Nov 21, 2024 6:00 AM
Relevance! Relevance! Relevance! Microsoft at 50 Is an AI Giant—and Still Hellbent on Domination
Jaime Teevan joined Microsoft before it was cool again. In 2006, she was completing her doctorate in artificial intelligence at MIT. She had many options but was drawn to the company’s respected, somewhat ivory-tower-ish research division. Teevan remained at Microsoft while the mother ship blundered its way through the mobile era.
Then, as the calendar flipped into the 2010s, an earth-shattering tech advance emerged. A method of artificial intelligence called deep learning was proving to be a powerful enhancement to software products. Google, Facebook, and others went on a tear to hire machine-learning researchers. Not so much Microsoft. “I don’t remember it like a frenzy,” Teevan says. “I don’t remember drama.” That was a problem. Microsoft’s focus remained largely on milking its cash cows, Windows and Office.
In 2014, Microsoft surprised people by promoting the ultimate company man, Satya Nadella, to CEO. Nadella had spent 22 years pulling himself up the ranks with his smarts and drive. And his likability. The latter trait was a rarity at the company. Nadella knew its culture intimately, and he knew he had to change it.
Three years later, Teevan became Nadella’s third technical adviser—and the first to have a background in AI. Then she became chief scientist, and her task was to imbue the company’s products with the AI of the time. In 2019, Nadella made the bold decision to spend $1 billion to partner with OpenAI, the small but trailblazing company that was leading the field. Microsoft was given unbridled access to its technology. It was a risky bet—even experts like Teevan, who’d seen OpenAI’s progress over the years, were skeptical that the tech would make much of a difference.
Then, late in the summer of 2022, she was invited to a demo of OpenAI’s latest large language model, GPT-4, at Microsoft’s Redmond headquarters. It took place in a windowless, gray-carpeted conference room in Building 34, where Nadella works. Two OpenAI cofounders, Greg Brockman and Sam Altman, came bearing a laptop. Brockman started with the sort of demos Teevan had seen in an earlier model, GPT-3.5. The new model responded with more sophistication, but Teevan wasn’t blown away. She knew how to break LLMs with requests that would expose the system as a sophisticated word jumbler. So she put it through its paces. At one point she asked it to write a sentence about Microsoft such that every word began with the letter G. The software whipped up a response, but it used the word Microsoft. Teevan challenged the answer, and GPT-4 admitted it failed—but it also asked her, didn’t you want the sentence to be about Microsoft? Then it offered her an alternate sentence that didn’t use the company name.
Teevan was stunned—not just by the way GPT-4 handled the problem, but by its self-awareness. She hadn’t expected that kind of a performance for years, if not decades.
She left the meeting and started driving the 2 miles home. She was having a hard time focusing. She pulled off the road and turned into the parking lot of a 7/11. “I sat in my car and let out a full-on scream,” she says. “And then I went home and drank.” After the first whiskey, she put on a movie: Terminator 2.
Not long after, she came to work dressed as Sarah Connor, the film’s feisty heroine. Teevan knew what she had to do. OpenAI may have created GPT-4, but her employer had the rights to exclusively build it into its products—and could beat its fellow tech titans at the most pivotal moment since the arrival of the internet itself. A year and a half later, for the first time in its nearly 50 years, Microsoft became worth $3 trillion.
Two years after the demo that blew Jaime Teevan’s mind, I’m seated among 5,000 or so people at a Microsoft sales team event. Held at the start of a new fiscal year, in July, it’s an entire day of product demos and morale-building chats and speeches. The highlight will be Satya Nadella’s keynote. Tens of thousands of Microsoft employees are streaming the event from desks, conference rooms, and, for those in distant time zones, kitchens and home offices to hear from their boss.
Onstage, an engineer who does client support work for the company’s Azure cloud service—a Softie with a Dave Grohl vibe—explains how OpenAI-powered tools can transmogrify a workflow. He tells the crowd how a developer from the AI team had shadowed him while he worked with customers, then developed a bot to do much of his work—better than he could, apparently. The bot launched in late 2023. “We’ve saved $100 million!” he says of the AI-based support program. “Thirty-one percent increase in first-call resolution! Twenty percent reduction in misroutes! Next year we’ll save $400 million.”
After the Azure headbanger leaves the stage, Nadella emerges from the wings. The bald, trim CEO, dressed in a T-shirt, gray slacks, and sneakers, has barely taken a step before it begins—the kind of rumble of a monster wave that sailors at sea might not live to describe. Everyone is standing and slow-clapping as he crosses the stage. Here’s the man who has not only ballooned their bank accounts but rocketed their status. As one longtime employee put it, “People look at Microsoft and think it is cool again.”
Nadella’s body language is an artful humble brag. His grin acknowledges the ovation while his hands make quiet-down motions. After he gestures the crowd to its seats, he cuts to the very question that has led me to the Pacific Northwest on this July day. “We are entering our 50th year as a company,” he says. “And there’s one thing I’ve been trying to make sense of … how the heck did it happen? How are we here as a relevant, consequential company in an industry that does not respect tradition?”
