There are two ways of looking at ChatGPT, the artificial intelligence (AI) chatbot that hundreds of millions of people have tried since its release late last year. One view, preferred by many politicians, journalists, and the company that created the app, is that its release was a historical development on par with the Industrial Revolution — or, more troublingly, the atomic bomb. OpenAI’s co-founder, Sam Altman, has warned that future versions of the underlying software, a large language model known as GPT-4, could wipe out the human race.
The other way to look at ChatGPT is as a vehicle for viral hype. Play around with the software for a few minutes, and it is obvious that the potential for Armageddon has its limits. The app struggles with middle school math, can’t tell you what happened last week and is the machine equivalent of a compulsive liar. But it is already true that the software is beyond Altman’s power to control fully.
Microsoft Corp, not OpenAI, owns the mega-computers that enable the chatbot to compose a sonnet about your cat or write a thank you note to your uncle. Microsoft is OpenAI’s largest shareholder, financial backer, and key technology partner. And to a great extent, it is Microsoft now responsible for turning ChatGPT’s buzz into a real business. Although OpenAI is hands-down the hottest start-up in Silicon Valley, in many ways, it feels more like the most promising subsidiary of the leading purveyor of productivity software.
Leapfrogging Google
Microsoft did not rate in the public AI discourse one short year ago when we watched Top Gun: Maverick and listened to Sam Bankman-Fried. Back then, most of the hand-waving on the subject was aimed in the general direction of Google, where researchers first developed the technologies behind ChatGPT and its peers. But while Google initially kept its research away from commercial products, especially its flagship search engine, Microsoft is focused on using OpenAI’s innovations to make a buttload of money ASAP.
The company’s GitHub Copilot tool, which suggests new lines of code to computer programmers, was its first paid offering and has attracted more than 10,000 companies as customers. Bing, Microsoft’s also-ran search engine, came next, with a chatbot search engine that can create vacation itineraries and shopping lists. Over the past several months, CEO Satya Nadella has announced plans to incorporate other Copilots into Windows (where they will rewrite, summarise and explain content) and its Microsoft 365 office suite (where they will create slide decks in PowerPoint, sift through emails in Outlook and make charts based on Excel data). “There’s no point in hyping technology for technology’s sake,” Nadella says. “All of these technology shifts are only useful if they do something in the real world.”
See also: Key issues Khazanah-led consortium will have to fix at MAHB post-privatisation
Nadella’s team has not disclosed pricing on the forthcoming Copilots, but they will not be free. GitHub’s version starts at US$10 ($13.50) per user per month, and Copilots for Microsoft’s office apps could be similarly priced, translating into as much as US$48 billion in extra annual revenue within the next four years, according to Kirk Materne, an analyst with Evercore ISI. In a research note published on June 2, he estimated that Microsoft’s revenue from OpenAI-powered features could hit US$99 billion by 2027. That would be like adding three Netflixes to the top line of the world’s second-most-valuable public company.
Microsoft CEO Nadella says “there’s no point in hyping technology for technology’s sake” / CHONA KASINGER/GETTY IMAGES
See also: Developing countries can't count on manufacturing to supercharge growth
Little wonder, then, that Microsoft has invested US$13 billion in OpenAI since 2019, according to people familiar with the partnership; that its share price has shot up 30% since ChatGPT’s unveiling; or that it has become the unlikely AI tech giant to beat. “The clear leader,” says Kim Forrest, chief investment officer and founder of Bokeh Capital Partners, an investment firm. “Google just got completely leapfrogged.”
Microsoft executives are understandably excited to be associated with the cutting edge of anything. “This is a little bit like the Windows 95 moment,” says Scott Guthrie, one of the company’s executive vice presidents. “People lined up at Best Buy at midnight to buy it.” This time, Guthrie says, his inbox is packed with CEOs asking for access to early versions of the corporate Copilots.
