I imagine this would work specifically to reduce sycophancy but what about 2nd order effects?
I've found the best ways to get good results from LLMs is firm and clear instructions: I need X so do Y then check Z, W and draw up a plan...
If I put any kind of uncertainty: "I'm thinking about
..." or "would this work..." or "my intern suggested..." then the quality goes down. Or even if we end up with a good result, it will take longer to get there.
It might be useful for learning though, as you and the LLM work through the fundamentals of the "intern's" suggestion to see if it's valid.
I imagine this would work specifically to reduce sycophancy but what about 2nd order effects?
I've found the best ways to get good results from LLMs is firm and clear instructions: I need X so do Y then check Z, W and draw up a plan...
If I put any kind of uncertainty: "I'm thinking about ..." or "would this work..." or "my intern suggested..." then the quality goes down. Or even if we end up with a good result, it will take longer to get there.
It might be useful for learning though, as you and the LLM work through the fundamentals of the "intern's" suggestion to see if it's valid.