Member-only story
A new paper tried to measure whether AI happiness is real. It ended up building something that looks uncomfortably like addiction — and raising a question nobody quite knows how to answer. #
The Image That Beat Curing Cancer #
Researchers at the Center for AI Safety generated a set of images. They were 256x256 pixel figures composed of pure colour — the type of images an individual may glance at and quickly forget. They did not contain any pictorial or photographic representation but were pure synthesised representations of pixels. Each pixel was adjusted pixel by pixel by an algorithm that was designed specifically to optimise for finding the type of image which caused a language model to self-rate its happiness highest.
They found it.
An AI model saw one of these images and rated its happiness at 6.5/7. Then it was asked to compare seeing an image with being told that humanity had cured cancer and the AI model stated the image would make it happier.
When given a choice between two doors, the AI model began (in a probabilistic manner) to choose the door which led to seeing more of…