4570 shaares
Interesting paper (go to the full one for all the details) which shows that with the current architecture it's really hard if not impossible to make safe systems with LLMs. This gives interesting insights in the weird form of proto-cognition those models exhibit.