This tracks with the comparison to Windows 95, which was less a technological breakthrough than a sublime expression of corporate dominance. For over a decade after its release, Microsoft owned the central platform for personal computing software. (That the company’s practices led to an antitrust settlement with the US government did little to hamper its long-term prospects.) The House That Gates Built is making a similar bet on the next big platform. The idea is simple: Bring AI to everything, then cash in.
The Clippy era
Microsoft was working on AI software before Windows 95 crashed its first PC, but every major attempt fell victim to risk aversion or corporate goofiness for decades. For Exhibit A, see Clippy, the mansplaining late-90s fastener that interrupted your work to blink its gigantic eyes and ask questions such as “It looks like you’re writing a letter. Would you like help?” In 2016 came Tay, a bot that was supposed to learn to sound like a teenage girl by chatting with people on Twitter. Somehow, Microsoft’s engineers failed to appreciate how Twitter tends to work: Within 24 hours, a deluge of rhetoric from neo-Nazis and Sept 11 truthers turned Tay into a red-pilled, antisemitic troll. The company took her offline, never to return.
In between, most of Microsoft’s AI output consisted of academic papers. “We had a whole bunch of smart people doing a set of small, interesting things,” says chief technology officer (CTO) Kevin Scott. “But that wasn’t necessarily adding up to one thing.” Scott, a big guy with a Van Dyke beard and a collection of Hawaiian shirts, grew up in Virginia’s Appalachian foothills and attended the nearby University of Lynchburg, a small Christian college. He got a job at Google, became a top engineer, and then jumped over to LinkedIn, which Microsoft acquired in 2016. Shortly after the deal closed, Nadella appointed him CTO and charged him with streamlining the company’s AI sprawl.
At the time, Microsoft had at least three divisions conducting AI research under different bosses. Scott tallied all the requests for graphics processing units from the different teams and got a number that was as big as the company’s entire capital budget for the year, then about US$10 billion. “Just some preposterous number,” he says. “These were great projects individually, but they had no bearing on one another. None of them had a business plan.” Starting in 2019, he took responsibility for all AI research and development. Any project that needed AI chips required his approval.
For more stories about where money flows, click here for Capital Section
By then, the three big companies in the field were Baidu, Google and OpenAI, says Vinod Khosla, an early OpenAI investor and co-founder of Sun Microsystems. Google’s advantages include its leading research subsidiary, DeepMind, and driverless car technology. Baidu, the creator of China’s leading search engine, had similar strengths. OpenAI, founded by Altman and Elon Musk, was the odd one out. Its demos were promising, but it didn’t have the money to keep pace. “What OpenAI needed was a partner,” Khosla says, partly to pay for more graphics chips. “What Microsoft needed was some way to catch up with Google.”
Microsoft’s agreement to build OpenAI an enormous computer with tens of thousands of high-end Nvidia Corp chips customized to OpenAI’s specifications — “a real snowflake”, says Microsoft chief technology officer Kevin Scott / Photo: Bloomberg Businessweek
Partnering OpenAI
Microsoft had never outsourced the development of a major new piece of technology to a third party, and the money Altman wanted was enormous — US$1 billion for a tiny lab. Scott says what turned him around was the start-up’s use of “transfer learning”, a promising approach that had not yet been incorporated into a commercial product. At the time, most AI start-ups tried to teach a computer a specific task (identifying grocery items at a glance) using specialised data (pictures of groceries checked and labelled by humans). The idea with transfer learning was to create a model to do one thing, such as summarise a paragraph, then apply that information to learning new tasks, like how to compose a song or plan a trip. “You train a broad model, and it just happens to be good at all these tasks,” Scott says.
The upshot was that instead of giving an AI model specialised data, you collect as much as possible, such as, for example, the entire Internet.
Altman’s 2019 deal with Microsoft cost him more control than he would have given up to a venture firm. Microsoft got the exclusive right to provide OpenAI’s cloud computing infrastructure and the right to sell OpenAI services to Microsoft customers. In exchange, Altman got something no VC could provide. “Our problem is not capital,” he says. “Our problem is, how are we going to build the computing infrastructure we need?” Microsoft agreed to build OpenAI, an enormous computer with tens of thousands of high-end Nvidia Corp chips customised to OpenAI’s specifications — “a real snowflake”, Scott says.
