Why individual neurons rarely mean one thing
If interpretability were as simple as 'each neuron represents one concept,' the field would be far more advanced already. In practice, when researchers examine what makes an individual neuron activate, they often find it responds to several unrelated concepts rather than one clean idea. This isn't a training failure, it's a predictable consequence of a network having far more distinct concepts to represent than it has neurons to represent them with.
The proposed explanation is superposition: a network represents more features than it has dimensions by encoding each feature as a direction in activation space rather than as a single dedicated neuron, and letting many of those directions overlap. Because any two features are rarely relevant in the exact same input at the exact same time, the network can tolerate the interference this creates and still recover the right answer most of the time, effectively trading a bit of accuracy for the ability to pack in far more concepts than raw neuron count would allow.