Arthur Samuel

In the [YOU] on AI Field Guide

The cycle that begins with [YOU] on AI reaches back to Samuel whenever it needs to ground an abstract claim about machine learning in the moment when the abstraction first ran on actual hardware. Samuel is the founding proof that the things we now argue about—whether machines understand or merely perform, whether self-improvement at superhuman levels is real or mythological, how to narrate a genuine but bounded achievement without inflating it into something it is not—were not introduced by the current generation of AI. They were present on a checkers board in 1959, in the first machine that learned.

Samuel’s career provides the cycle with its most precise account of the gap between a technology’s actual achievement and the legend that attaches to it. The 1962 game against Robert Nealey, which IBM publicized as a triumph over a checkers champion, became the template for a pattern that has repeated in every decade of AI history: a real and impressive milestone wrapped in a description that outstrips it, traveling faster and further than the truth, producing expectations the technology cannot meet, and generating a backlash that overcorrects. Samuel himself reported the win and the limits with the same flat honesty; the institution that employed him did not. The Nealey episode is Clarke’s Law of Revolutionary Ideas in miniature, and it is the most useful thing Samuel’s career has to teach a reader trying to think clearly about AI announcements today.

The deepest thing Samuel did—and the thing the cycle returns to when it needs to locate the question precisely—was to build a system that learned without understanding, and then to not claim otherwise. His checkers program formed a working theory of board strength from nothing but self-play experience and beat the man who wrote it. It had, by every behavioral measure, competence at checkers. And it plainly did not understand the game in the way a human master understands it. The gap between those two—competence without comprehension—is the central mystery of artificial intelligence, and Samuel installed it at the field’s foundation, in the cleanest possible form, before anyone had invented the vocabulary for it.

His founding move—the shift from telling a machine what to do to letting the machine work out what to do from experience—is also the founding move of the current era. The triumph of modern AI is the triumph of learned behavior over hand-written behavior. Samuel performed this shift first, on a checkers board, with a machine whose memory was measured in thousands of words. Everything since has been commentary on that one structural choice.

Origin

Samuel was born in Emporia, Kansas, in 1901 and trained as an electrical engineer, taking a master’s degree at MIT in 1926. He spent the better part of two decades at Bell Laboratories working on vacuum tubes and, during the war, on radar—a practical, hardware-first education that shaped his approach to AI in the most direct possible way. He was not a logician dreaming about the nature of mind; he was a builder of devices, and the checkers project was an engineer’s project: choose a problem with crisp rules, an unambiguous win condition, and a search space too large for brute-force lookup but tractable on a 1950s machine. Checkers was that problem.

He arrived at IBM in Poughkeepsie in 1949, nearly fifty years old, and wrote his checkers program for the IBM 701 and 704. The 1959 paper reporting the work introduced “machine learning” to the scientific literature, described the program’s self-play architecture, and demonstrated that a computer could be programmed to learn a better game than the person who wrote it. He retired from IBM in 1966 and joined Stanford’s faculty, where he continued working productively into his late eighties—including on the documentation of typesetting software, unglamorous work that is itself a corrective to the heroic narrative. He died in 1990, before the neural network renaissance that would make his term a household word. He did not live to see his founding vindicated at the scale it reached. He had, in any case, already demonstrated the principle.

Key Ideas

From programming to learning. The founding structural choice of the field: instead of encoding good play explicitly, encode a capacity to become better at play and let experience fill in the substance. Samuel’s program was given the rules and a vocabulary of features that might matter; the actual weights it assigned to those features—its induced theory of checkers strategy—emerged from self-play. This is, almost exactly, the philosophy of modern machine learning scaled by a factor no one in 1959 could have imagined: the most capable systems are not written but trained, their competence emerging from data and experience rather than from the foresight of any programmer.

Self-play and its limits. Samuel’s most prophetically prescient method was self-play: letting the program generate its own training experience by playing against copies of itself. The method is powerful in any domain with exact rules, full observability, and unambiguous outcomes. It is limited to exactly those domains, and the limitation is precise: you cannot learn to be a good doctor or negotiator purely by playing against yourself, because there is no perfect simulator and no clean signal at the end. The spectacular achievements of modern self-play systems—AlphaZero mastering Go, chess, and shogi without studying a single human game—live in the corner of the world where Samuel’s conditions hold.

Rote learning vs. generalization. Samuel’s program learned in two distinct ways that the field now debates at enormous scale. Rote learning stored board positions and their computed values for lookup; generalization tuned the evaluation function’s weights so the program improved on positions it had never seen. Samuel needed both and observed that they had complementary strengths: rote for recurring positions, generalization for the novel middle game. The question of whether a modern AI system is “merely memorizing” or “genuinely generalizing” is Samuel’s question, scaled to billions of parameters and stripped of the clean separability that his inspectable mechanisms permitted.

The learned evaluation function. Samuel’s central technical invention was a weighted combination of features whose weights were adjusted by experience—a parameterized function whose parameters are learned from outcomes. This is, with no exaggeration, the definition of a modern neural network at any scale. His program had dozens of weights on human-named features; contemporary networks have billions of weights on learned representations. The underlying architecture—a function that takes a situation as input and produces a judgment as output, with behavior shaped by tunable parameters set not by a programmer but by exposure to data—is Samuel’s architecture, scaled past recognition.