where we help each other assure that the things that we're releasing to the public have gone through a rigorous responsible AI review before they launch. So yeah, just a ton of work that we're doing together.

If you listen carefully, there's a qualitative difference in what you said from before, right? There are huge Azure customers with a big workload. Of course, like everybody works with their biggest customers closely. Yep.

Previously, it was we're interdependent and their models are powering every co-pilot across the company. It very much sounds like Microsoft has moved OpenAI to the category of big customer we work with closely from interdependent technology partner. Yeah, they are very different from any other big customer that we have. The models that they are training on...

Azure supercomputers are still very important to the things that Microsoft is building. Their components that they're building are important parts of the Azure platform. So they are both customer and they are platform building partner.

Look, they have a bunch of things that they're off trying to do on their own that are independent from us, like ChatGPT, for instance. That's awesome because their success with ChatGPT is helping put a bunch of super good pressure on the Azure platform. That is another consistent thing. I don't know whether I talked about this the last time we talked about the OpenAI partnership, but one of my

core thesis when we were doing the first deal, like five, six years ago now. It was, we need the best AI workloads in the world to be running on Azure so we can ensure that Azure is building itself in a world-class way for those AI workloads of the future. The more successful ChatGPT is, the better Azure gets. Yeah.

Yeah. Seeing those AI workloads, my colleague Tom Warren has reported that Elon Musk and XAI are preparing to host Grok on Azure. He tells me to ask you, there's some angst within Microsoft about working with Elon and whether you can trust that company, especially with these other dependencies. Do you feel that angst? Yeah, I'm actually not super plugged into that conversation. I know we're doing it. The thing that we're trying to do with the...

The model marketplace on Azure is like we want to make sure that all of the good open source models that developers want to use are available and easy for them to use. Everything that we can offer there, we do offer.

Do you still control the GPU budget at Microsoft? This is a thing you said to me several years ago, and I've never stopped thinking about it. You don't anymore? I do not. Thank God. What happened there? Was it just like, this is too much? Because you described it as a horrible job. Oh, it was a horrible job. Yeah.

Is the pressure on needing GPUs, is that lessened or is it increased? Yeah, we still need lots and lots of GPUs. Because there is also reporting from Reuters and others that Microsoft has slowed down some of its data center investment or reallocated it as the models have gotten cheaper to run, as things like DeepSeek have shown up.

We still are urgently deploying capacity. And the thing that I will say is if you are sitting inside of Microsoft and you are talking to all of your teams that are building AI products or doing AI research, I have seen no lessening because of any technology trend in the

desire for more GPUs. Do you think we can do AGI in the current hardware? This is a thing that's floating around that I keep hearing from people across the industry. I don't even know what AGI is, which is a thing I've been a little bit confused by since I wrote my book many years ago. I think first you'd have to define what exactly it is you think that means. I think...

I think if you look at the current generation of hardware that's rolling out, we're getting a big performance win from this next generation that's deploying right now. If you are thinking about what's going to happen over the next 12 months, it's going to be a pretty substantial leap forward in performance just across the board of everyone's systems because this current hardware generation and optimizations you can do on top of it are extraordinary.

Do you think that we're going to get more capacity because of the optimizations or because the hardware is more powerful? It is absolutely true that most of the performance wins are coming from the optimization. So you get, you know, order, you

2x improvement in price performance for every hardware generation, which is extraordinary, by the way. You never got that every 18 months with Moore's Law. It was a little bit slower than that. And so the hardware advancements here are just...

breathtaking in terms of like how good they are. But like the software performance optimizations on top of the hardware is like much bigger than that even. So we're just very reliably like getting order of magnitude every year or so when you combine those two things.

How would you categorize those optimizations? Because a lot of the early wins in model capability came from just ingesting more data, right? We just made the models bigger and that is how they got smarter. How would you characterize these optimizations? Yeah, look, it's a bunch of things. So like there are things that you do with how you're training the models. There's a whole bunch of things that like a lot of the wins have come from being able to effectively use

smaller data types to store model activations into both on the inference and the training side, which means that you can do a lot more arithmetic in parallel because you have like smaller numbers that you are using in your arithmetic operations.

