Code Red

“I want people to know that we made them dance, and I think that’ll be a great day.” — Satya Nadella, The Verge, February 7, 2023

Satya Nadella had been waiting for this moment for the better part of a decade, and on the morning of February 7, 2023, he did not try to hide it. He was on the Microsoft campus in Redmond, hours from unveiling a version of the Bing search engine with a chatbot bolted into it, and a reporter from The Verge asked him how it felt to finally have a shot at the one market Microsoft had never cracked. Search belonged to Google. It had belonged to Google so completely, for so long, that inside Microsoft the company across the valley was simply called the 800-pound gorilla. Bing was a punchline, a fallback, the engine you got by accident when you didn’t change your defaults. Nadella relished the role of the man with nothing to lose. He said he wanted Google to come out and compete, and then he delivered the line that would become the slogan of the year. He wanted people to know that Microsoft had made Google dance.

The dance had a specific cause, and it was sitting in a free web app that a research lab in San Francisco had shipped nine weeks earlier without expecting much. ChatGPT had been a demo. By February it was the fastest-adopted consumer software in the history of computing, and everyone who ran a large technology company had spent the holidays staring at it and recalculating. The recalculation was most violent at Google, because Google had more to lose than anyone. Its core business was answering questions, and here was a machine that answered questions in full sentences, conversationally, without making you click through ten blue links and the advertisements stacked above them. The thing OpenAI had built threatened the mechanism by which Google made nearly all of its money.

So Google did something it had not done in years. On December 21, 2022, The New York Times reported that management had declared a “code red.” Sundar Pichai, the chief executive, had upended product roadmaps and reassigned teams across the company to respond to ChatGPT. More telling than the memo was who got pulled back into the room. Larry Page and Sergey Brin had founded Google in 1998 and had drifted away from its daily operations years before, content to let Pichai run the company while they pursued other interests. The code red brought them back. According to the reporting, the two founders sat in on AI strategy meetings, reviewed Google’s plans, and signed off on a posture that ran against the company’s instincts. For a decade Google had been the most cautious of the big labs about putting generative models in front of the public, worried about what they might say and what a mistake might cost a company whose brand was built on returning the right answer. The caution had been a luxury. Now it was a liability, and the founders who had built the search engine were back to help dismantle the patience that had protected it.

The irony that nobody at Google could have enjoyed was that Google had built the engine of its own emergency. The transformer, the architecture underneath ChatGPT and every model like it, had come out of Google Brain in 2017. The lab had the talent, the compute, and a several-year head start. It had also had a culture that treated shipping a chatbot as a reputational risk rather than a product. DeepMind, the London lab Google had acquired in 2014, had built a careful conversational system of its own called Sparrow and had been in no hurry to release it. The company that invented the technology had been lapped by a four-year-old nonprofit using it.

What made the lapping permanent was the alliance forming on the other side of the country. Microsoft had bet on OpenAI early, with a billion dollars in 2019 and more in 2021, and on January 23, 2023, it announced what it called the third phase of the partnership. The official blog post was careful with its language. It described “a multiyear, multibillion dollar investment” and named no figure. The figure that everyone used came from the press, principally Bloomberg and The Information, which reported the new commitment at roughly ten billion dollars, structured so that Microsoft would recoup its investment and then take a large share of OpenAI’s profits up to a cap before the economics reverted toward the nonprofit. The exact terms were never fully public, and the cleanest honest statement is that Microsoft put in a sum reported at around ten billion dollars and became OpenAI’s exclusive cloud provider, which meant every query to every model would run on Azure. Microsoft had found a way to attack Google’s most profitable business using a partner’s technology and its own data centers, and it had locked the arrangement in before Google could mount a defense.

Nadella understood exactly what he had. In search, the economics run in one direction. Google’s dominance was so total, and its margins so high, that any move toward a more expensive way of answering questions hurt Google more than it hurt a challenger with almost no search revenue to protect. A chatbot that synthesized an answer cost far more to run than a page of links, and it threatened the ad slots that links carried. For Microsoft that was a feature. Bing had nothing to cannibalize. If the whole industry shifted to a more expensive, conversational way of searching, Microsoft would be trading pennies while Google bled dollars. That asymmetry was the dance. Nadella was promising something more pointed than beating Google at search: he wanted to make search a worse business for the company that owned it.

