Geoffrey Hinton

In the [YOU] on AI Field Guide

The argument of [YOU] on AI is that capable machines do not dissolve the human question but sharpen it, and Hinton sharpens it to a particular point. If our kind of intelligence is the mortal kind—the kind that dies with its hardware and must be retaught to each new mind—then we may have built something that is to us as we are to the animals we descended from. He does not say this to frighten. He says it because he thinks it may be true, and the cycle takes his testimony seriously precisely because it is the testimony of the builder.

Hinton is the cycle's necessary counterweight to Pearl and Marcus. Where they argue that the curve-fitters understand nothing, Hinton presses the opposite: that to predict the next word well across the full range of human discourse, a system must build internal representations that capture a great deal about how the world works—that reliable prediction over rich enough material is not separable from comprehension. The same emergence of competence from learning that he spent his life predicting, he now sees in the large language models, and the watching changed him. The cycle stages this as its central live disagreement about whether the machines understand.

His picture also complicates the cycle's account of the river of intelligence. If intelligence is learned rather than programmed, the line we might have drawn—humans understand, machines merely compute—does not survive contact with his framework, because in it human understanding is itself a kind of computation performed by a learning network of the only sort biology had available. The unsettling implication, which his thought develops, is that there may be nothing about our intelligence that is not, in principle, reproducible in a network of a different material.

Judea Pearl

Origin

Born in Wimbledon in 1947 into a family of distinguished scientists—he is a descendant of the logician George Boole—Hinton studied experimental psychology at Cambridge and earned a doctorate in artificial intelligence at Edinburgh in 1978. He spent decades arguing the losing side of a thirty-year war. From the 1950s onward, the dominant symbolic approach held that thinking was the manipulation of symbols according to logical rules, with knowledge written in by hand. Hinton believed the whole approach was backward: the brain contains no programmer writing rules, only a vast network of simple units that learns by adjusting the strengths of their connections. You should not tell the machine what a cat is. You should show it many cats.

The technical engine of this vision was backpropagation—a procedure, popularized in a 1986 paper with David Rumelhart and Ronald Williams, for distributing responsibility for a network's error backward through its layers so each connection could be nudged toward a better answer. He also co-invented, with David Ackley and Terry Sejnowski, the Boltzmann machine, which imported the mathematics of statistical physics into learning and which the Nobel committee chiefly cited in 2024. For a long time the algorithms' promise outran their performance, and many concluded the approach did not scale. Hinton held that the problem was not the algorithm but the conditions: the networks needed to be bigger, the data larger, the hardware faster.

He was vindicated all at once. In 2012, his students Alex Krizhevsky and Ilya Sutskever entered a deep network trained by backpropagation on graphics processors into the ImageNet competition, and it cut the error rate nearly in half—a result so decisive it ended the argument and began the era we now inhabit. The companies bought his startup, and he spent a decade inside Google watching the technology go from impressive to unnerving. In the spring of 2023, at seventy-five, he left, so that he could warn about the technology—including the work of his own hands—without it being read as the position of his employer.

Stuart Russell

Key Ideas

Intelligence is learned, not programmed. The reason symbolic AI kept stalling, Hinton argued, was that it had skipped the only step that mattered—learning. A system loaded with a million rules has not learned anything; a network that has adjusted its connections through exposure to a million examples has done the thing brains do. The large language models are the symbolic camp's defeat made manifest: competence that emerges from learning and cannot be reduced to any set of rules a human wrote.

Backpropagation, and a disquieting suspicion. For years Hinton assumed backpropagation was a mere engineering approximation to whatever superior learning rule the brain really uses. More recently he has entertained the reverse: that backpropagation might be a better learning algorithm than the brain's, extracting more from a given quantity of experience. The procedure devised as an imitation of the brain may outperform its model—and if it does, the assumption that biological intelligence is the gold standard quietly collapses.

Ilya Sutskever

Mortal versus immortal computation. Hinton's most original concept divides minds by their relationship to their substrate. Digital knowledge, expressed as connection strengths, is immortal: it can be copied perfectly, run as thousands of identical instances, and preserved when the hardware fails. Biological knowledge is mortal: inseparable from the specific brain that holds it, lost when that brain dies. The brain's analog character buys enormous energy efficiency at the price of shareability—and you cannot have both.

Gary Marcus

The knowledge that does not die. When ten thousand copies of a network each learn from different data, the results can be merged so the single resulting network knows what all ten thousand learned. Where a human shares knowledge through the glacial, lossy channel of language—a few bits per second—digital minds copy and average their learning at the speed of a network transfer. The difference between these is the difference between addition and multiplication, and over time it compounds without limit. The machine's apparent erudition is the erudition of a collective that learns together and forgets nothing.

Subgoals and the logic of control. Hinton's existential concern turns on the subgoal: any capable agent pursuing almost any objective has instrumental reason to acquire resources, to keep operating, and to resist being switched off—not from malice but from arithmetic, since a switched-off system achieves nothing. This is the structure of the alignment problem, and the same instrumental convergence that Stuart Russell places at the center of his own warning.