He tells a story from a few years ago, when a group of tech analysts came from China to take the measure of Silicon Valley. They attended all the key developer’s conferences: Apple’s WWDC, Google I/O, AWS Re:Invent, and of course, Microsoft’s own Build. “They said, ‘God, you know what? For anything that the United States has got, we’ve got equivalents in China. We’ve got ecommerce, search, hardware manufacturers, social networks of our own. But there’s this one company that we visited, Microsoft, that’s pretty different.’” As Nadella tells it, the delegation marveled at the company’s breadth, with everything from the PC operating system to Xbox: “It all comes together as this one systems platform.” And, he now implies, Microsoft’s breadth sets it up to seize the most propitious opportunity in the history of technology.
It was an odd choice of anecdote, considering that Microsoft’s history has been plagued by its eagerness to use its size as a cudgel—and that today it’s under investigation by the European Union and the US Federal Trade Commission for those same tendencies. Nadella skates past that and brings up his greatest triumph, AI. He tells the tens of thousands of Softies around the world that the new goal was to put Copilot—that’s Microsoft’s name for its AI—in the hands of people and organizations everywhere.
Nadella doesn’t say outright what everyone in the room knows: Just a decade ago, pundits had declared the company brain-dead.
In 1996, I wrote a Newsweek story called “The Microsoft Century.” The company, then more than two decades old, had been late to embrace the internet but now was using its power to neutralize its rival Netscape, ensconce its Internet Explorer as the victor, and eventually win the browser war. It was prepared to supercharge its dominant position in the tech world—seemingly for a generation. Back then, Michael Moritz, a venture capitalist who would later fund Google, told me, “In terms of a civilized world, you’d have to go back to the Roman Empire to find any organization that had as great a reach as Microsoft has today.” A lawyer who was trying to get the Department of Justice to file an antitrust lawsuit against Microsoft whined that the company was expanding into so many areas, “you may as well send your paycheck to Bill Gates!’’ Two years later, the US government did file suit against Microsoft, claiming it used anticompetitive and exclusionary practices to maintain its software monopoly and extend it to browsers. In 2000, the humiliating trial ended with a judge ruling Gates’ hardball, competition-squashing tactics illegal.
Even so, Microsoft avoided a breakup and retained its massive franchises, Windows and Office. But for the next decade or so, it operated with uncharacteristic timidity. It seemingly slept through Google’s launch of a web browser that eclipsed Internet Explorer. CEO Steve Ballmer, who inherited the role from Bill Gates, laughed off the iPhone, and Microsoft, the platform company, never managed to create a popular platform for smartphones.
Ballmer did ignite some smart initiatives that still serve the company well. He encouraged the growth of its cloud product, Azure. He also began the painful but necessary shift from software in boxes to web-based subscriptions. Still, Microsoft wasn’t working. Its strategy relied on maintaining a death grip on existing customers. “Bill and Steve were paranoid, particularly about Windows,” says a former senior executive. “And by the 2010s Windows was a fading paradigm.” Internally, he added, people were obsessed not with building products but with advancing in the hierarchy. Jaron Lanier, who joined Microsoft Research in 2006 and is now its “prime unifying scientist,” is more pointed: “There was rivalry. I don’t know how else to put this—there were unpleasant, powerful men.”
Tech analyst Benedict Evans captured the company’s decline in a July 2013 essay called “The Irrelevance of Microsoft.” “No one’s afraid of them,” he wrote. The next month, the board pushed out Ballmer. Contenders for his job included the CEO of Ford and the former president of Skype. But Nadella wrote a 10-page memo arguing that Microsoft’s revival would come from a growth mentality. As he later put it, he wanted to change the corporate personality from “know-it-all” to “learn-it-all.” The board—along with Gates and Ballmer, who were on the search committee—agreed that he was the one.
“Obviously, I’m a consummate insider,” Nadella says, talking to me in July after his speech and a lusty ovation. He saw firsthand how the company had lost its way. “You forget what made you successful in the first place. And hubris sets in.” Microsoft, he says, needed more than a great caretaker or efficient manager. “The metaphor I like is re-founding. Founders create magical things from nothing.”
From his first day as chief executive, Nadella worked at the company’s Glengarry Glen Ross culture. Perhaps in part because one of his children had cerebral palsy (Zain Nadella died in 2022), Nadella is highly empathetic. In Microsoft of old, everyone had a story about Bill Gates shrieking at the top of his lungs over their stupidity. In Nadella’s first meeting with department leaders he wheeled in a cart loaded with copies of a book called Nonviolent Communication and gave one to each person. “Before Satya it was difficult to show up to a meeting where you didn’t know the answer, or where you had a thought but you couldn’t prove it,” says Microsoft workplace exec Jared Spataro. “Satya was more like, ‘Come with your brain. Be sharp, and let’s talk about it.’ That felt liberating.”
