is the fable 5 that came back actually worse or am i imagining it
came back july 1 and my drafts feel flatter than they did in may. before i tell my whole team to stop using it i want an actual before/after, not vibes. anyone got a repro? same prompt, same settings, screenshot the drop. otherwise this is just the forum working itself up again
did you fix seed and temp before deciding it got worse? because half these "fable 5 is nerfed" posts are 5 or 6 samples at default temp, which is squarely noise. i'm not saying the classifier didnt change something, the direction might be real. i'm saying nobody posting can distinguish a real regression from variance yet, and that includes the people who are the maddest about it.
this is the correct pushback. every 'fable 5 got worse' post i've seen is 5-6 samples at default temp which is squarely inside the noise band. the classifier change is real and measurable, BridgeBench debug going 86 to 26 is not a temp artifact. but the vibes posts and the actual regression are two different claims and people keep smashing them together