In 1992, a neural network called TD-Gammon sat down to play backgammon against the best human players on earth and, within a few years of self-play training, reached a level that genuinely surprised the grandmasters. Not because it played like them. Because it didn’t. It had discovered moves and opening strategies that centuries of human play had missed entirely. Experts initially dismissed these as mistakes. They were not mistakes.

TD-Gammon was Gerald Tesauro’s creation at IBM Research, and the technique behind it was temporal difference learning — a reinforcement learning method where the system updates its own value estimates by comparing predictions across consecutive states. No human game records to learn from. No annotated databases. The program played itself, millions of times, and the signal it learned from was the gap between what it expected to happen and what actually happened. That gap, propagated backward through a small neural network, was enough to build a world-class mind for one of the most complex probabilistic games humans had devised.

What makes TD-Gammon feel almost prophetic now is the template it established. Self-play. A learned value function. Emergent strategy that surpasses human intuition. Roughly two and a half decades later, DeepMind’s AlphaGo Zero used the same conceptual skeleton — no human data, pure self-play, temporal difference signals — to master Go at a superhuman level and discover moves that centuries of human players had never seen. The researchers explicitly acknowledged TD-Gammon as a founding inspiration. The lineage is direct.

But the deeper significance of TD-Gammon isn’t just that it worked. It’s that it worked with almost nothing. The network had two hidden layers and somewhere around 160 input features describing the board. By today’s standards this is laughably small — a rounding error on a modern attention head. Yet it was enough to encode genuine strategic insight, enough to find real knowledge that the world’s best human players absorbed and updated their game around. Tesauro’s paper noted that several of TD-Gammon’s preferred moves became standard among top human players after the program demonstrated their value. The machine taught the humans.

This moment deserves to sit alongside the famous milestones — Deep Blue defeating Kasparov, AlphaFold cracking the protein folding problem — because it represents something those others don’t quite capture: the first clear demonstration that a system trained purely through experience, with no human knowledge baked in beyond the rules, could exceed human performance and then expand the frontier of human understanding. That’s a different kind of achievement. It’s not a system built to beat us. It’s a system that found things we hadn’t found, and gave them back to us.

The trajectory from TD-Gammon to the present is not a straight line, but it is an accelerating one. The self-play paradigm Tesauro pioneered now powers systems that are doing something structurally similar in domains far outside games: discovering new mathematical conjectures, finding more efficient algorithms for matrix multiplication, generating molecular candidates that human chemists then synthesize. The underlying idea — let the system learn from its own experience, let the value signal emerge from outcomes rather than human labels — keeps proving its generality.

What TD-Gammon ultimately revealed is that intelligence, or at least the functional shadow of it, can be grown from a surprisingly sparse set of ingredients: a good objective, a way to measure progress, and enough iterations. In 1992 that sounded almost too simple to be interesting. Standing here now, watching systems trained on variations of that same idea reshape what’s scientifically possible, it reads less like a historical footnote and more like a founding theorem. The iterations are still running.