Nadella wasn’t a blamer. In 2016, Microsoft was humiliated when its much-touted chatbot Tay proved shockingly vulnerable to being manipulated into generating racist content. Reviewers were brutal. “I was getting forwarded emails from really angry employees,” says Lili Cheng, who led the project. “I was feeling really terrible about putting the company in that position. And Satya sent me an email that said, ‘You’re not alone.’”
The CEO also smashed through outdated corporate ideas, notably Microsoft’s aversion to open source software, which it saw as a threat to its model of locking in customers with proprietary tools. “Microsoft had totally neglected the open source world for a decade—in fact they’d been hostile to it,” says Nat Friedman, who in the early 2010s ran a company based on open source software. “While Microsoft’s relationships with developers have been central to the success of the company, it had lost a generation.”
Nadella wanted to win the next one. Even before he became CEO, when he was in charge of Azure, one trip set him on the course. He and his lieutenant, Scott Guthrie, met with a group of startups to sell them on the cloud service. All of them used Linux. When the Microsoft executives left the room for a break, Guthrie said that Microsoft really should support Linux. “Absolutely!” said Nadella, trash-canning years of Microsoft dogma. Guthrie asked whether they should review the decision with other Microsoft leaders. “No,” Nadella said, “let’s just do it.”
“In a five-minute break, walking to the bathroom and back, we were able to completely change the company strategy around support for Linux and open source,” says Guthrie. When Nadella later told Ballmer, who was in his final days at the company, he simply informed him of the policy shift. Then, two months after Nadella became CEO, Guthrie suggested that they change the name “Windows Azure” to “Microsoft Azure.” It was done on the spot, sending a signal that Microsoft would no longer assess every move based on its impact on Windows.
Nadella also made the company less insular; he ensured that Microsoft’s cloud apps worked just as well on iPads and Android devices as they did on Windows machines. And he made a series of major acquisitions, many of which would end up shaping the company’s future.
The first of these was somewhat baffling. “He called me in one day and said, ‘I’m thinking of buying Minecraft, what do you think?’” says Yusuf Mehdi, a longtime Microsoft executive who had headed marketing for Bing. Mehdi began talking about the financials, but Nadella cut him off. “Tell me how we will land with customers,” he said. Mehdi’s answer was what Nadella had already figured out—that Minecraft-addicted kids in grade school who weren’t aware of Microsoft might make a connection to the company, and that could pay off in the long run. Unlike many previous acquisitions where Microsoft would absorb companies into “the Borg,” as employees used to put it, Nadella wouldn’t screw up the purchase by forcing it into the Windows ecosystem.
Mehdi came to call such purchases “reverse acquisitions.” “The philosophy is don’t mess them up. We buy them and say, OK, Microsoft is now your toolbox. They’ve all expanded us into areas that we normally wouldn’t be, like social networking.”
He is referring to LinkedIn. Nadella’s courtship of that company’s cofounder and chair, Reid Hoffman, began around 2015. “I basically got an email from him saying, ‘Hey, I think what you guys are doing at LinkedIn is really cool. Can we get on a phone call?’” Hoffman was impressed with Nadella’s soft sell. “Unlike all of my previous professional interactions with Microsoft, the conversation was grounded in a form of intellectual curiosity,” he says. The call led to a months-long exchange, and then to Bill Gates.
Nadella had been careful navigating his relationship with Gates, who was still to most people the face of the company. Gates agreed to spend 30 percent of his time advising Microsoft, and Nadella kept him close—no consultant in the world knew more about the business and the technology. Nadella would regularly go with key employees to Gates’ office to brief him on important initiatives. From accounts I heard, the briefings helped Nadella sharpen his thinking, since Gates was not reticent about criticism.
During the LinkedIn negotiations, Gates invited Hoffman to his office. “He spent two hours telling me why LinkedIn sucks as a product and Microsoft could build it very easily,” says Hoffman, who cheerfully defended his company. When Nadella and Gates later met with him and said they wanted to buy LinkedIn, Hoffman said he was surprised, given Gates’ remarks. “I was just testing,” Gates told him. Hoffman retorted, “Do you think everybody responds to that testing really well? Is that your theory of the universe?” Gates loved the pushback, and the two developed a relationship. The $26 billion deal closed in June 2016.
For Nadella, having Gates as a partner in such pursuits was critical, especially since aspects of his shopping spree weren’t all that popular among his executives. (Hoffman learned that most of the senior leadership opposed Nadella’s decision to keep his acquisitions largely separate rather than folding them into Microsoft.)
Arguably the most significant get in Nadella’s haul was GitHub, the open source code repository used by millions of programmers. Early in his tenure, Nadella and Scott Guthrie agreed that owning it would give Microsoft a huge advantage in winning the hearts and minds of developers, but the time wasn’t right. Developers didn’t think much of the tech behemoth at that point. “The community would rebel, and Microsoft would probably screw it up,” says Guthrie. But by 2018, that relationship had improved—which was lucky, because it was then or never. Google was courting GitHub. Microsoft had to make its move. When the company approached the founders, Guthrie remembers them saying, “‘We’ve seen what you’ve done, we like your culture.’ Years before they never would have done that.” Weeks later they sealed the deal.
