This reminds me of the latest Ken Stanley paper on representations.
“Questioning Representational Optimism in Deep Learning: The Fractured Entangled Representation Hypothesis“
https://arxiv.org/abs/2505.11581