Long-shot but as we rapidly move with Aquin's Interpretability work, It gets easier to determine and strengthen our systems produce surgical outputs! P.S. might open-source the weight editor.

At Aquin, We integrated 16 benchmarks to measure what's happening inside an LLM! 3 for SAE features: is the feature real, monosemantic, and does the model even use it? 13 for weight edits: does it generalize, survive future edits, show up in free-form generation? learn