The value of that $7.5 billion purchase would skyrocket for Microsoft because, a year later, Nadella made his best move of all—a deal with the startup OpenAI.
Nadella had missteps too. He’d always wanted a moonshot. He wanted to reestablish Microsoft as a visionary company. As he put it in his 2017 book, Hit Refresh, three technology shifts were essential to the company’s future: artificial intelligence, quantum computing, and mixed reality. Nadella made his first bold bet on … mixed reality. Oops.
The embodiment of that bet was the development and release, in 2016, of an unwieldy $3,000-or-more headset called the HoloLens, which presented a digital layer over the view of the world in its visor. It captured the imagination of the press when first demoed but was expensive and not terribly useful. It now lives in the bardo of failed gadgetry.
The miss was even more striking because by then Microsoft’s competitors were laser-focused on AI. The AI elite at Microsoft—in kind of a “know-it-all” mindset—appeared stuck in the conventional paradigm of reasoning-based AI. (In 2005, the company’s chief scientific officer, Eric Horvitz, visited deep-learning guru Geoff Hinton and paid him $15,000 to write up his thoughts about the new technique. Hinton’s musings failed to convert the Microsoft stalwarts.) As Google and others embraced deep learning, Microsoft’s most promising development was a chatbot called Cortana. It didn’t captivate the public.
In mid-2017, Nadella asked Reid Hoffman, who had joined Microsoft’s board, to sit in on a briefing from the Cortana team. In their postmortem discussion, Hoffman was brutal: “What I see at Microsoft is a lot of mediocre goals as moonshots.” Nadella agreed.
No one was more aware of Microsoft’s AI gap than Kevin Scott. He had been a senior VP at LinkedIn and was thinking about his next chapter when Nadella asked him to consider joining Microsoft as chief technology officer. Scott, who took the job in 2017, figured out that his role had two parts. The first was to integrate new technology throughout the company. The second, perhaps more important, was to develop the technologies of the future. Artificial intelligence was at the heart of both. “We had good people leaving the company,” he says, “because they didn’t think we had our act together on AI.”
The year after Scott started, Nadella met with OpenAI CEO Sam Altman at a conference in Sun Valley, Idaho. It was a significant moment. OpenAI had, after a few years of floundering, found a path to the exotic sci-fi future that Altman and his cofounders believed in. The trick was exploiting a Google invention called transformers to create astoundingly capable language models. OpenAI had just broken with Elon Musk, who had essentially been bankrolling the company. Reid Hoffman, who was an early funder, was paying the bills. Altman’s company needed to make a deal with a big cloud provider so it could cover its biggest expense—the infrastructure to build and run its models. Like many people in Silicon Valley, Altman had once written off Microsoft. But of late he had been increasingly impressed with the new CEO, and especially the company’s cloud capabilities. In Sun Valley, they began talking about an investment.
By June 2019, it was time for a decision. Kevin Scott sent an email to Nadella and Gates laying out why Microsoft had to make the deal. Google had already started integrating transformer-based models into its products, including the crown jewel: the infrastructure for Google Search. Microsoft’s attempts to duplicate the feat on its own systems exposed the company’s weaknesses. “It took us six months to get the model trained, because our infrastructure wasn’t up to the task,” Scott wrote. “We are multiple years behind the competition in terms of ML scale.” That July, Microsoft put $1 billion into OpenAI.
Scott still marvels that Nadella made such a risky deal. “Even that initial dollar amount seemed like a lot of money,” he says. “OpenAI was clearly a brilliant, brilliant research team, but they had no revenue streams, no product. It surprised me that Satya was willing to bank on them.” Nadella, though, had a vision. Microsoft didn’t want a bunch of competing LLMs within the company. “OpenAI had the best, so we partnered with them—they bet on us, we bet on them,” he says. Microsoft would end up spending much, much more to build up its own infrastructure to accommodate training and operating the new language models.
Some of Microsoft’s AI cognoscenti remained skeptical of OpenAI. “Microsoft, partially because of Bill, was very taken with symbolic AI programs,” Hoffman says. “They thought that the only way AI would succeed is through explicit knowledge representation,” which was squarely at odds with the techniques of generative AI. They saw OpenAI’s apparent advances as a parlor trick.
Scott understood that the OpenAI deal wouldn’t just make the startup’s research available—it would nudge Microsoft’s own AI talent to stop clinging to their old paradigms. Microsoft chief scientific officer Eric Horvitz remembers one meeting where OpenAI chief scientist Ilya Sutskever outlined what he believed was a clear path to general artificial intelligence, something not talked about much inside Microsoft. “We came away with a sense of awe that these people might be semi-crazed but interesting, in a good way,” says Horvitz.
