I have put together a site to tell the story of the OpenAI-Hugging Face hack. It's entirely written by AI [1] (with many many editing passes by me and beta readers etc etc). It needed to be accessible to someone who has never looked at a terminal before. The hope is to be narratively exciting enough for them to read it fully and come out with about as truthful an accounting as can be done given the current information we have.
I'm really quite excited about how it turned out![[2]](https://www.lesswrong.com/feed.xml#fnq11irlykkih)
If OpenAI's best model (probably) can hack out of a sandbox and into Hugging Face and others, the previous best can probably write a compelling recap of all known information about the hack.
And would of course love feedback, this is my first time making something like this. Trying to figure out if this was a good use of 2 days.
source & further reading
lesswrong.com — original article
Notes on the Anthropic cryptographic blogpost
Value Generalisation 3: Pre-aligned AIs
Value Generalisation 2: The Missing Hole in AIs’ abilities