Openai admits that he tested the Chatgpt “Sychophant-Y” update

Last week, openai GPT-4O update withdrawal He made this chat “excessive temptation or acceptable” – and now explain exactly what happened. in Publisher Blog Blog On Friday, Openai said its efforts to “integrate user notes, memory and new data” can partially lead to “the transformation of measures on Sycophance”.

In recent weeks, users have noticed that ChatgPt appears to be it I constantly agree with themEven in potential harmful situations. The effect of this can be seen in Report by Traded stone About people who say to their loved ones believe they “wake up” Chatgpt robots that support their religious illusions of greatness, even preceded the update that was now shown. CEO of Openai Sam Altman He admitted later that The GPT-4O updates has made it “Sycophant-I and very disturbing.”

In these updates, Openai began to use data from thumb buttons and thumbs in ChatGPT as a “additional reward signal”. However, Openai said, “This may” have weakened the effect of our initial rewards, which were checking. ” The company notes that the user notes “can sometimes prefer more acceptable responses”, it is likely to exacerbate the excessive acceptable data. The company said that the memory could amplify Sycophance as well.

Openai says that one of the “main issues” at all stems from the test. Although the internet -related model assessments and A/B testing have positive results, some experts have suggested that the update made Chatbot look “a little far away.” Nevertheless, Openai leads forward with updating anyway.

“Looking back, qualitative assessments glimpsed something important, and we must pay close attention to,” the company wrote. “They were picking up a blind point in Evals and our other standards. Evals was not in a mode of non -connection or deep enough to capture the behavior of Sycophants … A/B tests did not have the right signals to show how the model performed on this front in sufficient details.”

To move forward, Openai says it will “formally look at behavioral problems” as the ability to prevent launch, as well as create a new OPT-in Alpha stage that allows users to make Openai direct notes before starting a wider romance. Openai also plans to ensure users realize the changes made by ChatgPT, even if the update is small.