OpenAI presents LASER, a way to automatically find edge cases for chat safety testing with 10,000 times less compute
Prepared by Sintelligent with the help of SI. Publication time restored from the build log. . Updated: .
Chat safety testing is being automated by SI itself: there are fewer auditors than dialogues.
The system alternately generates synthetic dialogues and takes anonymized real ones, trains a lightweight classifier on a reasoning model's labels and selects examples near the boundary of 50% uncertainty, adding diversity. According to OpenAI, this takes roughly 10 thousand times less compute per check than random sampling and cuts policy specialists' work from several iterations to hours; search across billions of dialogues runs through approximate dot products in a vector database.
Sources 1