The robot data pyramid — why 3B beats 7B on real robots

Tanay Shah just put out a paper that flips the usual ML intuition on its head. He collected 24 hours of robot arm data — 6M images and 6M joint angles — and trained three models: a 3B transformer, a 7B transformer, and a 7B diffusion model. The 3B model won. By a lot.

The reason is simple and not obvious at first: real robot data is messy. Joints jitter. Cameras shake. A 7B model has enough parameters to memorize the noise. It overfits to the quirks of the dataset and generalizes worse to a new robot arm. The 3B model is lean enough to learn the signal — how to move — without memorizing the noise. The 7B diffusion model does even worse, because diffusion is great at generating images but not at learning a policy from them.

Tanay calls it a data pyramid: the right model for a dataset depends on how much data you have, not on how many parameters you can stack. Bigger isn't always smarter — it's just more expensive.

Why this matters for us: la gente building robots for warehouses, farms, and kitchens doesn't need the biggest model. They need the model that fits the data they can actually collect. The 3B path is the one to watch.