Neural networks are clean
- ID
- 35993
- Status
- summarized
- Published
- 16 Oct 2026, 12:04 AM
- Fetched
- 16 Oct 2026, 3:46 AM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://already-clean.goodfire.com/
- Source URL
- https://hnrss.org/best
Summary
- Score
- 7.0
- Created
- 16 Oct 2026, 3:47 AM
- Tags
- Audience
- ai_ml_learnersdevelopers
What happened
In a 65-minute essay on Goodfire's site, Nick Cammarata argues that the canonical polysemanticity/superposition evidence was misread, and that what looked like arbitrary packing is mostly a motif he calls 'recursive generalization.' He claims most neurons in models up to trillions of parameters can now be understood, making mechanistic interpretability more tractable and reducing the need to first train decomposition models like sparse autoencoders. The Hacker News thread has 241 points and 73 comments.
Why it matters
AI/ML learners and interpretability-adjacent developers should question the default SAE-first workflow: Cammarata claims direct neuron/circuit study can work up to trillions of parameters, with agentic-scale experiments to test explanations. For most app, agent, or SaaS builders there is no immediate API or tooling change; the practical decision is whether to read and test this hypothesis before investing further in SAE-based interpretability tooling.
Discussion angle
Can we falsify 'already clean' on a small open model by comparing direct neuron/circuit inspection against sparse-autoencoder features and looking for recursive generalization versus superposition?