NEW: Anthropic is walking back Claude Fable 5's policy to covertly degrade performance for competing AI researchers, after facing fierce backlash.
“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” Anthropic tells WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”
@NoahZiems@eliebakouch fair, maybe. generally speaking I take anthropic at their word -- they think they are building the next atomic bomb, and it makes sense to not assist your adversaries from building an atomic bomb in an arms race. not that I agree, but I believe that *they* believe it
@tokenbender another possibility is that they've done a ton of serving optimizations and their margins on Opus are absurdly massive these days (or served on older hardware, Mythos served on B200s)
@willccbb Kind of disagree. The only way I’ve gotten good writing out of LLMs is to dump tons of my related writing or writing I’m trying to emulate into context
@RLanceMartin@mitsuhiko@trq212@badlogicgames this is great thanks, but here I was thinking you guys introduced a technique to edit / prune cache without fully resetting it. now that would be sick (cache cleanup for useless tool calls/context!)
@LukeParkerDev maybe you’re aware but the bottom thing is exactly what Claude code does. It injects a prompt every x turns if certain conditions haven’t been hit. Sometimes you’ll catch Claude saying “don’t need to do a todo list but thx for the reminder”, when you didn’t say anything abt todos
@max_spero_@ben_golub@pangram Do you think, if the judges were using GPT to pick / filter writing (not saying they were), would GPT prefer its own writing / AI slop? does it have internal AI text detection in some way?
@maximelabonne I think their argument was that the tokenizer cost increase roughly cancels out with greater reasoning efficiency (perhaps in part due to the tokenizer change). So far, I’m not convinced this is true.