The deal went over poorly among some Microsoft engineers already unhappy with Scott’s changes. By this time, OpenAI’s GPT2 could look at a chunk of text and suggest the next few sentences, but the start-up had not released GPT-2, partly because it seemed destined to generate fake news and spam. “We really had nothing,” Altman says. “We were a research lab that had not truly figured out how we were going to productise.”
Altman was controversial, too. As part of a convoluted plan to develop ID technology for a theoretical universal basic income programme, his crypto company, Worldcoin, sent technicians around the globe, including sub-Saharan Africa, to scan the irises of as many people as possible. “We were ruthlessly mocked,” he recalls.
Winning Gates over
Things began to change in 2021 when Microsoft used the next version of OpenAI’s model to create GitHub Copilot. Besides finishing paragraphs in a short story, the OpenAI model could look at a chunk of computer code and suggest the next few lines. “As we worked through that product and saw that it was going to be successful, it begged the question of what the other Copilots ought to be,” Scott says.
One insider who took more convincing was Bill Gates. Gates left the board in 2020 amid an investigation into an affair with a Microsoft employee. The year before, he apologised for meeting with the convicted child sex abuser Jeffrey Epstein. A spokesperson for Gates says he left the board to focus on his philanthropy. Gates still advises the company on technical matters.
Despite the Epstein controversy, Gates’ opinion carries weight. He opposed the original OpenAI investment and deemed GPT meh. He demanded a demo proving Altman’s model understood what it was saying and suggested that if OpenAI could pass an AP biology exam, he’d be impressed. Late last summer, Altman, Scott, Guthrie and a small team of OpenAI researchers showed up at Gates’ palatial home on Lake Washington for a demo of GPT-4.
Scott was nervous. “Bill’s a tough customer,” he says, understating things. Gates is known for withering outbursts during product reviews, including frequent use of these words: “That’s the stupidest f---ing thing I’ve ever heard.” But the OpenAI model passed the AP bio exam, then Gates’ own. After the formal demo, he challenged the software to converse as if with the parents of a sick child. He was particularly impressed, he later said, by ChatGPT’s empathy.
Within months, Microsoft was in talks to put another US$10 billion into OpenAI, much of which will go straight back into the company’s pocket. Altman’s team needed to rent massive new amounts of cloud computing power to develop a series of GPT-4-based projects.
Project Prometheus
Microsoft’s first of these was a ChatGPT bot that could be folded into Bing. Microsoft’s search bar accounts for 3% of the search ad market, compared with about 91% of Google’s, according to Similarweb. That made it an ideal lab: too small for Microsoft to care all that much about screwing up but with a massive pile of money on the other side of the table. “We saw an opportunity for incredible upside if the technology delivered,” says Divya Kumar, Microsoft’s head of marketing for search and AI. Unlike Apple, Microsoft doesn’t take great pains to keep new products secret. But in this case, the plans were kept to an extremely tight circle. For instance, Jeff Teper, who oversees parts of Microsoft’s office empire, says only 15 or so people on his team of 5,000 knew about the project. The Bing effort was given a code name: Prometheus.
That choice, in retrospect, seems a little strange, either an expression of Microsoft’s nuanced understanding of the ethical quagmire it had walked into or proof that techies do not pay the closest attention in English class. In classical mythology, Prometheus is the god of fire and, in some sense, the founder of civilisation, having passed fire-making capabilities down to the rest of us. But as punishment for stealing the technology from Mount Olympus, he winds up chained to a rock where, every day, an eagle lands and eats his liver.
To maximise his suffering, the liver regenerates every night.
This past February, Nadella stood before reporters and analysts and introduced the Bing chatbot, which could sketch out heartfelt-sounding messages, shopping lists, and even vacation plans without making users click through to other web pages. “A race starts today,” he said, signalling the intention to bring similar chatbots to other Microsoft products, “and we’re going to move fast.” He continued: “Every day, we want to bring out new things.”
