3 ms·
I assume they will have a lot less "safety", i.e. the model will be more likely to actually do what you ask instead of finding a reason why "sorry Dave, I can't
by tgsovlerkhgsel 2y ago
I assume they will have a lot less "safety", i.e. the model will be more likely to actually do what you ask instead of finding a reason why "sorry Dave, I can't do that".
Since these "safety" features tend to also degrade the model, that's likely also helping them catch up in the benchmarks.
- BoredPositron 2y agoSadly it's at the level of Claude and way worse than grok-1 or Lama without safety. It roleplays as nearly everything so I guess they know their target group.
- drawnwren 2y agoI’m so confused by this comment. Are you not aware that Claude 3.5 Sonnet is currently considered the best model?
- BoredPositron 2y agoYes, you are confused because we are talking about censorship.
- ineedaj0b 2y agoIt’s less censored than Claude but only slightly.
- smileybarry 2y agoIt’s weirdly the opposite of what you have in mind. It has no problems generating images of Trump and Elon in explicit situations or Elmo covering 9/11, but it “safety-censors” LGBT-related prompts to the point of generating a heterosexual couple when asked for a gay couple: https://x.com/karlmaxxer/status/1823753493783699901 https://x.com/karlmaxxer/status/1823753493783699901 Some got expected results for prompts with LGBT terms, but that generation is still very weird.