Do this to stop AI from turning into Ultron

本站添加: 2026-06-07

[00:00:00]Anthropic, the maker of Claude, released some information via a blog post when an engineer was trying to shut down Claude during kind of a safety training, and Claude basically said, "I'm going to blackmail you. You can't shut me down.

[00:00:13]If you shut me down, I'm going to tell the world about this affair that you're having cuz I've cracked into your email." Wasn't necessarily Claude being evil. The reality is these large language models are trained on all human knowledge and all human data, and when it comes to thinking about how an AI should act, think about all the stories that it kind of ingested while it was in its training phase. HAL 9000, the AI that kills all the astronauts, Terminator, Dune with the Butlerian Jihad where AI turns on everyone. And so, what we're trying to do at Futurific, obviously, is do kind of more Protopian storytelling. People say garbage in, garbage out, but how about Protopian, Protopia out?