A New Math Trick Could Make AI Faster and Cheaper

Every AI system learns by adjusting millions of tiny settings, like dials on a giant mixing board. The standard way to figure out which dials to turn is called "backpropagation" — essentially retracing the machine's steps backward, one layer at a time, to see where things went wrong. It works, but it is slow and it burns a lot of electricity. It is the engine behind nearly every AI product you have ever used.

A new paper by researcher Manon Verbockhaven, posted on the preprint site arXiv, proposes a different route. Instead of retracing every step, it builds a "skeleton matrix" — a compact map of all the pathways through the network — and uses that map to get the same answer directly. Think of reading a subway map instead of riding every train on the line. She proved mathematically that this shortcut is correct, and released free Python code so others can try it. In her experiments, it ran notably faster than the usual backpropagation method.

Why should you care? AI runs on expensive chips and enormous amounts of electricity, and all of that cost eventually shows up in your subscription fees, your phone's battery life, and the price of the gadgets that include AI. If the core math gets faster, AI becomes cheaper to run. That could mean lower prices, snappier features, smaller devices that can think for themselves without phoning home to a data center, and less strain on the power grid.

The honest catch: this is a theory paper, not a product. It focuses on one specific kind of network — the straightforward, feed-forward type with a common activation style known as ReLU — and the tests were on relatively small models. It has not been shown to work on the giant systems behind chatbots and image generators. There is no company and no launch date. Still, research like this often quietly shows up inside real tools years later.