Kevin Scott: Treat AI Like a Hash Table, and Measure Platforms by What Others Earn

Lex Fridman introduces Kevin Scott as Microsoft's CTO, previously SVP of engineering and operations at LinkedIn, and before that the person who oversaw mobile ads engineering at Google. Scott also hosts Behind the Tech. The conversation was recorded more than a month before Microsoft announced its investment in OpenAI — a fact Fridman flags at the top, because people had already started asking him about it.

Fridman's first question is the impossible one: give a whirlwind of everything Microsoft is doing, research and product included. Scott says that once you include research, the task gets even harder. Then he tries anyway. There is a large cloud business and a stack of SaaS. There are the original productivity products everyone still uses, an operating system, and hardware from mice and headphones to high-end PCs and laptops. There is GitHub, LinkedIn, search, advertising, news, and Xbox.

The research group, he says, spans an awkward range: economics on one end, human-computer interaction and artificial intelligence in the middle. That is how Glen Weyl enters the hour. Scott calls him a really smart young economist his group works with a lot, the author of a technical book on radical markets. The fact that an operating-system company is funding that work is the tell. Scott does not treat "do markets work" as a question that belongs only to economics departments.

That morning he had been talking with Phil Spencer, who runs Xbox, and nerding out with Tim Schafer, the writer of LucasArts' *Day of the Tentacle* — a game Scott and his college friends played forever, now the subject of some Microsoft collaboration. Being CTO is serious most of the time. Occasionally he gets caught by how unlikely it is that he gets to have these conversations at all.

Radical markets, as Scott sketches them, start from a live argument: do capitalism and free markets actually produce benefits that are equitably distributed? If you doubt they do, one instinct is to tip toward socialism, central planning, and a government that decides where work and investment go. Weyl's instinct is the opposite. Lean harder into market mechanisms, and invent new ones.

The example Scott picks, and later calls the world's worst, is real estate. Imagine a continuous auction on the land you occupy. Someone can arrive, claim higher economic utility for the plot, and you either bid to stay or yield the asset. Speculative bidding would become expensive fast. Prices would be forced to sit close to the value the asset can actually produce. Once prices track value, you can tax them. You could not sit on a house you call worth ten thousand dollars while everything around it is worth ten million.

He is not sure he likes the prospect of being bid out of his own property. The design is intentionally provocative, and he says Weyl sells it much better than he does. The practical hook is a wealth tax of the kind Elizabeth Warren was then proposing. If you put a tax like that in place, you still have to set prices on assets. A real-time bidding mechanism might stop being a thought experiment and become a requirement.

Fridman ties the economics to Andrew Yang and to automation. Microsoft sits at the cutting edge of AI. Yang is campaigning on the fear that the same progress will cost jobs. How does Scott hold radical markets, innovation, and that worry in one frame? He calls it one of the most important questions in technology, maybe in society: how AI develops over the next several decades, what it is used for, who benefits, who is harmed, and who gets to steer.

The joy of the Microsoft job, he says, is the heritage of a platform company. Bill Gates said years ago that the measure of a successful platform is that it produces far more economic value for the people who build on top of it than for the owner or builder. Scott wants AI held to that same stick. It has to be a platform other people use to build businesses, fulfill creative goals, become entrepreneurs, and solve problems in their work and lives.

The failure mode is a handful of companies, sitting in a very small handful of cities, deciding what goes into AI and then capturing every commercially valuable use of the infrastructure. That is bad economics, back to whether markets work. It is also bad innovation. Scott says he has infinite faith that people with powerful tools will do interesting things. The people with the tools should not be tens of thousands. They should be millions.

He reaches for the late-18th-century steam engine, maybe the first large-scale machine substitute for human labor. At the start, the value went to people with capital to build factories and to the experts who could construct and maintain the engines. Access democratized. There is not one engine company that builds every engine and owns every product that uses one. Two people sitting in a room each have a MEMS gyroscope in a phone — a little engine, a component of the modern world. AI has to get there.

