Industrial policy for the real world

September 19, 2024

SEOUL – Google CEO Sundar Pichai has characterized AI as “the most profound technology humanity is working on. More profound than fire, electricity, or anything that we have done in the past.” The hype around “existential risk” in AI follows a similar narrative, analogizing it to Oppenheimer’s atomic bomb. Such grand pronouncements have stirred many a corporate board and government agency to develop AI deployment plans.

The problem, though, is that it’s not yet clear if the juice is worth the squeeze. The Head of Global Equity Research at Goldman Sachs recently threw a wet blanket on Promethean AI ambitions. As a company report summarized his findings: “to earn an adequate return on costly AI technology, AI must solve very complex problems, which it currently isn’t capable of doing, and may never be.” According to an Upwork survey, “Almost 80 percent of workers who use generative AI in their jobs said it has added to their workload and is hampering their productivity.” The core problem is that LLMs are language models, not knowledge models. They can predict what the next word in a text (or pixel in an image) is likely to be, but they’ve done no reasoning to make that prediction. That lack of reasoning limits their applicability in many situations.

Call it the “savage unicorn” problem, after a bizarre image brought to prominence by Gary Marcus. After being prompted to produce a picture of an old man hugging a unicorn in the style of Michelangelo, an AI image engine obliged — but showed the man looking deeply contented while being excruciatingly impaled by the unicorn’s horn. To be sure, such errors are rare. But in all too many contexts where serious money is involved, the kind of reliability assured