Anthropic taught an AI judge sixteen rules and trained a less harmful chatbot
Bai et al.'s 2022 paper measured that edge as a preference margin from 8,135 crowdworker harmlessness comparisons on Anthropic's own 52B models, the size behind a term now used far more loosely.
An image generator turns "six apples" into "some apples" before it draws a single pixel
Accuracy holds up for the first handful of objects, then collapses the same way in system after system tested.
Ten fine-tuning examples stripped Llama 2's safety training
RAND built a study specifically to test whether access to a language model helps plan a biological attack, finding no measurable difference between the teams that had one and the teams that didn't.
An ensemble beat SuperGLUE's human score before DeBERTa, the model credited with doing it first
DeBERTa's single model crossed the human score of 89.8 at 89.9, after Google's T5 + Meena ensemble had already cleared it at 90.2.
Bard's launch demo credited the James Webb telescope with a discovery ESO made in 2004
Google published the claim in its own promotional post two days before Bard's Paris launch, when the real first photograph of a planet beyond our solar system was already eighteen years old.