The tie back to radical markets is a project Scott's team is running with Weyl and Jaron Lanier. Scott calls Lanier the father of virtual reality, one of the most interesting human beings on the planet, and a sweet guy. Together they are working on "data as labor," also called data dignity. If data is the raw material the AI machine consumes, we are doing a poor job of running a transparent marketplace for those contributions.

Everyone contributes, in two ways. Explicitly: you set up a LinkedIn profile, you know what you put in, and you have some nominal sense of what you get back, though the service is free and the perceived cost is low. Implicitly: you give off data exhaust just by living with technology, and you do not know where it goes. The hope, in the limit, is an economy of controls and incentives in which people might earn part of a living from the data they create.

Some of this is already priced. Scale AI, and a cluster of labeling companies in China, pay workers to annotate training data for supervised learning. That labor is explicit. A label is a label, and those people are compensated. The rest is the hard part. Fridman asks whether you can actually put a number on everyone's ambient contribution. Scott does not pretend. "I don't know. It's hard. It really is."

Technologists, he says, already know the subtlety. A single record, a single click, is not valuable atomically. Value shows up only when data is aggregated at large numbers, then combined and recombined along paths that are hard to see. Even paid labelers are in that boat: you need lots of labels before a model performs. Until you can trace how value flows through those combinations, a fair market is a wish, not a system.

Fridman asks how Microsoft has changed since Office in 1989 and Windows 3.0 in 1990. Scott was not there for the whole history, but the through-line he cares about is that it started as a platform company and still is one. The thriving parts of the business are the parts that build platforms. The original mission was a PC on every desk and in every home — democratizing access to personal computing when microprocessors were new and people built machines from kits the way people still build ham radios.

In success, he says, a platform vanishes. You stop paying attention to it. Operating systems are still critical; when one fails you notice. You are not waiting for the next one the way you waited in 1995, when the Rolling Stones were on stage for the Windows 95 rollout and people lined up the way they later lined up for iPhones. Ubiquity means taking it for granted. That is not necessarily a bad thing.

Satya Nadella's mission, articulated when he became CEO, is to empower every person and every organization on the planet to be more successful. Still a platform mission, executed differently: a hyperscale cloud, AI infrastructure, a productivity suite, Microsoft Dynamics automating business processes. The vision is broader because technology is now a basic fact of life, and a platform player has many more directions in which it can transform.

Fridman notes that Scott had not mentioned mixed reality. Scott adds it, along with quantum, as a potential computing platform of the future, next to the obvious one, AI. HoloLens is not a century-scale toy. It is finding traction with technicians, service people, and factory workers who never used a computer to do their jobs — mobile, hands busy, no laptop, no desk. For them, he says, the headset is the same productivity jolt an office worker got from a first PC, and units are selling.

How does AI actually infuse Microsoft products? The company already offers neural-network training in the cloud, pretrained models, compute, and inference. Scott and Nadella now run review meetings with Nadella's leadership team and a cross-functional group from across the company, looking at AI infrastructure and at teams whose products depend on AI in a significant way.

His rule for how it should appear in an experience is blunt. You do not want the "AI-ness" to be the first-order thing. The product is trying to help you do something; the AI just makes that better. People get super excited about where AI shows up in products. His comparison is rude on purpose: do you get that excited about where you used a hash table in the code? It is a tool. An interesting programming tool. An engineering tool.

It is already everywhere if you look. Office has dozens of features powered by fairly sophisticated machine learning. The search engine would not work at all if you took the machine learning out. Content moderation on Xbox and xCloud uses it too — not as a recommender of what to watch next, Scott is careful to say, but as anti-bullying for a gaming audience, where the line between playful banter and legitimate bullying is subtle.

Fridman has a philosophical problem. Scott ran engineering at LinkedIn. Xbox is also a social network. The talk on the two platforms is different, and so is the line for bullying. Nobody, he says, knows where that line is — Jack Dorsey and Twitter were under fire at the time for trying to find it. How do you stop abuse and still let people be playful?

Vertical social networks, Scott says, are a little easier. If you have a clear notion of what the network is for, and you design a community around that, the content-safety problem has fewer dimensions than it does on a general-purpose platform. LinkedIn is about connecting people with opportunity: a job, a mentor, a sales lead, or a professional identity broadcast to peers. In some ways that is very broad. In others it is narrow enough that machine-learning systems can make better automated calls about what is inappropriate, dangerous, or illegal.