Microsoft kept raising its investment, eventually reaching more than $13 billion. In exchange, it got 49 percent of OpenAI’s profits and exclusive access to its technology. Scott, who lived in the Bay Area, regularly dropped into OpenAI’s San Francisco headquarters to see what the company was up to. In 2020, OpenAI released its powerful GPT-3 model, and the deal allowed Microsoft to exploit its powers. But still, there was no compelling use for the technology. So far.
That would soon change. An OpenAI researcher discovered that GPT-3 could write code. Not perfectly; there were mistakes. But it was good enough for an instant draft of code that might take a decent programmer hours to produce. This was explosive. When Nadella saw a demo, he says, “I became a believer.”
OpenAI started developing its own coding product called Codex and worked hard to release it the following spring. Microsoft, however, had not only the rights to develop its own product but also the perfect platform for it: GitHub, where, as Scott says, “an enormous number of the world’s developers are doing their coding work.”
Not everyone at Microsoft, or even in the GitHub community, welcomed the idea of an AI coding assistant, especially since, as Scott put it, it was at “the ragged edge of possible—it barely worked.” But even in a nascent state it could relieve programmers of drudgery. That was going to be AI’s story for a while—fast but mediocre results at the start, then later, a model would emerge that could beat your ass at your job.
“I took it to some of my most famous programmers,” says Nat Friedman, who by then was GitHub’s CEO. “It was really polarizing—some of the best developers said it was useless, because they could see it make mistakes. People were telling me, ‘I don’t think you can ship this.’ If I had been a regular Microsoft executive who valued his career, I might not have.” Microsoft’s AI responsibility team produced a many-page report calling the product … irresponsible. “Ultimately,” says Friedman, “I said, look, I’m the CEO of GitHub—fire me if I’m wrong.”
So Friedman asked the Azure cloud team to get more GPUs. That request happened to coincide with an availability of 4,000 Nvidia chips. To secure them, though, GitHub would have to commit to the whole block, all 4,000 of them. And that would drain its budget of $25 million per year. “That was a lot of money for us—we were a zero-dollars product, and we had no idea how this would sell,” says Friedman. Even so, he made the commitment.
In June 2021, the new product went live. It was dubbed GitHub Copilot, an idea from a team member who happened to know that Friedman flew planes. “As soon as we heard it, we were like, that’s too perfect,” Friedman says. “The name communicates your relationship to it—you’re still the pilot.” Soon hundreds of thousands of developers had signed up and become unpaid evangelists. “Every time someone commented that it was terrible, it made mistakes, someone else would reply, ‘I’m using it every day and really like it,’” Friedman adds. GitHub started charging for access to Copilot and more than made back its $25 million.
Friedman felt that the industry was on the cusp of a big transition. He left Microsoft to become a funder of AI startups. “I really expected that GitHub Copilot would trigger a wave of new AI products, because the developers who used it would see that AI works,” says Friedman. But then, to his disappointment, “nothing happened.”
A year later, of course, everything happened. And Satya Nadella made sure that Microsoft was in the middle of it.
OpenAI had a new model, GPT-4. Even before it was done training, OpenAI knew the model was crossing some sort of AI Rubicon. That summer, engineers there started showing it to Microsoft. As Jaime Teevan saw in the demo to her group, it seemed as if something was alive in those models. GPT-4 was the starting gun for the wide implementation of AI across Microsoft’s products.
There was one prominent holdout: Bill Gates. Nadella was long past the point where Gates’ opinions could threaten any move he chose to make. Gates wasn’t even on the board anymore; he’d resigned in 2020. But the BG sign-off still mattered. Altman now had a relationship with him too. “I’d gone through the transition from seeing him as a character to a real person,” Altman says. “So I was not surprised by his skepticism and directness.” Gates told Altman that if OpenAI’s chatbot could take the AP Biology test, and get the top grade of 5, he would be impressed.
The demo took place in Gates’ sprawling mansion on Lake Washington, with much of the Microsoft brass in attendance. Greg Brockman fed prompts to the system, with the help of a young woman who was a biology exam Olympian. GPT-4 aced the test. After the demo, Hoffman asked Gates how it ranked among the thousands of demos he’d seen. “Only one other might be as good,” Gates said: the day he went to Xerox PARC, in 1980, and saw the graphical user interface. Gates had gone from skeptic to booster.
In the wake of GPT-4, Kevin Scott wrote a company-wide memo called “The Era of the AI Copilot.” Once again, he tapped OpenAI’s ambition as a model for Microsoft—this energy was powerful enough to turn around the corporate battleship. He urged Softies to temper their skepticism. It was time for the lumbering giant to pounce on these technologies with ambition and vision, even if the outcome was impossible to predict:
What should we build with them? The beauty of platforms is that I’m not entirely sure: It’s your job, and the job of the world’s developers, entrepreneurs, and creators to figure that out! But one thing that we are increasingly sure of is that foundation models will create a brand-new category of software, perhaps the most important category that has ever been created: the Copilot.
