When you type a question into a chatbot, you might wonder how it arrives at the answer—and it turns out, so do those who built AI.
A host of “interpretability” startups are trying to figure out the inner workings of models in order to make them safer and curb their tendency to hallucinate. The $1.25 billion-valued AI research lab Goodfire has made a recent breakthrough: AI models use shapes to represent concepts, also referred to as “neural geometry.”
Any fix must nudge the model based on the shapes it prefers. “If you don’t respect this kind of geometry … it generally just makes it dumber,” says Goodfire cofounder Tom McGrath.
This is a published version of the Forbes Daily newsletter, you can sign-up to get Forbes Daily in your inbox here.
First Up
