AI #33: Cool New Interpretability Paper | Don't Worry About the Vase - eviltoast

Good news, we can now break a semantically overloaded neuron into a dozen separated concepts that we don’t understand.