All That Machinery Was a Workaround

When you give someone directions, "turn left at the bakery" works no matter where they're standing. "Go to 40.71 north, 74 west" does not, until you've agreed on a starting point. One instruction describes a move; the other names a fixed spot. It feels like a quirk of language. It isn't. Nearly every artificial intelligence that reasons about space — a robot arm reaching for a cup, a model folding a protein, a self-driving car reading a street from any angle — picks the second style. It hands the network a list of numbers for where each thing sits, then, in a separate step, tries to teach the network to handle geometry correctly. Describe space first; make the description behave second.

That second step is where the cost hides. A scene shouldn't mean something different just because you turned the whole thing a quarter turn; the answer has to come out the same. Getting a network to honor that obvious fact has worked, but only by piling on machinery — scaffolding bolted on after the design was done, because the original never accounted for it. The scoring mechanism alone can eat thousands of a network's adjustable settings, each one a layer the design never planned for. And even then, certain distortions — the stretch and skew of scaling and shearing — stayed out of reach.

Skip the scaffold, though, and things fall apart. Hand the network bare position numbers with no special machinery and it never learns the rule at all: rotate the scene and its answers don't drift a little, they shatter, off by many orders of magnitude.

A new paper finds a third way, and it sounds too small to matter. What if the transformation were the thing the network processes, instead of a description of where something sits? The geometric move itself — the turn, the slide, the tilt — rather than a list of numbers standing for a position. When two of these meet inside the attention step, the part that asks "how strongly do these two relate?", the answer is simply how far apart the two moves are, read straight off the math. It's the difference between measuring the gap between two dancers and the gap between two dance steps. Symmetry stops being a rule you enforce from outside. It's built into what the pieces are.

The payoff is stark. This approach matches the heavy-machinery method's near-perfect symmetry — the expensive thing's whole reason for existing — while using fifty to eighty times fewer settings for that core comparison. Dozens of numbers where the old way needed thousands, and none of the scaffold. It even handles the stretch-and-skew distortions the older tools could never touch.

The proof is small-scale — tidy, controlled sequences, not a robot in a real kitchen — and it has real limits: it can stumble near certain extreme rotations, where the underlying geometry turns ambiguous, and it only handles relationships that run both ways. None of that dents the idea underneath. A large share of the complexity in this corner of AI was never a law of mathematics. It was the downstream tax on one early choice. Call the move "data about a thing," and you inherit a hard problem. Call it "the thing," and the problem was never there.