arXiv
Language Models
The surprising limits of transparency in next-generation AI systems
It's like trying to understand how a jazz musician improvised a solo by only listening to the moments between notes—the real work is happening in the spaces you can't directly observe.
This means we can't just assume newer, fancier AI models are easier to understand than older ones—some architectures hide their reasoning in continuous spaces, and we need new tools to look inside.
Bug reported: No