The language models of OpenAIin addition to the big updates, they have the little ones, which They do not represent a change of nomenclature, but they can become quite significant. It is what happened in March, when GPT-4Othe predetermined model of Chatgptwon a native image generator that has been very well received. Last week, however, The opposite happened. OpenAI updated the same model, which now He has reverted to the previous version because ‘it was too flattering or pleasant, often described as flatterer’. With this outdated, Chatgpt has ‘a more balanced behavior’ again.
The problem is flattery
According to the company in its blog, the last update consisted of a series of settings in the predetermined personality of the GPT-4O model ‘To feel more intuitive and effective in a variety of tasks’.
However, the result was not expected and observed how ‘GPT-4O leaned towards answers that were Too much support, but false‘. Too much ‘flattery’ of chatgpt to the user.
How the Personality of Chatgpt is formed
Openai explains that, to form the behavior of the model, they begin by the principles and instructions described in the Model specifications. To that are added User signalsas the thumb options up or down that chatgpt gives to assess their answers.
Openai focused on these short -term valuations, instead of the long term and without taking into account how users’ interactions evolve with AI over time. Something Openai is correcting while introducing more customization functions.
A personality that can be “uncomfortable, disturbing and causing anguish”
The company indicates that the predetermined personality of the chatbot is a fundamental element of the experience of use. ‘Flattening interactions can be uncomfortable, disturbing and cause anguish. We fell short and we are working to do it well, ‘he says.
The difficulty facing chatgpt is a huge user base – with more than 500 million per week– Belonging to all cultures and contexts. It is very difficult to satisfy everyone. ‘Each of these desirable qualities, such as trying to be useful or support, can have unwanted side effects‘, they explain.
In addition to the internal measures they are taking to realign the behavior of the model, OpenAI points out that users They already have the option of instructions to the model so that it has a personalized behavior. Soon, they can comment on real time to directly influence their interactions and Choose between multiple predetermined personalities‘.