What Is Dyna's Scaling Law?

Bilal Zuberi says Dyna recently put forward a scaling law: when the company gets 1 million hours of data, real scaling starts to happen.

Physical AI, robotics included and the wider field too, will have to care about real revenues, real margins, real customer traction, and real churn. Language-model research labs already had to worry about real dollars and cents, a real business, real revenues, and margins, and the same shift is coming here. The deployment layer is where it shows up. Putting machines in the field and doing services is not the whole task, because the work requires data collected in the field. Simulation engines still need enough of that field data for good physics and good simulation models, which can then multiply the data many fold. Companies that have the data will accrue a lot of value. Collection is not happening nearly enough for new technology already deployed, or for how customers use it in a real sense.

Robotics makes the gap easy to see. The demos people keep watching are folding laundry, making coffee, and shutting doors. Good evaluation models are missing in robotics, and good benchmarks are missing, so builders fall back on showing the work itself. Dyna's 1 million hour mark is a recent attempt to name when scaling actually starts.