- Openai addressed what did not work with its sycophandry update
- OPENAI refines basic training techniques to avoid sycophance in the future
- Its model of the model’s behavior suggests that the Chatppt could have several personalities in the future
Openai addressed the decline in the most recent update of Chatgpt-4O due to the complaints of users concerning his boring sycophantic personality. In a blog article of April 29 Openai States:
“We fell the GPT-4O update last week in the Chatppt, so people now use a previous version with more balanced behavior. The update we have deleted was too flattering or pleasant – often described as sycophantic.
We actively test new fixes to solve the problem. We revise how we collect and incorporate comments into long -term satisfaction of long -term users and we introduce more personalization features, giving users greater control over chatgpt behavior. »»
The blog post explains other reasons why the problem has happened, but the node of the problem is that Optai focused too much on short -term comments, and has not fully explained how user interactions with Chatgpt evolve over time.
OPENAI promises to refine its basic training techniques and systems prompts to explicitly keep the sycophance model in the future, as well as allow more user tests before deploying models.
Ask me
By reading the blog post, you have the impression that Optai has the impression of having dropped the ball on it and does everything it can to make amends.
In addition to publishing the Declaration, Joanne Jang, her modeling manager of the model, has just made a “ Ask me anything ” (AMA) on Reddit, to answer all the other questions that people can have, in particular on the personality, sycophance and future behavior of Chatgpt. His answers revealed interesting information on what the future could contain for the personality of Chatgpt:
Reddit User Responsible_COW2236 asked: “Where do you see the future of the behavioral of the model? Let’s move on to more customizable personalities, like giving users tools to shape how Chatgpt sounds and interacts with them over time?”
Jang responded by offering the possibility that Chatgpt could come with several predefined personalities:
“We have worked on two things: (1) to happen to a default personality that could be acceptable so that all users can start (not possible, but we have to arrive somewhere) and (2) instead of counting on users to describe / find personalities by themselves, offering presets which are easier to understand (for example, personality descriptions against 30 sliders on lines)”
The idea of personalities predefined in Chatgpt is an intriguing and something that I could see well combine with the different voices that the advanced vocal mode already offers.
As for what we can expect at the next update for Chatgpt-4o, Openai says nothing right now beyond “we are failed and that we are working on things.”