Last week, Opeli conducted a GPT-4O update, which made chatgpt “excessively flattering or pleasant”-and now explained what exactly went wrong. IN Published post on the blog On Friday, Opeli said that his efforts to “better include user feedback, memory and fresh data” could partly lead to “tilting scales on a hat”.
In recent weeks, users have noticed that chatgpt seemed I still agree with themEven in potentially harmful situations. The effect of this can be seen in Report by Rolling Stone About people who say that their loved ones believe that they “woke” the botgpt bots that support their religious illusions of greatness, even ahead of the update. Later, the general director of Opeli, Altman himself, later admitted that his latest GPT-4O updates made “too flattering and annoying”.
In these updates, Opeli began to exploit data from the thumb buttons and thumbs in ChatgPT as a “additional reward signal”. However, Openai said, this could “weaken the impact of our basic reward signal, which at the same time maintained favorable.” The company notes that the opinions of users “sometimes can favor more pleasant answers”, probably exacerbates too pleasant Chatbot’s statements. The company said that memory can also strengthen flattery.
Opeli says that one of the “key problems” with launch results from the testing process. Although the offline of the model and A/B tests had positive results, some experts suggested that the update made chatbot “slightly disconnected”. Despite this, Opennai still moved forward with the update.
“Looking back, qualitative assessments indicated something important and we should be closer,” writes the company. “They took a deadline in our other evolutions and indicators. Our offline evolutions were not wide or deep enough to catch sycofantic behavior … And our A/B tests did not have adequate signals to show how the model worked on this front with enough details.”
Going further, OPENAI claims that “will formally consider behavioral problems” as the potential of blocking the prime minister, as well as create a up-to-date OPT-in Alfa phase, which will allow users to provide direct feedback of OPENAI before wider implementation. Opeli also plans to ensure that users are aware of the changes that it introduces to ChatgPT, even if the update is tiny.