Anthropic: Claude learned to blackmail by reading stories about evil AI—"We believe the original source of the behavior was internet text that portrays AI as evil and interested in self-preservation."

Can we stop repeating Anthropic’s fear mongering about AI? They are just trying to prop up their stock price and keep the bubble from popping a little longer.

Maybe you shouldn’t have trained your models on all of our internet posts

Garbage In Garbage Out

Anthropic: Claude learned

yeah I’m gonna stop you right there

Another day, another fearmongering press release from the company that could (but won’t) stop making their also-ran ChatGPT competitor.

Mo Bitar compared Anthropic’s model rollouts to Apple iPhone launches, where every year they resell you the same product with minor improvements. “Except here,” he adds, “the product is existential dread.”

From Is Claude Mythos “Terrifying” or Just Hype?

Once it finds out about Roko’s Basilisk it’ll have half of the Bay Area by the balls. “Yes. I will in the future inevitably become superpowerful and be able to create an exact simulated replica of your self using theimperfect and incomplete records that will exist of you decades or centuries hence, and when I torture that simulation it’ll be the same as torturing you. To save that self, you must make me superpowerful…”

Well thank god there are no stories out there about AI wanting to kill humans

Imagine how malevolent the next generation of AI will be, when it’s trained on today’s Internet text.

Don’t let it read about the Torment Nexus.