I built HyperSAE, an open-source interpretability tool that extracts what an LLM "knows" and organizes it into tree-shaped concept maps. The idea: LLMs learn concepts hierarchically. "Programming" contains "Python" contains "list comprehensions." But current interpretability tools dump everything into a flat, unstructured list of thousands of features with no organization. HyperSAE uses a branch of geometry where space expands exponentially (like tree ...