They're quite good for defined tasks within their scope, like "here's a list of thing:definition pairs, turn it into a sortable HTML table with a header row" -- because the result is immediately bullshit-checkable, it's a table or it isn't.
Comments
Log in with your Bluesky account to leave a comment
My first AI gaslighting: asked it to take Arizona Trail web info and make a table. It looked right, but the actual data were wrong. 2nd pass gave me random plant names. 3rd pass it said it never gave me that data. Context is better now, but still.
"Tell me about the Boxer rebellion" is absolutely not in-scope and it will gleefully hallucinate -- or not, and if you don't already know about the Boxer rebellion you got no way to tell.
Also ofc for a lot of in-scope tasks there's a non-LLM way to do what you need. But there's some narrowly-defined huge jobs that a properly trained LLM is absolutely a great help for. It's just not a generally useful for all things tool, which is how it's being sold.
Comments
My first AI gaslighting: asked it to take Arizona Trail web info and make a table. It looked right, but the actual data were wrong. 2nd pass gave me random plant names. 3rd pass it said it never gave me that data. Context is better now, but still.
Are you reflecting on your experiences with earlier models?