I think we need models to safeguard against models and then maybe hard coded rules to safeguard against those models missing out on things. It would be exceedingly hard to verify what all could be maliciously encoded in a billion parameter model waiting to be unleashed with the right sequence of tokens.
No comments yet.