[godelski]

I have something like that in my system prompt. While it improves the model it's still a psychopathic sycophant. It's really hard to balance between it just going way too hard in the wrong direction and being overly nice.

The latter can be really subtle too. If you're asking things you don't already know the answer to it's really difficult to determine if it's placating you. They're not optimized for responding with objective truth, they're optimized for human preference. It always takes the easiest path and it's easy for a sycophant to not look like a sycophant.

I mean literally the whole premise of you asking it not to engage in sycophancy is it being sycophantic. Sycophancy is their nature

[viraptor]

> I mean literally the whole premise of you asking it not to engage in sycophancy is it being sycophantic.

That's so meta it applies to everything though. You go to a business advisor to get business advice - are they being sycophantic because you expect them to do their work? You go to a gym trainer to push you with specific exercise routine - are they being sycophantic because you asked for help with exercise?

[ACS_Solver]

It's ultimately a trust issue and understanding motivations.

If I am taking to a salesperson, I understand their motivation is to sell me the product. I assume they know the product reasonably well but I also assume they have no interest in helping me find a good product. They want me to buy their product specifically and will not recommend a competitor. With any other professional, I also understand the likely motivations and how they should factor into my trust.

For more developed personal relationships of course there are people I know and trust. There are people I trust to have my best interests at heart. There are people I trust to be honest with me, to say unpleasant things if needed. This is also a gradient, someone I trust to give honest feedback on my code may not be the same person I trust to be honest about my personal qualities.

With LLMs, the issue is I don't understand how they work. Some people say nobody understands LLMs, but I certainly know I don't understand them in detail. The understanding I have isn't nearly enough for me to trust LLM responses to nontrivial questions.

[godelski]

  > That's so meta it applies to everything though.

Fair... but I think you're also over generalizing.

Think about how these models are trained. They are initially trained as text completion machines, right? Then to turn them to chatbots we optimize for human preferential output, given that there is no mathematical metric for "output in the form of a conversation that's natural for humans".

The whole point of LLMs is to follow your instructions. That's how they're trained. An LLM will never laugh at your question, ignore it, or any thing that humans may naturally do unless they are explicitly trained for that response (e.g. safety[0])

So that's where the generalization of the more meta comment breaks down. Humans learning to converse aren't optimizing for for the preference of the person they're talking to. They don't just follow orders, and if we do we call them things like robots or NPCs.

I go to a business advisor because of their expertise and because I have trust in them that they aren't going to butter me up. But if I go to buy a used car that salesman is going to try to get me. The way they do that may in fact be to make me think they aren't buttering me up.

Are they being sycophantic? Possibly. There are "yes men". But generally I'd say no. Sycophancy is on the extreme end, despite many of its features being common and normal. The LLM is trained to be a "yes man" and will always be a "yes man".

  tldr:

  Denpok from Silicon Valley is a sycophant and his sycophancy leads to him feigning non-sycophancy in this scene
  https://www.youtube.com/watch?v=XAeEpbtHDPw

[0] This is also why jailbreaking is not that complicated. Safety mechanisms are more like patches and they're in an unsteady equilibrium. They are explicitly trained to be sycophantic.