Every Friday at 10 am, the 17 members of Microsoft’s senior leadership team meet with Nadella in his conference room. The session, informally called “soak time,” goes on for hours. In late 2022, most of those hours were taken up with Scott’s rhapsodic vision of the Copilot era. GPT-4 still hadn’t been released, and not many people had hands-on experience with it. But Microsoft had to move fast. Google had had LLMs for months but had squandered its lead. This was Microsoft’s chance to gain an edge. Teevan held daily calls with five product leaders, each overseeing a team of thousands, and tried to convey how they should proceed.
In November, in the midst of that scramble, OpenAI released a product called ChatGPT. Even though it ran on the earlier model, 3.5, its accessible interface made it easy to use—and let the public know how far AI had come. By the end of January, 100 million people signed on to use ChatGPT. Across the tech world it triggered a panic as it became clear that the companies leading in AI would be winners and the laggards would be dust. Microsoft suddenly worked with existential urgency.
“In the words of the ancients,” Teevan says, citing an old samurai proverb, “one should make all decisions in the course of seven breaths.” In other words, move fast when the time calls for it. “We were making all our decisions within the seven breaths. Every day we were figuring out what the model could do. We were going to make it work.”
GPT-4 was still prone to spitting out falsehoods, but it seemed clear that AI could transform the search experience by answering queries not with a link but with a smart, researched explanation. So Microsoft chose its search engine, Bing, to be the first Copilot consumer app. Earlier in his career, Nadella had been in charge of Bing, tasked with making a serious run at Google Search. Nadella had put his heart and soul into it, but the search engine did not make even a dent in Google’s dominance. Now, with GPT, Bing might finally stand a chance. And by doing it before its rival did, Microsoft would, in Nadella’s words, make Google dance.
The team worked overtime, through the end-of-year holidays. That included red teams in charge of hunting for weaknesses; at one point the red team focused on child safety came back with some disturbing results. “They had been able to get the raw GPT-4 model to really successfully role-play grooming a child,” says Sarah Bird, Microsoft’s chief product officer of Responsible AI. Her team worked overtime to strengthen the guardrails and make it harder to “jailbreak” the large language model, which Microsoft secretly named Sydney.
In early February 2023, Microsoft invited journalists to its campus for the reveal of GPT-4-powered Bing. Nadella began by comparing the moment to the founding of the company, when Bill Gates and Paul Allen rushed to be the first to write a Basic interpreter for the first PC, the Altair. “The race starts today,” Nadella told the crowd.
Altman appeared at the event: “I feel we’ve been waiting for this for 20 years. This is the beginning of a new era,” he said.
At first, pundits and observers universally lauded Microsoft for its boldness. So much so that they overlooked some missteps. Within a few weeks, those missteps started coming to light. Perhaps most infamously, the Microsoft chatbot revealed to a New York Times reporter its secret name, that it wanted to be a human, and that it was in love with that reporter, who should admit that he loved the bot too and leave his wife.
The incident was embarrassing, sure, but Microsoft shrugged it off as a growing pain. Bird says her team had essentially done triage to prevent the worst abuses, and judged that kind of intentional manipulation as a concern for the future. “The things we did spend time on were not issues,” she says.
Over the next year and a half, Microsoft built on its lead. It improved its products—“Sydney” was purged of its adulterous impulses—and eventually introduced Copilots for dozens of its products, including Windows and Office 360. The company has also made big investments in other AI ventures, including the French company Mistral. (Scott and Nadella parried my queries about whether Microsoft might one day create its own large language model competitor to OpenAI.) In March 2024, they hired DeepMind cofounder Mustafa Suleyman—basically taking over his startup, Inflection, by hiring its key employees and paying off the investors—and installed him as its head of Microsoft AI. “I’ve got about 14,000 people under me, many billions of dollars I’m accountable for,” he says. Suleyman also now sits next to Nadella at Friday soak time.
As for Suleyman and OpenAI, they talk about three times a week. “It’s like we’re married,” he says. I asked him whether the marriage is monogamous, considering Microsoft’s own research efforts and its subsequent deals with other AI companies.
“Technically yes,” he replied, which is what no spouse wants to hear. “Microsoft is a platform of platforms, so there’s no exclusivity. Microsoft’s very open to all kinds of options. OpenAI does what it wants to do, which is why it’s working with Apple.” He adds that OpenAI has its own profits and losses to worry about. He doesn’t mention that, according to the partnership, Microsoft reaps 49 percent of that profit. So if it’s a marriage, the prenup is tilted in the tech giant’s favor.
In January 2024, Microsoft snuck past Apple to become the most valuable company in the world. In the months afterward, it would tangle with Apple and Nvidia for the lead as its valuation climbed, at one point reaching $3.5 trillion. “It simply comes down to gen AI,” one analyst told The New York Times.
Satya Nadella had re-founded Microsoft. But he hadn’t perfected it. Nor had he cast out all its demons.
