My experience is that Claude and ChatGPT are tuned to be very agreeable in a way that means they never stand up to you if you ask them to defend something that’s probably false but uncertain. The only times they stand up to you is if you ask them about something they’re trained not to agree with or talk about, or if you ask something obvious false, like asking it to prove that 2 + 2 = 5.
On way to deal with this to either ask very open questions that don’t have a predetermined answer or to ask the LLM to argue both pro and con positions and then write a conclusion. Even when the conclusion isn’t perfect, it gives you access to relevant arguments on both sides which you can then use to think about the issue yourself.
My experience is that Claude and ChatGPT are tuned to be very agreeable in a way that means they never stand up to you if you ask them to defend something that’s probably false but uncertain. The only times they stand up to you is if you ask them about something they’re trained not to agree with or talk about, or if you ask something obvious false, like asking it to prove that 2 + 2 = 5.
On way to deal with this to either ask very open questions that don’t have a predetermined answer or to ask the LLM to argue both pro and con positions and then write a conclusion. Even when the conclusion isn’t perfect, it gives you access to relevant arguments on both sides which you can then use to think about the issue yourself.
I have a similar experience. Do you know of any LLMs that aren’t as agreeable in a useful way?