🌈 Research blog: thesephist.com/posts/prism/ Text embeddings contain semantic and structural features that we can now automatically discover with SAEs, and use to build rich, interactive new information interfaces. In this write-up on my SAE experiments and a prototype called Prism, I share everything I learned about training text embedding SAEs and what kinds of features they contain. I also share: 1. A new more precise feature intervention/edit method using inference-time gradient descent to minimize interference. 2. How applying warmup to the sparsity penalty leads to more stable early SAE training. I also talk about two primitives that may be a part of future AI interfaces: detailed decomposition of concepts and styles from an input, and precise steering of high-level semantic edits to text (and in the future, any media). A lot of this work was done Dec-Feb, and in the intervening 5 months there's been so much new momentum in the space particularly around applying SAEs to production models and multimodal models like CLIP. But I think this piece has a unique focus on how interpretability enables new interfaces, and its application to text embeddings.