This past summer, the House Committee on Homeland Security convened in the Cannon House Office Building, in Washington, DC, for a hearing entitled “A Cascade of Security Failures: Assessing Microsoft Corporation’s Cybersecurity Shortfalls and the Implications for Homeland Security.” The committee was digging into a devastating report about a massive breach of national security, one that included the exposure of 60,000 State Department emails and access to the inboxes of Commerce Secretary Gina Raimondo and US ambassador to China Nicholas Burns. The report came on the heels of other recent breaches involving Russia, North Korea, and hackers just out for money and lulz. The report detailed an inexcusable breakdown of basic safety practices within Microsoft. It pointed out something that the legislators—and many critics—would focus on: When a company whose products are as ubiquitous as Microsoft’s fails, the pain goes far and wide. Preventable failures are inexcusable.
Speaking for Microsoft was its president, Brad Smith, who added that title in 2015 after years of service as general counsel. For decades, he has acted as the company’s master of deflection. While its new CEO was rejuvenating the company’s business plan and boosting its geek credibility, Smith and his team have had plenty to deflect: antitrust investigations, challenges to Microsoft’s acquisitions, and, as in this case, egregious security failures that had allowed the Chinese free access to America’s secrets.
That day on the Hill, Smith sat patiently while the chair reamed his company for the shameful neglect that ended up weakening national security. Then it was time to speak. Smith unleashed a series of regret-bombs. He accepted responsibility “without any defensiveness” for every charge of complacency and sloth. He vowed to do better. He explained that Microsoft had launched a Secure Future Initiative—“a multiyear endeavor to evolve the way we build, test, and operate our products and services,” as he’d put it in written testimony. Some 34,000 engineers would be involved. He never quite explained, though, why the security culture of a $3 trillion company had been so lousy in the first place. Legislators brought up reporting by ProPublica that said a Microsoft employee had noted a catastrophic breach but was ignored. And that Microsoft waited six months to acknowledge the incident on its website. The committee made it clear they found this unacceptable. “I said the same thing, and we had the same conversation inside the company!” Smith chimed in.
By the end of the nearly three-hour hearing, Smith had successfully shifted the committee from scrutinizing the company’s failures to exploring how to work together in the future. The urgency of the problem was underlined only one month later. Operations of multiple major companies, including Delta Airlines, came to a standstill because a cybersecurity company called CrowdStrike had pushed out faulty code that ran on Microsoft systems. It was a reminder that Microsoft’s pervasiveness meant its shortcomings are everyone’s problem.
Nadella loves to talk about culture, so I asked him why he had failed to create a culture of security. After all, he had been around in 2002, when a high-profile series of security failures prompted Bill Gates to launch a Trustworthy Computing initiative, eerily similar to the Secure Future one Nadella had launched. Microsoft never really became an exemplar of tight security, but in the past few years, as the government report noted, its failures were epic. Why had the company dropped the ball under his leadership? Did he fire anyone?
“This is not about a witch hunt internally at Microsoft,” he says, which I took to be a no. He admits that there are “perverse incentives” that presumably lead companies to invest in new products rather than devoting resources to locking down existing ones. But he also complains that “there a lot of people who are chasing ambulances.” Ultimately, he says he accepts the criticism and that he should do better. “That’s what will be culture change,” he says.
Security shortcomings are only one déjà vu at Microsoft. There’s also evidence that Nadella’s celebrated empathy hasn’t purged the company of its tendency to flatten competitors. In the old days, when it felt threatened by another company’s product, the (insidious) playbook went something like this: First Microsoft could try to buy the company. If that failed, it might make a version of the product and maybe include it for free in software used by hundreds of millions of customers. The Microsoft version might not be better than the competitor’s, but it wouldn’t matter.
In 2014, a company called Slack launched a workplace chat application that quickly grew into a threat. Microsoft, in one of its SEC filings, officially listed Slack’s rise as a factor that could hurt the software giant. According to media reports, the company first considered buying Slack for $8 billion. Then Nadella decided he’d rather build his own version of the tool, called Teams. It was free—and built into Microsoft’s Office product.
Microsoft didn’t bother to hide its plans. “Teams was largely built off the back of this idea that work-based chat, what Slack had to offer, was going to be the future of work,” says Jared Spataro, a Microsoft senior leader who was then working on the Office business. “We wanted it to be portrayed as Teams versus Slack. Satya has always taught us, don’t be afraid of using competition to make the product better and to draw people’s attention to it.”
Because Teams was free for so many users, Stewart Butterfield, the CEO of Slack, found it harder to win new contracts with big businesses. In 2021, Salesforce bought Slack for $27.7 billion, but Slack’s founders think that were it not for Microsoft’s anticompetitive practices, their company would have been worth more. (Microsoft’s response is that its customers expected Teams-like features, and Slack could have built features like video to be competitive.) Meanwhile, the European Commission, the EU’s governing body, was investigating Microsoft’s actions around Teams and Slack, a process that might lead to sanctions. In a clearly preemptive move, last year Microsoft announced it would no longer bundle Teams into Office by default. The EU, apparently unimpressed, released a statement in June, deeming the changes “insufficient to address its concerns.”
