I predict that even less-than-1-bit viable quantization methods will be found eventually. Of course, the bits-per-weight figure should be interpreted as somethi

I predict that even less-than-1-bit viable quantization methods will be found eventually. Of course, the bits-per-weight figure should be interpreted as something average: e.g., group-quantizing 32 weights into 24 bits would be 0.75 bits per weight.