QLabs

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.