Games, weather, chips: superhuman where the reward is clean.

Here's a rule that predicts where AI goes superhuman and where it stalls. Give it a clean score — a cheap, trustworthy way to tell a good answer from a bad one — and let it try a million times.

The one idea

Narrow superhuman AI follows one recipe — a clean, cheap reward plus self-play or simulation — from AlphaGo/AlphaZero/MuZero to weather, chips, and faster algorithms; remove the clean score and the trick evaporates, which is why open-ended judgment stays hard.

Wherever that holds, machines don't just match us. They leave us behind. The clearest example is a board game. In 2016, DeepMind's AlphaGo beat Lee Sedol, a top Go player. Go has more positions than atoms in the universe; experts thought superhuman play was a decade away. One move, move thirty-seven, was so alien it stunned the professionals. Then it got purer. A year later, AlphaZero threw out human games and learned by playing itself from scratch — surpassing every program at Go, chess, and shogi. MuZero went further, mastering games without even being told the rules. The engine is always the same: self-play, plus a clean win-or-lose signal.

How it works — the demo

A Go board and move 37; a self-play spiral climbing without human data; the pattern jumping the fence to a spinning weather globe, a chip floorplan, and matrix-multiply gears; the reward-beacon and the fog where it fails.

Now the pattern jumps the fence into science. Weather forecasting: models like GraphCast now match or beat physics-based forecasts on many of the tested measures — and run in minutes, not hours on a supercomputer. Same shape as a game: a clear target, cheaply scored against reality. It keeps generalizing. Designing the floorplan of a chip is a brutal search problem, and reinforcement learning has been used to help lay out real silicon. AlphaTensor discovered faster ways to multiply matrices — the operation under all of this. The common thread: when the objective is crisp — speed, area, correctness — search wins. Now the shadow side. Every one of these wins needed a cheap, trustworthy score. Take it away — was this essay wise? was this decision fair? — and the superhuman trick evaporates.

The trap to avoid

Assuming a superhuman game-player implies superhuman judgment — without a cheap, trustworthy score there's nothing to search against.

Why it matters — and what’s next

There's nothing to play a million times against. Clean reward isn't a detail. It's the entire game. And that lands us back on the spine. This is why the course keeps returning to verification and reward. The frontier of chatbots is the hard version of the same rule: turning fuzzy, open-ended judgment into a signal you can check. That's the map — four kinds of AI the headlines skip. Back to the main line.

This is a supplement in AI: Zero → Frontier — a side-trip that deepens the act it sits beside, one file and one loop at a time.