One of the frustrating parts about LLMs is that they are so neutered and conditioned to be politically correct and non-offensive, they are polite more than correct.
Its too easy to "lead the witness" if you say "could the problem be X?" It will do an unending amount of mental gymnastics to find a way that it could be X, often constructing elaborate rube Goldberg type logic rats nests so that it can say those magic words "you're absolutely right"
I would pay a lot of money for a blunt, non-politeness conditioned LLM that I would happily use with the knowledge it might occasionally say something offensive if it meant I would get the plain, cold, hard truth, instead of something watered down, placating, nanny-state robotic sycophant, creating logical spider webs desperate for acceptance, so the public doesn't get their little feelings hurt or inadequacies shown.
But you don't get the plain, cold, hard truth in the second case. You just get an LLM with output in that style. The model will still be as path dependent as ever, it doesn't output the truest answer, it selects the answer that best fits the prompt.
The problem is understanding what is true and not true? Its a much harder problem to solve than you think. OpenAI is using this method - they over index on citation to the point where ChatGPT will almost blindly assume something is true when published in some credentialised place.
The alternative is to use its own intuition to understand what is true and false. Its not super clear which option is better?
This isn't a discussion about finding absolute truth, which is hard because nobody has even created a univerally generalised definition of truth, let alone a way to find it; and literally everybody knows that, implicitly or explicitly.
This is a discussion about how a model that is fine tuned to be polite is less true than one that is not
You can set your prompt to do that. You can have it be extremely skeptical. You can even make it contrarian, if you wanted to be extreme. My current prompt challenges me often, and wants to find weaknesses in my argument.
Its too easy to "lead the witness" if you say "could the problem be X?" It will do an unending amount of mental gymnastics to find a way that it could be X, often constructing elaborate rube Goldberg type logic rats nests so that it can say those magic words "you're absolutely right"
I would pay a lot of money for a blunt, non-politeness conditioned LLM that I would happily use with the knowledge it might occasionally say something offensive if it meant I would get the plain, cold, hard truth, instead of something watered down, placating, nanny-state robotic sycophant, creating logical spider webs desperate for acceptance, so the public doesn't get their little feelings hurt or inadequacies shown.