AI, ChatGPT, and the Race That Will Change the World

The story of the AI industry over the past seven years or so is the story of Prometheus in California—the story of humanity receiving the gift, or perhaps the curse, of a new kind of fire. Here the fire is a technology on the threshold of humanlike artificial general intelligence; the role of the Greek god is played by a group of mortal entrepreneurs and researchers, together with the chief executives of powerful tech companies. In “Supremacy,” Parmy Olson offers a history of how we got to this moment. Ms. Olson, a Bloomberg Opinion technology columnist, centers her tale on Sam Altman, OpenAI’s co-founder and CEO, with various allies and rivals of Mr. Altman in supporting roles; among them are DeepMind co-founder Demis Hassabis and Mr. Altman’s rich donor turned adversary, Elon Musk. The essentials of Mr. Altman’s story are, by now, well known. A computer[1]science student and poker enthusiast at Stanford University, he dropped out at 19 to start a social-networking company with the support of the Silicon Valley start-up incubator Y Combinator. His start up didn’t work out—but he impressed Paul Graham, the head of Y Combinator, ultimately becoming Mr. Graham’s successor in 2014. From there, he amassed wealth and contacts. All the while he maintained an interest in AI, one that he had developed at Stanford. Messrs. Altman and Musk, both of them concerned about the risks posed by future AI technology, made a fateful decision to join forces; in late 2015, Mr. Altman would start a non-profit, with millions in funding from Mr. Musk and others, to carry out AI research and development more responsibly than companies like Google would, and more openly. But within a couple years, Mr. Musk split with OpenAI and Mr. Altman over the organization’s direction. To raise more money, Mr. Altman devised a hybrid scheme in which the non-profit would own a for-profit company with capped profits for its investors—what Ms. Olson calls “a byzantine mishmash of the non-profit and corporate worlds.” It was this company that would, with a $1 billion investment from Microsoft, release the AI-driven services GPT-3, DALL-E and ChatGPT to a mostly appreciative world. Ms. Olson has done her homework on AI technology, offering careful but accessible explanations of such concepts as neural networks, deep-learning models and diffusion models. (The last are at the heart of image-generating AIs like DALL-E.) The book is thought-provoking on the dilemma faced by entrepreneurs who want funding for expensive leading-edge research while also wanting to maintain control over what they view as ethically fraught technology. Mr. Hassabis, at DeepMind, took the approach of striking a deal with Google that he believed would allow his company independence, including its own ethics board; in Ms. Olson’s telling, Google essentially reneged on its pledges, though there’s no indication she asked Google for comment. Mr. Altman worked out a quite different arrangement with Microsoft’s CEO, Satya Nadella. In the men’s first conversation, in a stairwell at the annual Sun Valley conference, Mr. Nadella “was struck,” Ms. Olson says, “by how big Altman wanted to go” with AI. The eventual result was a strategic partnership rather than an acquisition, a deal that gave OpenAI the independence that Mr. Altman wanted, leaving Microsoft without even a board seat. In return, Microsoft got AI technology that could differentiate its products from its competitors’. Ms. Olson also offers convincing, if conventional, reasons why Google let its own pioneering AI technology languish at first while OpenAI raced ahead. A research unit of Google known as Google Brain achieved a foundational breakthrough in 2017, with an invention known as the transformer (the “T” in “GPT”). For Ms. Olson, Google’s failure to capitalize on its invention more aggressively was mainly the result of “lumbering bureaucracy” and an imperative to protect its enormous search and advertising business. Another reason OpenAI pulled ahead, as Ms. Olson notes, came down to one engineer, little known outside AI circles, named Alec Radford. It was Mr. Radford who played the pivotal role in making the leap from transformers to a far more capable subset of them known as generative pretrained transformers, which could be trained on large bodies of text and then learn new tasks from a few examples. When Mr. Radford’s efforts showed promise, OpenAI’s leadership recognized what it had on its hands and quickly changed the company’s direction, focusing on GPT models and turning them into usable products. While Ms. Olson tells a clear and well-researched story, “Supremacy” has some nontrivial problems. The least among them is that the prose often tends toward the tired. (“Silicon Valley was the land of crazy thinkers.”) Of greater note, her narrative is distorted by her peremptory rejection of concerns about destructive behaviour by AI systems, a threat cited by many who are close to the work. Harm from misanthropic AI—for example, takeover of critical infra[1]structure—is an outcome that can’t be dismissed, given that no one really knows why large language models engage in reasoning-like behaviours or what’s going on inside them. For Ms. Olson, anyone who expresses worry about large[1]scale dangers from AI is a crank (“Musk went down the rabbit hole of AI doom”)—or, if not a crank, then a cynic trying to divert attention from what she views as the real problem with AI, namely race and gender bias. While bias is an important concern, it’s a non sequitur to insist that there’s a choice between one concern and the other. ChatGPT wouldn’t have made that mistake. And the mistake could be a big one. As Mr. Altman observed in a tweet in July 2014, a year and a half before OpenAI: “AI will be either the best or the worst thing ever.”

Mr. Price is the author, most recently, of “Geniuses at War: Bletchley Park, Colossus, and the Dawn of the Digital Age.