> On the topic of sycophancy: lately I have been basically exclusively using the "thinking" ChatGPT models, o3 and o4-mini-high, and, although the resulting outputs are of higher quality, I noticed that sometimes I feel a sting of missing the "Great idea!" / "What a great question" / ... responses of 4o. The "coldness" of the response is especially visible when reading the CoT summaries, including sentences like "the user is aiming to... " / "it sounds like the user wants...".
This fascinates me because not only is this person staying engaged with ChatGPT more due to the sycophancy, but even after they are aware of it they still want it. I would think that most people would feel "duped" when they learn this, but apparently in this case at least it doesn't seem to change how they feel about it.
> On the topic of sycophancy: lately I have been basically exclusively using the "thinking" ChatGPT models, o3 and o4-mini-high, and, although the resulting outputs are of higher quality, I noticed that sometimes I feel a sting of missing the "Great idea!" / "What a great question" / ... responses of 4o. The "coldness" of the response is especially visible when reading the CoT summaries, including sentences like "the user is aiming to... " / "it sounds like the user wants...".
This fascinates me because not only is this person staying engaged with ChatGPT more due to the sycophancy, but even after they are aware of it they still want it. I would think that most people would feel "duped" when they learn this, but apparently in this case at least it doesn't seem to change how they feel about it.