"Already Clean" Neural Nets Might Enable Continual Learning

I’ve been digesting Nick Cammarata’s new essay at Goodfire ("Neural Networks Are Already Clean") and thinking about what it implies for model size and continual learning rather than just safety. Four findings in the essay stood out to me: Neurons act like bipolar (+/−) control-flow switches rather than static noun labels. Outside of early "dial" layers, the weakest 75% of activations can be zeroed out with almost no accuracy loss. Foundational building blocks recur across very different models (e.g., Qwen3-32B and Claude 3.5 Haiku converging on the same 4-neuron hub for preeclampsia). Memorized facts sit right at the output and route through existing circuits (like the 9-step umbrella-to-bee experiment, or clipping 99.9th-percentile activations in Qwen3.6 to erase verbatim recitation while keeping reasoning intact). The speculative question: If models converge on a Spelke-like "semantic skeleton" of bipolar switches, could we encode specific facts as sparse ternary recipes (+1 for positive pole, −1 for negative pole, 0 for inactive) and store them outside the model as hypervectors in a Vector Symbolic Architecture (VSA)? In theory, a much smaller, frozen model could translate inputs into that vector space and query the external memory—meaning adding or deleting a fact is just adding or deleting a vector, without retraining. I know there are real obstacles here: early layers still need continuous signal strength, memorized facts are wired through existing semantic routes that might not pull apart cleanly, and multiple models would have to agree on a standard coordinate registry. I put together a short write-up here: [https://markwaltz.substack.com/p/already-clean-neural-networks-might](https://markwaltz.substack.com/p/already-clean-neural-networks-might) For folks who have worked with hyperdimensional computing/VSAs or mech interp: is the boundary between the "semantic skeleton" and late-layer memorized facts actually clean enough to pull apart like this, or does the routing tangle them too tightly?