What's this about?
A paper arguing that neural networks develop internal structure corresponding to symbolic relations — that what looks like smeared noise across layers contains distillations of conceptual relations, including grammatical ones, that can be described mathematically.
Why does "symbolic" carry so much history here?
Because it names the field's oldest argument. Symbolic AI built systems from explicit rules and representations and stalled; connectionism learned distributed representations from data and won. The claim that symbolic structure *emerges* inside the connectionist system would mean neither side was quite right — the representations are learned rather than programmed, but they are not the unstructured statistical soup the strong version of the connectionist story implies. It is also the precondition for interpretability actually working.
What's the community saying?
- The most useful comment is someone honestly summarising what they understood and asking whether they got it right — and getting a substantive reading list in reply.
- The scepticism that matters: does this hold generally, or in the examples the authors found? The thread does not resolve it.
- The best joke in today's digest, in reply to a comment about distillation: "distillation is now illegal tho."