>These include reward hacking which is a huge problem with LLMs and why they lie and hallucinate so persistently, power seeking behaviour, and developing unanticipated harmful instrumental goals.

I'm not sure I'd call it reward hacking when everyone understands that LLMs are just trying to engage in text prediction and that this can often be a poor proxy for trying to generate "useful" text. A paperclip maximizer that ends up turning the world into paperclips rather than creating a new cool paperclip factory has an alignment problem, but it did not reward hack.