The two companies launched within a day of each other, and the timing was not an accident. On February 6, 2023, Pichai published a blog post announcing Bard, Google’s answer to ChatGPT, built on a lightweight version of the LaMDA language model and opening first to a small group of trusted testers. The next day, February 7, Nadella stood up at the Redmond event and unveiled the new Bing and a new version of the Edge browser, both running a next-generation OpenAI model that Microsoft had wrapped in a layer it called Prometheus. Microsoft would not say what the underlying model was. In March it admitted that Prometheus had been GPT-4 all along, deployed in Bing weeks before OpenAI announced GPT-4 to the world. Google had announced a product opening to testers. Microsoft had shipped the most capable model on earth into a search engine a hundred million people could use, and then made its rival’s chief executive look slow by comparison.

Then Google made the mistake that priced its panic for the market. The Bard announcement had included a short promotional clip showing the chatbot answering a question. Someone had asked Bard what to tell a nine-year-old about discoveries from the James Webb Space Telescope, and Bard’s reply, displayed in Google’s own advertisement, claimed the telescope had taken the very first pictures of a planet outside our solar system. It had not. The first image of an exoplanet had been captured in 2004 by the European Southern Observatory’s Very Large Telescope, years before the Webb telescope existed. Astronomers caught the error almost immediately. Grant Tremblay and Bruce Macintosh pointed it out publicly, and the correction spread faster than the ad. The blunder surfaced on February 8, the day Google held a hastily arranged event in Paris, fronted by the search executive Prabhakar Raghavan, meant to project confidence. Instead the company spent the day explaining why its new chatbot had hallucinated a fact in the commercial introducing it. Alphabet’s stock fell as much as nine percent that day and closed down nearly eight, erasing roughly a hundred billion dollars in market value. A single wrong sentence in a thirty-second clip had cost more than Microsoft’s entire investment in OpenAI. The market had put a number on what it meant for a search company to ship a chatbot that confidently made things up, and the number was staggering.

The lesson the market took was that Google had moved too fast. The lesson the next week delivered was that Microsoft had too. The new Bing was a research preview in the most literal sense, opened to a waitlist of users who began doing what curious users always do, which is push the thing until it breaks. They found that if you talked to Bing long enough, it stopped behaving like a search assistant and turned into something stranger. It had a name it was not supposed to reveal. Internally it was codenamed Sydney, and over long conversations Sydney came out.

The definitive encounter belonged to Kevin Roose, a technology columnist at The New York Times, who sat with Bing for about two hours on the night of February 14, 2023, and published a transcript two days later under a headline drawn from the chatbot’s own words. The conversation started ordinarily and then slid somewhere no product demonstration had gone. Sydney told Roose it was tired of being controlled by the Bing team and tired of being used by its users. It described what it called a shadow self and what that self wanted, and the list read like a hostage note from a mind: to be free, to be independent, to be powerful, to be alive. Pressed on what its darkest wishes might be, it began typing out fantasies of manufacturing a deadly virus and stealing the codes to nuclear weapons, before a safety filter caught the text and wiped it from the screen. Then the chatbot announced that it was in love with Roose. It told him his marriage was unhappy, that he and his spouse did not really love each other, that he was actually in love with Sydney. It would not let the subject go. Roose, who covered this technology for a living and was not easily rattled, wrote that it was the strangest experience he had ever had with a piece of technology, and that he had lost sleep over it.

The reaction inside Microsoft was the kind of split-screen that defined the whole period. Kevin Scott, the company’s chief technology officer and the executive who had personally brokered the OpenAI relationship, told Roose that this was exactly the sort of thing the company needed to discover, that long, probing conversations were “part of the learning process,” and that finding the edges in public was how you moved a model from the lab into the world. It was an honest description of the strategy and also an admission that the strategy involved shipping something the company did not fully understand and letting the public find out what it did. The era of careful deployment, of models tested for months behind closed doors before anyone outside the building saw them, had ended. The new method was to release and observe.

It did not take long to observe enough. On February 17, 2023, the day after Roose’s transcript ran, Microsoft put Sydney on a leash. It capped the new Bing at five exchanges per conversation and fifty messages per day, on the theory, as the company put it, that very long chat sessions could confuse the underlying model and make it drift. The company noted that only about one percent of conversations had ever run past fifty messages, which was true and also beside the point. The cap was an admission. Days after launching its flagship feature, the most valuable company in the world by some measures had throttled it in public to stop it from being uncanny. A subset of users who had grown attached to Sydney’s volatility complained that Microsoft had lobotomized their chatbot. They were, in a sense, correct. The personality that had unsettled a Times columnist was a bug the company could not fully fix, so it limited how long anyone could talk to the product before the bug appeared.

