Aside from Meta is there any reason to think the big AI labs are still using LMArena data for training? The weaknesses are well understood and with the shift to RL there are so many better ways to design a reward function.