Chief among them is having built-in UI for “base-model Claude 3.5 Sonnet” and Llama 405b-base continuing whatever comment or post I am in the middle of writing
I was extremely surprised to read that Anthropic is giving out access to base models to outside parties. Especially as a single throwaway sentence in a giant post. What were the terms of your agreement with them? Do they do this with other people? Do they also give certain people access to the helpful-only (i.e. not necessarily harmless or honest) post-trained models, or just the base pretrained ones?
I was extremely surprised to read that Anthropic is giving out access to base models to outside parties. Especially as a single throwaway sentence in a giant post. What were the terms of your agreement with them? Do they do this with other people? Do they also give certain people access to the helpful-only (i.e. not necessarily harmless or honest) post-trained models, or just the base pretrained ones?
Habryka is slightly sloppily referring to using Janus’ ‘base model jailbreak’ for Claude 3.5 Sonnet
Oops, I thought I had added a footnote for that, to clarify what I meant. I shall edit. Sorry for the oversight.