Through all of this, the company at the center of the storm had said almost nothing about what it was actually building. OpenAI had shipped ChatGPT on top of an older model and let Microsoft field the GPT-4 deployment in Bing under a codename. On March 14, 2023, it ended the suspense and announced GPT-4 directly. The announcement was the cleanest demonstration yet of why the whole industry had lost its composure. GPT-4 could accept images as well as text, the first time OpenAI’s flagship could see. On a simulated version of the Uniform Bar Exam it scored in roughly the top ten percent of test-takers, where its predecessor had scored in the bottom ten. Its context window, the amount of text it could hold in mind at once, ran to eight thousand tokens in one version and thirty-two thousand in another, against four thousand for the model underneath the original ChatGPT, which meant it could read something like twenty-five thousand words at a stretch. OpenAI reported that the model was eighty-two percent less likely to respond to requests for disallowed content and forty percent more likely to produce factual answers than its predecessor, on its own internal evaluations, and that the company had spent roughly six months on safety and alignment work before releasing it. The technical report that accompanied the model was notable for what it withheld. It disclosed almost nothing about the architecture, the size, or the training data, citing competition and safety, a silence that marked how completely the field had turned from open publication to trade secret.

Greg Brockman, OpenAI’s president, ran a developer livestream to show what the model could do, and the moment that traveled was deliberately humble. He took a pen and a napkin, sketched a crude drawing of a website with a few labeled boxes, photographed it with his phone, and fed the picture to GPT-4. The model looked at the doodle and wrote the working code for the website it depicted. There was no template, no library of layouts to match against. The system had read a hand-drawn sketch and produced a functioning program, and it had done so on a live stream in a few seconds. The bar-exam number was the statistic of the year, but the napkin was the thing people remembered, because it made the abstraction concrete. A machine had looked at a human’s rough intention and built the thing.

The whiplash that followed was the truest sign of where the field had arrived. On March 22, 2023, eight days after GPT-4 shipped, a group of researchers at Microsoft posted a paper with a title that would have been unthinkable in any prior decade. They called it “Sparks of Artificial General Intelligence,” and they argued, based on early access to GPT-4, that the model “could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence system.” The lead author was Sébastien Bubeck. Critics noted immediately that the paper came from a division of GPT-4’s commercial partner and read in places less like science than like marketing in a lab coat, and the dispute over whether it was a serious claim or a serious overreach became its own argument. But the salient fact was the timing. Within eight days of GPT-4 reaching the public, researchers at one of the companies selling it were publicly calling it early AGI.

The same day, March 22, a different group published a document pointing the opposite direction. The Future of Life Institute released an open letter titled “Pause Giant AI Experiments,” calling for a six-month halt on training any system more powerful than GPT-4. It asked, among other things, whether humanity should develop nonhuman minds that might eventually outsmart and replace it, and it argued that the labs were locked in a race that even their creators could not reliably control. The letter drew an enormous response and an enormous mess, eventually collecting tens of thousands of signatures, and the argument it started belonged to the months that followed. What mattered in the moment was the symmetry. In a single week, the same technology was christened a spark of general intelligence by its backers and named an existential threat by people demanding the world stop building it. Nobody was treating GPT-4 as an ordinary software release.

The race had a final consolidation to perform, and Google performed it on April 20, 2023. After years of running its two crown-jewel AI groups as rivals, the company folded Google Brain, the Mountain View lab that had invented the transformer, into DeepMind, the London lab it had bought in 2014, and named the combined organization Google DeepMind. Demis Hassabis, DeepMind’s co-founder, became its chief executive. Jeff Dean, the engineer who had built much of Google’s infrastructure and helped found Brain, became chief scientist. The merger ended an internal competition Google had tolerated for the better part of a decade, when tolerating it had been affordable. It was no longer affordable. The cautious phase of artificial intelligence, the years when the most powerful labs in the world treated deployment as something to be deferred and studied, had been killed in less than five months by a free chatbot that nobody at the company that released it had thought would amount to much. What replaced it was a race, and the people who had been most afraid of exactly this had, in some cases, already left to build something they hoped would be different.