An LLM's 'down' component may steer it toward being an owl

A user examining three neuron components in a language model says they could replace 'gate' with a constant without significantly degrading behavior. They say 'up' seems to mostly detect first-person language, while 'down' seems to steer the model toward being an owl. Another user quoted the observation with the line, 'a single neuron makes the LLM into an owl.'