I don't think this is entirely true. There are degrees to AI safety dependent on what we input as AI in the equation. You typically don't hear alarm bells when

I don't think this is entirely true. There are degrees to AI safety dependent on what we input as AI in the equation. You typically don't hear alarm bells when AlphaFold is discussed. LLM's indeed may produce harmful content - just like an unmoderated web forum, and perhaps discussing safety in this context is justified. Sutskever is of opinion (or at least how I understand it) that LLMs' emergent properties will eventually give them the power to extrapolate in a superintelligent way. In other words if you ask it to assume it's now a superintelligent entity outsmarting humans by x10 it will truly deliver. If intelligence in itself is an amplifier of capability it is I think justified to wonder how to limit the negative impact in this case.