This is the best AI news #blackbox #AI

For as long as I’ve worked in AI, we’ve called it a black box. New research from Goodfire suggests that’s starting to change. The classic example is a neuron in a vision model that fires for both cats and cars. It looks like nonsense, the kind of thing that made us say we don’t know why AI does what it does. Look closer, though, and it makes sense at a different level: both share the same pointed shape, and that shape is what the neuron is actually tracking. The AI isn’t random. It just doesn’t think in our categories. In an early audit of a language model, the researchers estimate that around