A conceptor by any other name

According to LessWrong, researchers keep rediscovering the same elegant technique for steering large language models and calling it by different names. It's called a conceptor—a solution to a core interpretability challenge. Here's the problem it solves: you've identified which directions in a neural network's activation space correspond to a specific concept—say, a persona or expertise. You want to adopt that concept without changing everything else. Make the model act like a pirate, but keep it speaking Chinese if the input is in Chinese. A conceptor is what researchers call a soft projection operator. Instead of a binary yes-or-no for each dimension, it smoothly weights them based on how much they vary with the concept. Directions that strongly encode the concept get fully preserved. Barely-relevant ones get mostly ignored. Everything in between gets proportional weight—eliminating arbitrary cutoffs. The post's real contribution: cataloging that this idea has been independently discovered so many times it goes by at least a dozen names across the literature, from soft projection to projection pursuit to factor analysis. A Rosetta stone for anyone who's stumbled into it on their own.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton