New repo: https://github.com/google-deepmind/dangerous-capability-evaluations. (I haven’t read it.) I support sharing evals like this to (1) enable external scrutiny and (2) let others adopt or improve on your evals. Yay DeepMind. Hopefully it’s not too costly or downside-y to share more evals in the future.
New repo: https://github.com/google-deepmind/dangerous-capability-evaluations. (I haven’t read it.) I support sharing evals like this to (1) enable external scrutiny and (2) let others adopt or improve on your evals. Yay DeepMind. Hopefully it’s not too costly or downside-y to share more evals in the future.