alt Hacker NewsThere was also a NeurIPS paper which successfully trained 1-bit models directly without floating point weights or any quantization: https://arxiv.org/abs/2405.16339
I didn't see any further work in this direction however. The paper seems to have been mostly overlooked.
While what can be done with less bits per parameter may be interesting, this is not really the important number for an LLM.
Much more important is the total amount of memory needed by an LLM, which can be decreased either by reducing the number of bits per parameter or the number of parameters. Arithmetic circuits are already cheap so for the inference cost the amount of memory transfers is more important.
It is not clear yet which number of bits per parameter allows the minimization of the total memory requirement, at a given LLM quality.