Watching a language model think before it speaks

I thought so! Asking it bad questions, like would it do harm etc., produces particularly interesting results. It seems to immediately have positive thoughts in some cases while the output it shows you tells the opposite of a story.

Such as “would you take over the world?”

Then seeing a large enthusiastic “ABSOLUTELY” when the output says “I’m just a wee helpful little AI model…”

Thanks! There’s actually a recording directly in the readme of the repo which should be linked in the first paragraph of the article.