Xbox is simpler still. It is about playing games and having fun. Bullying is not fun, so you use everything in your power to stop it. The joke-versus-abuse line in gaming is still subtle, but the direction is clear. General-purpose platforms do not get that gift. Scott's judgment is that policymakers eventually have to say where the lines are, through a democratic process, because ordinary people should have a say. You do not want a non-representative group making unilateral decisions with national and global impact.

He also admits the awkward present. Some platforms already have to make those unilateral decisions, because policy will not move fast enough to prevent very bad things. The policy side, he says, needs to catch up as quickly as possible. The digital world is not the physical world in which nations were built, and what bullying, healthy talk, and global policy even mean is still an open question.

Fake news and deepfakes, Scott says, add another layer of complexity, but even the written word is a cautionary timeline. Papyrus is about three thousand years old. The printing press is five hundred. Truly ubiquitous print did not arrive until the end of the nineteenth century, with the offset press, then exploded with the industrial revolution's need for educated workers. Literacy and the volume of words rose together.

In those three millennia, humans invented journalism, editorial integrity, scientific peer review — a whole apparatus for filtering the noise technology had made possible down to something society could cope with. The PC did not exist fifty years before this conversation. Mary Meeker's latest deck, released the week of the taping, put global internet penetration at 50 percent: three and a half billion people connected. The compression is what worries him.

It is not surprising, he says, that we have not figured out what to do yet. It is mandatory that we lean into the problem, because three millennia of work probably has to be done in the next decade. GANs make the time gap worse. They have good uses — Microsoft is already generating synthetic visual data to train face-recognition models and reduce bias. The same technology is about to make fabricating a photorealistic fake video fantastically cheap. The default human assumption that what you see is roughly trustworthy is about to fail, and he does not think we are ready.

Fridman asks how Microsoft thinks about face recognition as a platform for governments and companies, given powerful uses and ugly ones. Scott points to a blog post Brad Smith, Microsoft's president, wrote the previous fall. Two claims: there are uses to which face recognition should not be put, and there is a need for regulation, with the lines drawn by a democratic process. In the short term Microsoft has already drawn some of its own and pushed back on certain uses.

San Francisco had banned government agencies from using the technology. That may prove a little overly broad, Scott says, but for certain law-enforcement cases he would personally rather be overly cautious until a reasonable, democratically determined framework exists. Bias is not a side issue. Underrepresented minorities missing from training data produce strange model behavior. Public research has also turned up weirder artifacts, including models that learn things like "all doctors are men."

His operational ask is small and serious. Before anyone publishes an experiment, pushes code, or even ships a paper, they should start thinking about potential negative consequences. Deepfakes worry him for the same reason the timeline does. Detection may help, though the fakes are already convincing even under machine scrutiny. Cryptography is the other half. Sign content. Attach a chain of custody. Click a video and see whether it traces to an identity you trust — the White House, a candidate's office, Satya Nadella, LinkedIn CEO Jeff Weiner.

"That's not super high tech. We've had all of this technology forever." The genie will not go back in the bottle; the underlying fakes will only get better. Fridman, who grew up in the Soviet Union, notes that nobody trusted the media there because it was propaganda, and the skepticism sent people looking further. Scott takes the analogy to science. Hypothesis, experiment, peer review, reproducibility: you do not have to trust anything. If you are skeptical enough, you can try to reproduce it yourself. A lot of engineers, he says, are wired that way, and we need a little more of it.

Fridman is a PowerPoint person, still using it to make graphics instead of Illustrator. He remembers the excitement of waiting for XP, and the shrug when Windows 10 arrived: nice improvement. What does the future of Windows and Office look like? Scott's answer is a pile of small productivity wins from AI-powered features. The products get smarter in a subtle way. There will not be a big-bang moment when a reimagined Clippy descends as a full assistant. Fridman jokes that he will wait. Scott does not close the door on Clippy, and then gets serious.

