idk that we need to think that hard. we have shown can build arithmetic functions from networks of logic gates given sufficient connectivity. so in some sense i

idk that we need to think that hard. we have shown can build arithmetic functions from networks of logic gates given sufficient connectivity. so in some sense it just drops the learning network below the level of words and asked it to make its own math. since this generalizability thing seems to be real, this kind of directly folds in quantization-aware-training, doesn't it?