The data-locality inversion

The edge-AI story just split into two categories, and only one of them is new. Inference at the edge is the well-known story: a pretrained model is shipped, the chip runs it, the network stays asleep. Carloscodix's Qapla' project belongs to the other one. A model is born where the data is born, on an $8 ESP32-S3, with gradients hand-written in C and a finite-difference gradcheck that lands at 1.07e-08 relative error.

The mechanism is the data-locality inversion. A vibration sensor bolted to a specific pump cannot see that pump's failure modes until it lives on that pump, and there is no internet to fetch a model from. So the model is trained there. Qapla' is a Klingon-language demonstration of a class of device that learns its own environment instead of importing someone else's summary of it.

The Qapla' project README makes the inversion explicit: forward pass, backprop, and weight updates all run inside the chip, with LoRa or short-range radio as the only backhaul. The caveats are real. The parameter count is small, the use cases are framed as "picture this" hypotheticals, and this is not a substitute for serious model training. But the cheap part matters because the data cannot leave, and that is a structure, not a spec.

The on-device model is the one that knows that machine. The off-the-shelf model is the one that knows the world. They are not the same product.