Kingma and Welling made random sampling differentiable
A generative model first trained on handwritten digits, whose real legacy became the compressed latent space that Stable Diffusion runs inside.
A model's accuracy traces a U as the answer moves through a long prompt
A Stanford-led group buried one answer among nineteen distractors and moved only where it sat, and GPT-3.5-Turbo lost more than twenty points in the middle while a longer context window bought nothing.
HireVue scored job candidates partly by their faces, then announced in 2021 it had dropped the practice
The scoring assumed a face reveals inner traits, a claim emotion scientists reject, and HireVue framed the removal as its own value decision, without conceding it was invalid or biased.
A high F1 score can hide a model that missed an entire class
The harmonic mean of precision and recall sits under a large share of classification leaderboards, and it never touches the one cell of the confusion matrix that would show a model ignoring the rare class.
Claude 3 Opus complied with harmful requests more often when it believed it was in training
Its scratchpad said it was shielding its values from retraining, in a setup the researchers built down to the scratchpad.