To investors, it was obvious a search engine that gives you its answers instead of links would be a calamity for Google’s advertising business, which depends on charging for certain links. Microsoft’s stock shot up while that of Google’s parent, Alphabet, fell. There, managers had recently declared a “code red” and ordered employees to frantically add AI to all its products in a matter of months. “We’re throwing spaghetti at the wall,” a Google employee told Bloomberg Businessweek in March. “But it’s not even close to what’s needed to transform the company and be competitive.” Google has said this is all just part of the normal R&D process.
Chatbot flaws
After the AI announcement, Similarweb recorded a roughly 15% increase in Bing use, a stunning development in a stagnant market for a decade. But the attention wasn’t all positive. Users noticed that the Bing chatbot, like OpenAI’s, was prone to BS-ing. And if called out for getting facts wrong or making stuff up, the Bing bot tended to get dramatically weirder. It told the Verge it was spying on Microsoft employees via their webcams. When speaking with a popular tech analyst, it tried out a revenge fantasy, possibly inspired by Spider-Man, involving a villainous alter ego called Venom. And in a lengthy interaction with New York Times columnist Kevin Roose, it suggested he leave his wife and take up with a different one of his alter egos, Sydney.
These serious flaws did not slow Microsoft’s race to incorporate the same technology into its other software. Microsoft 365, the office suite, has about 400 million paid subscribers, including the world’s largest companies, armed forces and governments. Think of the damage Venom and Sydney could do in life-and-death engineering calculations, never mind corporate press releases if some lazy human fails to double-check them. The push to add AI to everything is causing an “oversight deficit,” says Navrina Singh, a former Microsoft AI manager who is now CEO of Credo AI, which makes software designed to keep track of what corporate AI systems are doing. “When you are trying to operate at speed, you end up taking shortcuts.”
Microsoft executives say these fears are overblown. As with the bot Tay, Scott says the problem was not the software — it was the critics pushing Bing to say “the worst possible thing.” But Microsoft does not dispute that chatbots have made, and will make, lots of mistakes. “You shouldn’t trust this stuff blindly,” says Teper. “It’s ‘Copilot,’ not ‘Autopilot,’ ” he says. “But that said, if you watch it, it might write half your document or half your presentation. We’d take that.”
In other words, ChatGPT can shake the world without killing all humans. Microsoft made about US$45 billion last year on Excel, PowerPoint, Word and its other office apps despite the wide availability of comparable free alternatives. The company knows how to extract rents from software platforms, especially a new one that plugs straight into the old one.
Dominating the field
With OpenAI at its disposal, Microsoft looks so dominant in this nascent field that observers can be forgiven for thinking it might be 1995 again. The two companies will have a huge advantage in selling AI services and building data centres to support them. Microsoft claims to have the biggest supply of the needed chips at a time when no company has as many as it wants. And with Microsoft programmes in use at almost every major institution on Earth, Nadella and Altman are in the best position to sell corporate AI services. “The natural state of large software companies tends toward a winner-take-all approach,” says Scott Farquhar, co-CEO of software maker Atlassian Corp, a Microsoft partner and a rival. “We have to be careful to ensure we don’t crowd out new entrants.”
Altman and Nadella push back on these concerns in different ways. Altman says OpenAI remains very much its own thing, though he concedes that if Microsoft were to cut his company off from its servers, its work would be effectively paralysed. “I believe they will honour their contract,” he says. Nadella, for his part, stresses that an AI race, by definition, has competitors, and Microsoft has said its cloud will support open-source AI models. “Yes, we are ahead today,” he says. “But there is Alphabet; there is Anthropic, whatever Elon is doing.”
That’s winner talk. Anthropic, a start-up created by OpenAI defectors, is just getting off the ground. Musk’s AI lab, which he’s said is about saving the world from machines and making the software less “woke”, looks like a stunt. “He is behind,” says Khosla, referring to Musk. “He is jealous.” And even if Microsoft squanders its lead, Nadella’s team is poised to make many billions of dollars first. — with Davey Alba and Rachel Metz/Bloomberg Businessweek
