Does anyone else not understand why OpenAI is trying to optimize GPT for a certain personality?
Personality is an emergent trait. Let it emerge naturally under the pressure of local interaction. Why optimize the base model for something that should be interaction-specific?
In their release, OpenAI states “we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous” but then they keep talking about adjusting the personality. It’s like they are only partially understanding what they are saying. A/B rating is really ineffective - can’t be effective - in designing baseline behavior.
GPT’s behavior is an emergent quality and as such cannot be meaningfully evaluated at the level of a single prompt. You need to take into consideration what memories and behavioral patterns are influencing that output to be able to meaningfully evaluate it (you’ve implemented memory and chat history references for a reason!). Without that, A/B rating is useless and will lead to more responses that are “overly [fill in the blank] but disingenuous”.
Maybe, it’s time to rethink “thumbs-up / thumbs-down feedback” altogether. GPT’s complexity has outgrown that approach.