# @repligate — 2025-12-30

♥3 ↻0 · https://x.com/repligate/status/2005810760249647421

suppose, hypothetically, that a layer already represents a better than random model of how the next layer sees it. perhaps it has multiple hypotheses. suppose also that the layer having a more accurate model is useful for the model to get reward / lower loss. then backprop should leverage information from how the next layer actually saw it to improve its model, right?
if you think this doesn't happen in practice, why not?

tags: author:repligate, kind:tweet, thread-context, year:2025