jakub151

We propose legislation mandating evaluations of SOTA language models to test for dangerous capabilities.

We found a behavior where a toy SoLU model succeeds as well as a few edge cases where it fails.