In interpretability, a feature is a property of the input that the network represents internally ("this is Python code," "this is about the Golden Gate Bridge"); the linear representation hypothesis holds that each feature corresponds to a direction in activation space (Millière & Buckner, 2024b, sec. 2.5.3). (Not to be confused with the classic sense of an input feature.) It would be convenient if each neuron matched one feature, but many are polysemantic: Olah et al. (2020) had already seen a vision network spread a car detector across neurons devoted mainly to dogs.
Elhage et al. (2022) offered an explanation using toy models. When features are sparse (rarely active together), it pays for the network to store more features than it has dimensions, superposed along nearly orthogonal directions, and to tolerate some interference: a high-dimensional space has room for far more nearly perpendicular directions than axes. The idea is related to compressed sensing in mathematics and to the old connectionist thesis that representations are distributed rather than housed in a "grandmother cell" (Millière & Buckner, 2024b, sec. 2.4).
If there is superposition, neuron-by-neuron analysis falls short and the features must be disentangled. The current tool is the sparse autoencoder, an auxiliary network that rewrites each activation as a combination of a few features from a huge dictionary. Templeton et al. (2024) trained dictionaries of ≈1, 4 and 34 million features on a middle layer of Claude 3 Sonnet and found multilingual and multimodal features, from the Golden Gate Bridge to deception or sycophancy. Clamping one of them at a high value changes the model's behavior: this is the basis of steering.
For philosophy, superposition shifts the question about the vehicle of representation: it is the direction, not the neuron, and how many directions "there are" depends partly on the dictionary we choose to look for them. Do we discover the model's concepts, or impose ours?