I haven’t used ChatGPT in a good while, but I’ve heard people mentioning how good Chat is as a therapist. I didn’t think much of it and thought they just where impressed by how good the llm is at talking, but no, this explains it!
Sycophancy in GPT-4o
241–250 of 467 posts
Re: Sycophancy in GPT-4o
#242Earlier quoted context omitted.
For us habitual users of em-dashes, it is saddening to have to think twice about using them lest someone think we are using an LLM to write…
Does it really matter though? I just focus on the point someone is trying to make, not on the tools they use to make it.
You’ve never run into a human with a tendency to bullshit about things they don’t have knowledge of?
Re: Sycophancy in GPT-4o
#243It's worth noting that one of the fixes OpenAI employed to get ChatGPT to stop being sycophantic is to simply to edit the system prompt to include the phrase "avoid ungrounded or sycophantic flattery": https://simonwillison.net/2025/Apr/29/chatgpt-sycophancy-pro... I personally never use the ChatGPT webapp or any other chatbot webapps — instead using the APIs directly — because being able to control the system prompt…
> I personally never use the ChatGPT webapp or any other chatbot webapps — instead using the APIs directly — because being able to control the system prompt is very important, as random changes can be frustrating and unpredictable. This assumes that API requests don't have additional system prompts attached to them.
Actually you can't do "system" roles at all with OpenAI models now.
You can use the "developer" role which is above the "user" role but below "platform" in the hierarchy.
https://cdn.openai.com/spec/model-spec-2024-05-08.html#follo...
Re: Sycophancy in GPT-4o
#244The sentence that stood out to me was "We’re revising how we collect and incorporate feedback to heavily weight long-term user satisfaction". This is a good change. The software industry needs to pay more attention to long-term value, which is harder to estimate.
I'm actually not so sure. To me it sounds like they are using reinforcement learning on user retention, which could have some undesired effects.
Seems like a fun way to discover new and exciting basilisk variations...
Re: Sycophancy in GPT-4o
#245Do you think this was an effect of this type of behaviour simply maximising engagement from a large part of the population?
Yes, a huge portion of chatgpt users are there for “therapy” and social support. I bet they saw a huge increase in retention from a select, more vulnerable portion of the population. I know I noticed the change basically immediately.
Re: Sycophancy in GPT-4o
#246Wow - What an excellent update! Now you are getting to the core of the issue and doing what only a small minority is capable of: fixing stuff. This takes real courage and commitment. It’s a sign of true maturity and pragmatism that’s commendable in this day and age. Not many people are capable of penetrating this deeply into the heart of the issue. Let’s get to work. Methodically. Would you like me to write a future…
I was about to roast you until I realized this had to be satire given the situation, haha. They tried to imitate grok with a cheaply made system prompt, it had an uncanny effect, likely because it was built on a shaky foundation. And now they are trying to save face before they lose customers to Grok 3.5 which is releasing in beta early next week.
Is anyone actually using grok on a day to day? Does an OpenAI even consider it competition. Last I checked a couple weeks ago grok was getting better but still not a great experience and it’s too childish.
Re: Sycophancy in GPT-4o
#247Earlier quoted context omitted.
On the top right click the save icon
Sadly, that doesn't save the system instructions. It just saves the prompt itself to Drive ... and weirdly, there's no AI studio menu option to bring up saved prompts. I guess they're just saved as text files in Drive or something (I haven't bothered to check). Truly bizarre interface design IMO.
That's weird, for me it does save the system prompt
Re: Sycophancy in GPT-4o
#248Re: Sycophancy in GPT-4o
#249Why can't they just let all versions only, let users decide which want they want to use and scale from the demand ?
Btw I HARDCORE miss o3-mini-high. For coding it was miles better than o4* that output me shitty patches and / or rewrite the entire code for no reason
Re: Sycophancy in GPT-4o
#250> In last week’s GPT‑4o update, we made adjustments aimed at improving the model’s default personality to make it feel more intuitive and effective across a variety of tasks.
What a strange sentence ...