I mean, it really is kind of crazy to see the breadth of the optimizations, you know, from, you know, what, what you're doing and in a training side to, you know, just people fundamentally rewrite rewriting the numeric kernels, uh, for the inference stack. And, and then there, there's a bunch of things that you can do using your standard bag of computer science techniques around, uh,

prompt optimization and caching and using more than one model to service prompts. You don't always need to send every prompt to the most expensive model. We now have a big enough portfolio of models where you can choose to handle certain things with models that are hyper performance optimized but less general.

you know, send the more complicated things to the, you know, to the bigger, more expensive models. It's funny when I talked to the other agentic AI CEOs, they described that kind of orchestration as the key. And you're talking about MCP is the key. And I'm curious which one has to come first, right? Like we got pretty far in orchestration. Yeah. I look, I think,

Again, going back to this capability overhang, I think we have more reasoning power in these models right now than we're effectively using. And so my hypothesis is that one of the things that is preventing us from having more useful things is just action taking and the entirety of the action space is too constrained right now. And so I'm not saying it's an either or. I just think it's going to be super hard work to get...

that action space opened up. And so like, we just need to get on it. Like it's a ecosystem right now.

I want to end by talking to you about my favorite thing to talk to you about, which is the relationship of technology and art. You did write a book. It's called Reprogramming the American Dream. The first time we ever talked, it was about that book. Somewhat notably, the foreword to that book was written by J.D. Vance, who is now the vice president of the United States. I don't think you saw that one coming at that time. You did see a lot of stuff coming about how AI would reshape the economy or at least threaten to reshape the economy. Yeah.

As we talk about the models getting more capable and the ways they're getting more capable, the very notion that we could make the models more capable by just ingesting more data has hit a limit. We've ingested all the data. And now there's a lot of lawsuits about whether that ingestion of data was legal, whether it should be compensated. You're an author. I'm going to put these to you as simply as I can. If I stood outside the bookstore and everybody who came looking for your book said,

I stopped them and said, I can just make you a podcast about that book. You just send me a note, like just text me and I'll send you a full podcast summarizing that book. Do you think that would increase or decrease sales of your book?

Well, look, I will say something that's about Kevin that I don't know whether you can or should generalize to authors in general. I'd be fine with people doing whatever they want to with the content of my book. If you had to make your living based on sales. Yeah, yeah, yeah. Which is why it's super different. Look, I think authors...

If you spend all of your time and dump your heart and soul into making a thing, you deserve to be compensated for it. Now, I think there's a lot of different ways to be compensated for a thing. I'm not even following what's going on with a bunch of this litigation closely enough that even if I could comment on it, it would be useful commentary.

I think going back to the conversation that we started with, open protocols for this agentic web is that people who are making things get more, particularly here at the beginning where the whole landscape hasn't really sorted out what the business model is. I think people can jump in and have more agency around business model. Did you make a Studio Ghibli meme?

when I released that tool? No, I didn't. You didn't? Okay. I'm curious. A lot of people did. I'm not saying I didn't. No comment from me on that. My point is, as somebody who makes creative work, you're a maker yourself, there are lots of people who their livelihood depends on economic exchange for their creative work. Yeah. And their criticism of the AI industry broadly is you've created all of this capability, maybe more capability than we're even using, as you're saying. Yep. But we have received nothing in return.

And we're just, we're in it now, right? Act two, middle innings. And it doesn't feel like that has changed but for some litigation. I'm just wondering if your thinking has evolved or matured. Well, I think the way that I was thinking about this in the beginning is the, you know, the, the,

way that I'm thinking about it right now. So, like, I certainly wouldn't want to see anything cause Hayao Miyazaki to make less beautiful things. Like, I'm, like, maybe one of the biggest fans in the world of, like, what he and Studio Ghibli have done over the years. Like, I think it's just...

i mean it's like some of the most beautiful art that was created in the 20th and early 21st century and so yeah like i like people like that i want them to have every incentive in the world to like do more of what they're doing i mean the the thing though too with the platform is like i i will just sort of you know come out and say it like i'm

not really super interested in these image generator things. Like what I'm interested is a model that can do medical diagnostics for my mom who lives in rural central Virginia and doesn't have access to really super high quality healthcare. And there are just, you know, tens of millions of people like her in the United States who are in the same predicament. It doesn't get better over time because of, you know, the demographics and,

You know the United States, you know Studio Ghibli Content has nothing to do with whether or not AI is like good at medical diagnosis or not and so like that's the thing I want to make sure of in this debate that you know, we can have the part of the debate which is

There's this creative economy that I don't want disrupted at all because I'm a fan of and I appreciate those folks. I was having this conversation with Reid Hoffman and J.J. Abrams a few months ago where, if anything, I would love for it to be the case that AI made it easier for people like J.J. Abrams to

to do more of what they do. And I think is, you know, like you just sort of have more stuff floating out there that like, that's what people will want. They'll want more of the, you know, Hey, here's, I am a fan of JJ Abrams and his voice and his work. Like, give me more of that. Like not, you know, this, this,

random crap, you know, like what some teenager has gotten out of an image generation model. That's a very important debate to have. I don't want that to overshadow like this other thing, which is like these tools can be enormously useful for solving some very important problems. And we don't want to

have this conversation that we're having over here, which is important to have to impede our ability to push forward with this other stuff, which is also super important, if that makes sense. It does. It does. It's interesting because you are describing what more or less is the framework for

The Copyright Office just released before Trump fired the Librarian of Congress in the Register of Copyrights, which appears to have backfired. And now there's an even more copyright maximalist person in that position. Because that report said, look, some of these uses for training data are obviously fair use. Scholarship and research and all this stuff. And then I'll just read the quote from the Copyright Office preliminary report that came out last week.

