Openai says that the GPT-4O update can be “uncomfortable, worrying, and causes hardship”

Openai has declined to the GPT-4O update for ChatgPT, which caused the virtual Chatbot to be “very mired or acceptable-often described as Sycophanty” and that “cycuttle reactions can be uncomfortable, worrying, causing distress,” He says in a blog post.

The company made a GPT-4O update last week that included modifications “aimed at improving the virtual personality of the model to make it easier and effective through a variety of tasks,” according to humans. Openai says it begins to form typical behavior first with what has been determined in Model specifications The forms teach how to apply the principles contained in the specifications “by combining user signals such as thumb / thumb on the reactions to ChatGPT responses.”

But with a back update, Openai says: “We have focused a lot on short -term comments, and we are not completely account on how user interactions with Chatgpt developed over time.” This means that “GPT -4O is deviant towards responses that were excessively supportive but deceitful.”

Openai designs the virtual Chatgpt character “to reflect our mission and be useful, supportive and respectful for different values ​​and experience,” says the blog post, but he adds that “each of these desired qualities such as an attempt to be useful or supportive can have unintended side effects.” The company says that “one virtual cannot capture every preference” for the 500 million cahatgpt users.

Openai will “take more steps to reorganize the behavior of the model”, including “improving basic training technologies and system demands to directly direct the model away from Sycophance” and “expanding methods” for users to make notes. “We also believe that users should have more control of how ChatGPT behaves, and the extent that it is safe and possible, to make adjustments if they do not agree with the virtual behavior,” the company says.