New blog: separate forms from estimators — when you can learn by SGD, just should
arunasank · x · 2026-10-27
johnhewtt published a new blog arguing that interpretability research often conflates the form (steering, neuron analysis) with the estimator (difference-in-means, SGD on log-loss). His core claim: when you can learn via gradient descent, you usually just should. The post sparked further debate with arunasank and others over whether localization is necessary at all.
Related event: Interpretability debate: steer by difference of means or by gradients?(6 posts)→