Excerpt from Chapter Eight

Agency

If Yann LeCun and Geoffrey Hinton were subatomic particles, LeCun might be referred to as the anti-Hinton. He is roughly Hinton’s equal in magnitude but is oppositely charged.

Both men have won the Turing Award—the highest honor in computer science. In fact, they received the 2018 Turing Award jointly, for their contributions to the deep learning revolution. As of 2025, if you typed “Godfather of AI” into the Google search bar, the first auto-complete suggestion you got was Geoffrey Hinton and the second was Yann LeCun.

But LeCun—a professor at New York University who was chief AI scientist at Meta until he left in 2025 to form his own company—counters Hinton’s dark pessimism about our AI future with sunny optimism. And he seems committed to his side of the argument at least as intensely as Hinton is. LeCun once posted on social media that all the “hyperventilating about AI risk” is based on a “preposterously stupid” premise that in turn is grounded in a “*complete* misunderstanding of how everything works.”

Among the people LeCun thinks completely misunderstand how everything works is Eliezer Yudkowsky. In early 2024, LeCun shared his opinion of Yudkowsky with Yudkowsky. Replying to a social media post in which Yudkowsky had defended the idea of “p(doom)”—the idea of estimating the probability of a catastrophic AI outcome—LeCun wrote: “To you, it may look like AI ‘just happens’ like a natural phenomenon you can do nothing about. So you are scared. But to us in [the] trenches, it is something that *we* build. We have agency. That gives us a level of certainty that you don’t have.”

Agency. That word is, of course, the noun underlying an adjective that figured in the previous chapter: agentic. But LeCun is drawing on a different sense of the word agency than the sense implied when we talk about “agentic AIs.” An agentic AI is just an AI that does something—maybe something as pedestrian as, say, ordering your standard lunchtime fare via DoorDash. To have agency in that sense isn’t, by human standards, all that impressive. In fact, it’s not much more impressive than what a thermostat does when it decides whether to turn the heater on or leave it off, depending on what the temperature is.

But sometimes the word “agency” means more. To remind someone who seems to be passively accepting their fate that they have agency is to tell them that they have the capacity to intervene in the flow of events and shape the future. They can exert influence not just in the sense that a thermostat exerts influence but in the sense that whoever set the thermostat exerted it. That person had the power to choose 71 degrees Fahrenheit or 72 degrees Fahrenheit, and they chose 71. They in some small measure changed history. They exerted mastery over the future. They exhibited agency in the strong sense of the term.

And they were able to do that because they understood the behavioral algorithm of the thermostat well enough to be able to predict how it would respond to their promptings. LeCun is saying that, though AI is way more complicated than a thermostat, we can understand it well enough to anticipate—broadly speaking, at least—how it will respond to our promptings. We can engineer it to be safe and so keep it under control. It will always be our servant and never our master. We, not it, will have the kind of agency that ultimately matters.

So, in a way, the debate between LeCun and Yudkowsky—and much of the debate between AI optimists and sci-fi AI doomers—boils down to the future application of the word agency. LeCun believes that, though AIs will have agency in some sense of that term—in the sense that they take actions—humans will have agency in the strong sense of the term; we’ll be the ones who make the big decisions, the ones in the driver’s seat; we’ll understand the AIs we build well enough to keep them in their place.

Yudkowsky, in contrast, envisions a day when the AIs will understand us much better than we understand them. And that means they’ll be able to manipulate us. So, should their goals diverge significantly from ours—perhaps because, as in the paper clip thought experiment, we don’t see all the goals they might choose in pursuing the goals we give them—we could be in trouble.

In the Yudkowsky scenario, the kind of agency we’re assigning AIs now is just the camel’s nose under the tent…

Excerpted from The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning by Robert Wright. Copyright © 2026 by Robert Wright. Published by Simon & Schuster.