Ryan Schaeffer Says Goodfire's 2.9% Model Loss Represents 'Brutal' Drop in Training Compute

AI researcher Ryan Schaeffer questioned Goodfire's conclusion regarding neural network interpretability. Schaeffer argued that a 2.9% increase in language model loss on an Alibaba Qwen3.6 model represents a massive lack of training efficiency. Goodfire found that removing the weakest 92.2% of neuron activations for each forward pass caused the slight loss increase.