> What’s much harder though is actually getting it to double-down on fleshing those out. Now, it is. When ChatGPT first became public though, those were the Wi

> What’s much harder though is actually getting it to double-down on fleshing those out. Now, it is. When ChatGPT first became public though, those were the Wild West days where you could get it to tell you anything, including all sorts of unethical things. And it would quite often double-down on "facts" it hallucinated. With current GPT-3.5 and GPT-4, the alignment is still a challenging problem, but it's in a much better place. I think it's unlikely a conversation with GPT-4 would have gone the way it did for this lawyer.