I saw a paper/article on hacker news at one point about making LLMs where they did not use floating point gpus to do their calculations so you wouldn't get the non-deterministic problem (ask same question get diffrent response). How is that going? …