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.