My understanding is that they are using increasingly small precision on the floats used by LLMs. So we just have to wait for them to reach 1-bit precision.