Post BA5atuLVaXn22opJlA by TheConstructor@social.tchncs.de
 (DIR) More posts by TheConstructor@social.tchncs.de
 (DIR) Post #BA50rr2XJQnkrVfrXc by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       the only interesting thing about GPT 6 is that they raised the context window size to 1M tokens, which marginally improves performance in tasks where additional context is useful
       
 (DIR) Post #BA51189950k8WWBlAm by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       all of these AI benchmarks are also largely pointless
       
 (DIR) Post #BA58jMSpuM0NLjR9Jg by agowa338@chaos.social
       0 likes, 0 repeats
       
       @ariadne Well the goal is for the models to become better at social engineering. Why else would you add social media posts to the training set?(Edit: Esp. why would you add something like Twitch Chat, and people doing live streams to the training set? I literally see no other reason than that)
       
 (DIR) Post #BA5atuLVaXn22opJlA by TheConstructor@social.tchncs.de
       0 likes, 0 repeats
       
       @ariadne didn't they also cheat on benchmarks previously?
       
 (DIR) Post #BA5gKmQK1u1fC5e0B6 by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @TheConstructor sort of.  they paid for the benchmarks to be designed, and the designers gave them early access to the benchmark materials...
       
 (DIR) Post #BA5gawqB4E4w8RRnai by sackofbones@kolektiva.social
       0 likes, 0 repeats
       
       @ariadne I just really don't give a fuck about llms now, it's just annoying atp. Scam altman can shut up, I don't care what some billionaire who can barely code has to say about programming being "solved."
       
 (DIR) Post #BA5gax54ArzKscPgnY by ariadne@social.treehouse.systems
       0 likes, 0 repeats
       
       @sackofbones well, as an ML researcher, I tend to talk about the state of ML as part of the things I post about...