Then perhaps a method emerges out of this to make training faster (but not inference) - do early training on highly quantized (even ternary) weights, and then s

Then perhaps a method emerges out of this to make training faster (but not inference) - do early training on highly quantized (even ternary) weights, and then swap out the weights for fp16 or something and fine-tune? Might save $$$ in training large models.