Machine-learning assists are already in the product in small ways. Auto-reply is getting better: it can tell that someone is trying to schedule a meeting, look at your calendar, and find a time that works for both. Microsoft Search is not just web search; it searches the information in your Office 365 tenant, and potentially in other products. The Microsoft Graph is an API federator that weaves together what used to be information silos. Scott often emails people asking for a paper or a document. There is no reason, he says, the system should not fetch it for him.

Inside Microsoft, more and more of the Office portfolio is happening inside Teams as a canvas, with collaboration at the center. A framework they had been building for a couple of years was about to go open source. Internally he still uses the code name; at launch he thinks it will be called Fluid Framework. You can sit in a Teams conversation, embed part of a spreadsheet someone else is editing in Excel, and watch the object update in real time. You work in whichever canvas is convenient. The same object is live in all of them.

Fridman asks the civilian question: when you collaborate in Word, don't people still send the file around? Scott says they are already a little better than that. Then he adds the sentence that makes the hour honest: the fact that Fridman does not know means Microsoft has a better job to do helping people discover the features. A man who leads tens of thousands of engineers treats undiscoverable collaboration as a product failure, not as a user error.

Fridman knows what it is to lead ten or fifty engineers. Microsoft has tens of thousands. What does it take to keep a group that large innovative and productive? Scott says scale teaches you that three things become very, very important. The first is forethought about what you will need to build over long periods — not betting the company on one product, but knowing the capabilities the future will require, then investing in that infrastructure.

Infrastructure, as he uses the word, is wider than storage systems and cloud APIs. It is what the development process looks like, what tools you want, and what culture you build around making complicated technical things together. Have an opinion, and invest. You can wing it at small scale; a startup does not need to be specific on day one. The worst miseries he has seen as an engineering leader came from not having a clear enough opinion soon enough, then inheriting technical debt and culture debt that is excruciating to clean up.

The second bundle is a mission that is not cute copy you say because you think you should have one. It has to clarify where you are headed together. He cites Yuval Harari's *Sapiens*: storytelling is the quintessential mechanism for coordinating large groups. Scott has watched it in engineering teams. Below Dunbar's number, about 120 to 150 people, you can brute-force it. You know the personalities, the tensions, the dynamics. Past that line, without a set of shared goals, things start to fail in a catastrophic way.

He knows it sounds touchy-feely. Technical people balk. He still calls the mission very, very, very important. Fridman says stories are the fabric that connects a society. Scott pushes further. A currency is a story. A constitution is a story. Laws are stories. We believe in them hard, thank God, but they are abstract — just words. If no one believes, they are nothing. In some sense those stories are platforms, some of which Microsoft is building, on which people define a future.

Last question, bigger than Microsoft: what do the next twenty or thirty years look like for computing, technology, devices? Scott's frame is that technology is progressing as fast as it ever has, while society-scale problems sit in front of it, and the two have to intersect. Climate is the first. Population is the second. The rest of the population boom over the next fifty years is largely in Africa, and warming will make feeding the planet harder, especially there. The two biggest variables collide.

The third is health care. If AI is pushed in the right direction, it can help people live better lives, which for him means using it to raise quality, lower cost, and widen coverage. That gets more important every day in the United States, the rest of industrial Europe, China, Japan, and Korea, where a bubble of working-age people will retire over the next two or three decades. More retired people than working-age people raises a blunt question: who takes care of the old, and who does the work?

His historical reference is the Green Revolution. As a kid in the 1970s and 1980s, everyone talked about a population boom the planet could not feed. They were in the middle of a technology-driven jump in crop yields — some of it Western knowledge distributed to the developing world, some of it smarter biology. We do not talk about overpopulation that way anymore, he says, because we more or less figured out how to feed the world. "That's a technology story."

He is super hopeful, and he is spending time trying to get other people hopeful too. Back to Harari: we are the stories we tell. If we get overly pessimistic about technology's future, we may fail to put in place the things we need for the best possible one. Fridman is glad the person leading those engineers is an optimist, and says he agrees with everything except that he still hopes Clippy comes back. Scott laughs, thanks him, and that is the hour. A few weeks later Microsoft announced the OpenAI investment. The argument in the room — AI as platform, engines as parts, data with a price, Gates's test — reads less like a values speech than like a caption on the bet.