My anons are pretty convincing. Too bad normal GLM output is both really good and really bad at the same time. I feel like it can sort of be compared to how people explained image gen with QWEN locally. Where prompt (in this case immediate preceding text) is a far bigger influence on output then seed. I find spamming rerolls gives me the same ideas and concepts just worded differently unless I significantly adjust the last 1-3 lines. Upping randomness to over 2.0 helps but introduces incoherent replies as well.