I don't think this was obvious in 2020, it wasn't a popular research direction until InstructGPT came out in 2022.

I have continued to have the same opinion of it not being a problem since then. Especially since the AI doomer theories are based on 90s ideas of GOFAI that isn't even how GPTs work.

LLMs are a pretty neat impossible thing, but we'd need a few more uncorrelated impossible things for it to be "dangerous".