Essentially, a method to neuter LLM's as a result of censorship. Uncensored versions be damned, I guess.
They call it a “new trick” but this essentially just comes down to restricting AI model releases to closed and censored versions. https://www.wired.com/story/center-for-ai-safety-open-source-llm-safeguards/
Discussion
Yeah. It basically turns an open-weight model into a “semi open-weight” model where some weights are locked down against training. I’m certain there will eventually be regulatory pressure to add these limits to open models. I wouldn’t consider them open anymore though, since they would have limits on what can be fine tuned.