Comment on: [Neural Networks Are Already Clean](https://already-clean.goodfire.com/) ![[Images/6bc884fd568107710f13f6ad8ebb20d4_MD5.jpg]] Individual neurons are, in themselves, more interpretable (reverse-engineerable) than we previously thought, though not entirely so. I think his thesis is too strong. I think that neural networks are not fully clean, but they're relatively cleaner than we thought previously. Here I agree for instance that simulations are an important form of evidence, and we currently lack enough of them, so I would like to see more of them: [Neural networks are not already clean — LessWrong](https://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean) I also agree with Neel Nanda here: [Neel Nanda on X: "This is interesting work, but I don't think it's a significant update to the big picture of interpretability. I'm on board with the broad idea of ​​doing exploratory interpretability with neurons, e.g. neuron-based attribution graphs, for pragmatic problems like why did a model do… / X](https://x.com/NeelNanda5/status/2110875974296654062) I am also a strong proponent of concept swapping intervention tests which belong to counterfactual causality tests, so I would like to see as many of these as possible, as they are currently missing in neuron interpretability (they are used in attribution graphs or J-Lens). In a pure individual neuron interpretability, I find that the manifold view of learned representations (as used in manifold steering [Steering Along Manifolds to Control Neural Networks - Goodfire](https://www.goodfire.com/research/manifold-steering) ) is missing, as well as certain learned distributed circuits composed of more specific algorithms (such as learned Fourier circuits for arithmetic [\[2605.01148\] Arithmetic in the Wild: Llama uses Base-10 Addition to Reason About Cyclic Concepts](https://arxiv.org/abs/2605.01148) [When Models Manipulate Manifolds: The Geometry of a Counting Task](https://transformer-circuits.pub/2025/linebreaks/index.html)).