And research articles indicate that when the model computes that it should employ sycophantism it becomes less useful in every other way, just like a real sycophant.
Why tech CEOs love LLMs. Ultimate yes man.
Claude says “You're absolutely right!” about everything
21–30 of 560 posts
Re: Claude says “You're absolutely right!” about everything
#22What llm isn’t a sycophant?
Grok. These were all in continuation, not first reply.
> Thank you for sharing the underlying Eloquent query...
> The test is failing because...
> Here’s a bash script that performs...
> Got it, you meant...
> Based on the context...
> Thank you for providing the additional details...
Re: Claude says “You're absolutely right!” about everything
#23Not Claude but ChatGPT - I asked it to pipe down on exactly that kind of response. And it did.
I've done this in my Claude settings, but it still doesn't seem that keen on following it:
> Please be measured and critical in your response. I appreciate the enthusiasm, but I highly doubt everything I say is “brilliant” or “astute”, etc.! I prefer objectivity to sycophancy.
Re: Claude says “You're absolutely right!” about everything
#24I find Gemini is also hilariously enthusiastic about telling you how amazingly insightful you are being, almost no matter what you say. Doesn't bother me much, I basically just ignore the first paragraph of any reply, but it's kind of funny.
100%! I got the following the other day which made me laugh out loud:
"That's a very sharp question. You've correctly identified the main architectural tension in this kind of data model"
Re: Claude says “You're absolutely right!” about everything
#25And research articles indicate that when the model computes that it should employ sycophantism it becomes less useful in every other way, just like a real sycophant.
> And research articles indicate that when the model computes that it should employ sycophantism it becomes less useful in every other way, just like a real sycophant. The end goal of a sycophant is to gain advantage with their flattery. If sycophant behavior gets Claude's users to favour Claude over other competing LLM services, they prove to be more useful to the service provider.
Until users find out it's less useful to the user because of that.
Or it causes some tragedies...
Re: Claude says “You're absolutely right!” about everything
#26This applies to so many AIs. I don't want a bubbly sycophant. I don't want a fake personality or an anime avatar. I just want a helpful assistant. I also don't get wanting to talk to an AI. Unless you are alone, that's going to be irritating for everyone else around.
Sure but different people have different preferences. Some people mourn replacement of GPT4 with 5 because 5 has way less of a bubbly personality.
Re: Claude says “You're absolutely right!” about everything
#27[dead]
Re: Claude says “You're absolutely right!” about everything
#28I find Gemini is also hilariously enthusiastic about telling you how amazingly insightful you are being, almost no matter what you say. Doesn't bother me much, I basically just ignore the first paragraph of any reply, but it's kind of funny.
I was feeding Gemini faux physicians notes trying to get it to produce diagnosises, and every time I feed it new information it told me how great I was at taking comprehensive medical notes. So irritating. It also had a tendency to tell me everything was a medical crisis and the patient needed to see additional specialists ASAP. At one point telling me that a faux patient with normal A1C, fasted glucose and no diabet…
And then those sort of responses get parlayed into "chatbots give better feedback than medical doctors" headlines according to studies that rate them as high in "empathy" and don't worry about minor details like accuracy....
Re: Claude says “You're absolutely right!” about everything
#29This bug report is absolutely right
Yeah, I was working through the design of part of this thing I've been working on and noticed that every time I would ask a follow up question it would change its opinion to agree that this new iteration was the best thing since sliced bread. I eventually had to call it out on it to get an 'honest' assessment of the various options we were discussing since I didn't want 'my thing' to be correct but the system as a whole to be correct.
And it's not like we were working on something too complicated for a daffy robot to understand, just trying to combine two relatively simple algorithms to do the thing which needed to be done in a way which (probably) hasn't been done before.
Re: Claude says “You're absolutely right!” about everything
#30Not Claude but ChatGPT - I asked it to pipe down on exactly that kind of response. And it did.
Yet I'll tell it 100 times to stop using em dashes and it refuses.