QLabs

A new research paper from QLabs introduces Dust, a method for pretraining transformer models without using backpropagation, the algorithm that underpins nearly all modern deep learning. The claim is drawing attention because backpropagation is considered essential to training large language models, and any workable alternative could reshape how AI models are trained and at what cost. Details of the method and its benchmarks are published on the QLabs research site.

A research project called Dust claims a method for pretraining transformer models without backpropagation, the algorithm at the core of modern deep learning. If the results hold up, the approach could challenge assumptions about how large models must be trained, but independent verification and details of its performance are not yet established.