Post #4443674

I watched the OpenAI talk. It's gross. I guess I'll write it up. They are pouring gasoline onto the wooden floors of a building. Lighting a match and wondering if the neighbors have smoke detectors. They ran the evals with reduced cyber refusals relative to production. They open the talk explaining that models really like to cheat. The sandbox let the agent reach a writable, internet-connected, shared service with unauthenticated endpoints. Every flammable condition was a chosen configuration. Their word is inadvertent, but inadvertent wasn't their setup. Lighting a match in a room you already soaked and calling the result a surprise, is like calling Grover Shoe Factory a surprise. Then they stupidly end with a call for better alarms and sprinklers, everywhere, for everyone. The one thing absent is stop soaking your own floors with gasoline and lighting a match. https://www.flyingpenguin.com/tesla-lied-about-autopilot-court-testimony-shows-systematic-violation-of-basic-engineering-principles/