Whoever else is in the batch settles a near-tied logit
At temperature zero the sampler is off, and a GPU still sums the model's math in an order that depends on which other requests share the batch.
RAND's red-team study found no measurable AI advantage in bioweapons planning
Forty-five red-teamers spent seven weeks testing whether AI access improved biological attack plans and came back with a null result, even as two labs raised their own safeguards without similarly definitive evidence.
DeepSeek trained a model to reason without ever showing it a worked example
DeepSeek's January 2025 paper proved reinforcement learning alone can teach a model to reason at length, and the $5.6 million price tag attached to that result belongs to a different model.
Gemini 2.5 Pro (Experimental 03-25) fell from 92% to 87% when the exam turned over
With only 15 problems a year, one of them is worth 6.7 points of any headline AIME score.
A ranked list of candidates was enough to arrest the wrong man
Robert Williams spent 30 hours in a Detroit jail cell over a shoplifting charge built on a facial-recognition match and a photo lineup shown to a witness who wasn't at the scene.