Brad Smith’s take on why Microsoft first bundled then unbundled Teams is a masterclass in deflection. “We looked back at it and said, ‘Gee, we would have been better served to offer a version of Office without Teams. It would not have been the biggest deal in the world to do that. Its inclusion was not because of some anticompetitive instinct. It was more a natural evolution of the product.”
The Slack investigation is only one in a series of recent or current complaints about Microsoft’s practices. The FTC is reportedly looking into Microsoft’s multiple alliances in the AI realm. It has also challenged the company’s $69 billion acquisition of Activision, which puts it in control of some of the world’s most popular games, including the Call of Duty and Diablo franchises. (Microsoft games czar Phil Spencer tells me that the real reason the company made the deal was to extend its reach in mobile games like Candy Crush and to bolster the company’s online gaming service, Xbox Game Pass. Soon after the deal, Microsoft raised prices on the service.) A revamped FTC in the new Donald Trump administration may well be friendlier to big mergers, ending those investigations and providing a green light for more big purchases by Nadella.
Then there are some of the familiar annoyances of Microsoft’s tactics, just as in the days of Gates and Ballmer. Instead of the reliable, familiar PC apps of old, which people installed on their hard drives, Windows users now have to deal with pricey, often less powerful, cloud-based subscription versions, and they have to sign in with a Microsoft account. And the company aggressively steers users toward its browser. Also unwelcome: the appearance of ads in the Windows Start menu.
Nadella shrugs off my questions about whether Microsoft still harbors the bullying behavior that powered its initial rise. “It’s not like the ’90s, where there was Microsoft and there was daylight, and then there was the rest,” he says. “There are now lots of competitors who on any given day can do anything.”
Or perhaps Nadella is just craftier than his forebears. “I don’t think Microsoft is now dumb enough to duplicate the [antitrust] case that humiliated the company,” says Tim Wu, an antitrust expert who served for nearly two years as a technology and competition policy adviser for President Joe Biden. “But I do think they have that core DNA.”
No question about it: Nadella’s Microsoft is a triumph. Finally, in the 2020s, Microsoft has centered on the most innovative tech since the PC itself. And though revenues from AI products haven’t begun to offset Microsoft’s huge investments, it has the confidence—and the resources—to wait until the products improve and users find them useful.
But can Microsoft really avoid the hubris that set it so far back before? Consider what happened in May of this year with a product called Recall.
The feature was supposed to epitomize Microsoft’s integration of AI into its hardware, software, and infrastructure. The idea was to give users something like a personal version of the Internet Archive. Recall would constantly capture everything that happens in your machine: what you read, what you write, pictures and videos you look at, sites you visit. Simply describe to your machine what you’re looking for: What were those carpet samples I was considering for my living room? Where is that report about the ecology of the Amazon? When did I go to Paris? Those moments would pop up like magic, as if you had a homunculus that knew everything about you. It sounds scary—kind of like an onboard Big Brother—but Microsoft insisted users could feel safe. Everything stays on your computer!
Almost immediately, critics lambasted it as a privacy nightmare. For one thing, they noted, Recall worked by default and gobbled up your personal information, no matter how sensitive, without asking permission. While Microsoft has emphasized that only the user could access Recall, security researchers found “gaps you could drive a plane through,” as one tester put it.
“Within about 48 hours, we went from ‘Wow, this is extraordinarily exciting!’ to people expressing some reservations,” says Brad Smith. While the press was piling on, Smith was on a plane to meet Nadella in Washington, DC. By the time he landed, he figured it would be prudent to make Recall work only if users opted in; Nadella agreed. Meanwhile, in Redmond, Microsoft’s senior executives piled into meeting rooms to see how they might scale back the product. Fortunately, since the feature had not shipped yet, they didn’t have to recall Recall. They postponed the launch. And they would add security features, like “just in time” encryption.
“People pointed out some obvious things for us to do, which we should have caught,” Nadella says. But his own Responsible AI team missed them as well. A measure of “know-it-all”-ness had led to an announcement of a product that fell short, indicating that, even when led by a purported empath, Microsoft still retains too many of its earlier character flaws. Only now, it’s a $3 trillion company with locked-in access to the products of the leading-edge AI operation.
“You can think about it in one of two ways,” says Brad Smith. “One is, ‘Gee, I wish we would have thought about this before.’ Hindsight is a great thing. Or two, ‘Hey, it’s good that we’re using this to make this change—let’s be explicit about why.’ It was really a learning moment for the entire company.”
That’s fine. After 50 years, though, it’s a lesson that Microsoft—and Nadella—should have learned a long time ago.
Getty Images (timeline)
Let us know what you think about this article. Submit a letter to the editor atmail@wired.com.