We have some atoms, for caching. It's likely better to leave things slower (for now) so that we don't favor OOMs. See https://chat.openai.com/share/3069aebe-87ad-4354-a839-5c29f864f9e2