3 ms·

I had an interesting conversation with ChatGPT yesterday. I asked what the heredoc operator was in Terraform that removes white space. It replied that I should

I had an interesting conversation with ChatGPT yesterday. I asked what the heredoc operator was in Terraform that removes white space. It replied that I should use <<~ as the operator and gave an example. I tried it and got an error. A bit of Googling and I found the correct operator is <<- with a hyphen and not a tilde. I told ChatGPT and it apologized and said I was indeed right. Then it proceeded to give me another example, again with the incorrect operator as <<~ to which I replied again with the correct one. The next example used <<~- as the operator. Clearly incorrect, but it was interesting that it was learning on the fly. Another couple attempts and it started to respond correctly with the <<- operator in the explanation and examples. I was really surprised that it learned so quickly on the fly. I hadn’t thought of the implications until I read this article. Can LLMs be gaslighted in realtime?

fnordpiglet 4y ago
To my understanding the “chat” part means the entire context of the conversation becomes a conditional to the next prompt (forgive me if I used the wrong term), so in a sense it can be trained in the convo. You can definitely gas light it, including building up a context that subverts it’s guard rails. A lot of focus is on the one shot subversion but you can achieve it too more subtly.