Making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, especially where this is accomplished through illegal access, taking the stuff without permission, goes beyond established fair use boundaries. Here's the distinction. There's some stuff where we say medical imaging, the utility of this is so high and the work is so transformative, maybe fine. And then some of the stuff where you've just copied every YouTube video in the world and you're just letting people make more YouTube videos, probably not.

Can you envision a framework where that would apply to the tools Microsoft is building? Where you would say, this is stuff we're going to do and this is stuff we're not going to do? Yeah, look, I think we're open to having any kind of sensible conversation. I think you just have to sort of show up and there's some...

technical limitations and constraints on what's possible and what's not possible with the fundamental technology. But I think there's a super rich dialogue to be had here. There's also this interesting thing where I think increasingly you accurately identified that

we've kind of exhausted all of the data that is available to train models. And so, you know, we're now in a regime where a bunch of these systems are being trained with a set of techniques where they don't depend as much on data as they once did. And so, like, there may be all sorts of technical ways to, like, enhance the resources

reasoning capabilities of models that aren't quite as dependent as they might have been at some point on ingesting a bunch of sort of organic tokens of data. It's also true, I think we talked about this last time, that there's an increasingly...

good understanding of the quality of data, like how much a token of data contributes to the reasoning power of the model. And then the biggest thing, my bugaboo in general with all of this stuff is thinking of models as...

databases, like his information retrieval systems is kind of like you want to talk about a sub optimized system, like they're kind of terrible as databases, just from an efficiency perspective. And so again, you know, you go back to things like in a web, like there is the, you know, here's a model and like it has learned how to reason the same way that you have

you might've taught a biological brain how to reason. And then like once you have a certain level of reasoning capability, like the interesting thing is like, you know, prompt by prompt, task by task, like what information do you have access to, to reason over and like, you know, how you monetize those two things and like what the share of the, you know, business is between those two things can be very, very different. Yeah. With, with one, for instance, like you could have, you know,

if you need a model that's reasoning over breaking news, like if you have...

you know something like in a web and a subscription to a bunch of news outlets you can provide the agent access to those subscriptions if the publisher wants to allow that uh using the user's authorization tokens and then like let the model reason over that information and like you're sort of paying a subscription fee to like have this ephemeral content to reason over

So I think there are all sorts of ways we may be able to sort the business model stuff out over time. Yeah. I mean, I want to bring this all together. It sounds like with the new search project and our web and with the investment in MCP and wanting it to be more widespread –

It just feels like you're trying to create a wholesale architecture shift for the web, right? Like this is the new kind of web and you're trying to just sort of incept and incentivize its creation because the deal of the old web appears to be up. Is that a fair characterization? I don't know whether the deal of the old web is up, but like it's kind of time for us to be thinking about some new, I think. Yeah. And, you know, like I think as we're all collectively thinking about the new, like we should do what every –

Yeah.

the creators and the consumers are like, have their interests balanced. And, you know, there aren't a bunch of like weird, you know, intermediaries, like constraining, like how utility and value gets exchanged. Well, I wish you luck because so far the creators have a very clear point of view on where their incentives are. Kevin, I could talk to you forever. Obviously, you got to come back soon. I want to keep an eye on this web project and see how it's going over time. Yeah. Thank you so much for having me.

I'd like to thank Kevin Scott for taking time to join Decoder today, and thank you for listening. I hope you enjoyed it. If you'd like to let us know what you thought about this episode or really anything else at all, drop us a line. You can email us at decoderattheverge.com. We really do read all the emails. You can also hit me up directly on Threads or Blue Sky, and we have a TikTok and an Instagram. They're both at decoderpod.

If you like Decoder, please share it with your friends and subscribe wherever you hear podcasts. If you really like the show, hit us with that five-star review. Decoder is a production of The Verge and part of the Vox Media Podcast Network. Our producers are Kate Cox and Nick Statt. This episode was edited by Xander Adams. The Decoder music is by Breakmaster Cylinder. We'll see you next time.