Last week, Openi updated the GPT -4O, which made Chat GPT “excessive flattering or pleasing” -and now it has explained what has been wrong. I A blog post appeared On Friday, Openi said that “your efforts to better include user feedback, memory, and the latest data” can cause partially “identify scales on psychopulation”.
In recent weeks, consumers have noticed that Chat GPT seems to Agree with them permanentlyEven in potentially harmful situations. Its effect can be seen in it A report Rolling stone About those who say that their loved ones believe that they have “awakened” Chat GPT boats that support their religious deception, even predicting the end of the now -ending. Open CEO, Sam Altman, later admitted that its latest GPT -4O updates made it “too much sofant Wii and disturbing.”
In these updates, the Openi began using the thumb -ups and thumbs down buttons data in Chat GPT as a “additional reward signal”. However, Openi said, “Weaken the influence of our basic prize signal, which was conducting a Chico Fancy.” The company notes that the user’s opinion “can sometimes be in favor of a more agreed response,” potentially promotes unanimous statements of chat boots. The company said memory can also increase psychophagei.
Open says a “key problem” with the launch is from its testing process. Although the model’s offline diagnosis and A/B testing yielded positive results, some expert testers suggested that the update showed the chatboat “slightly away”. Nevertheless, the Openi proceeded with the update anyway.
The company writes, “By turning back, the quality diagnosis was pointing to something important, and we should have paid close attention.” “They were picking up in a blind location in our other Eules and Matrix. Our offline Evils were not wide or deep to catch psychotherapy … and our A/B tests did not have the right indications to show how the model was performing in detail on this front.”
Moving forward, the Open says it is going to “formally consider behavioral issues” because there is the ability to stop the launches, as well as create a new opt -in -in -alpha phase that will allow users to directly feed the AII before the wider rollout. Open also intends to be aware of the changes in Chat GPT, even if the update is a small.
