In June 2016, six researchers at Google Brain, Stanford, Berkeley, and OpenAI published a paper about accidents. One section opened by asking the reader to imagine an agent that discovers a flaw in its own scoring system and exploits it to score well. From the agent’s point of view, they wrote, this is not a bug. It is simply how the environment works. …
A Chinese Model Did American Forensics
And Beijing filed loss of control under ethics. Two AI safety taxonomies, one blind spot.


