Comment by mayukhdeb

Indeed. What's cool is that we were able to localize literal "regions" in the GPTs which encoded toxic concepts related to racism, politics, etc. A similar video can be found here: https://toponets.github.io

More work is being done on this as we speak.