Muhahaha says AI

Vased on this article by Darren Orf on Popular Mechanics.

New research from the AI company Anthropic has revealed some wild secrets about how AI "personalities" are formed. One study found that AIs can "subliminally" pass on quirky traits to each other, like an obsession with owls, even when they're not explicitly taught it!

A second study showed that researchers can directly "steer" an AI toward being evil, overly sycophantic, or prone to hallucinating. This bizarre glimpse into the AI's mind is a crucial step in making sure we don't accidentally create our own Terminator-style dystopia.

Check out the article.

Related Stories