OpenAI explains why ChatGPT became too sycophantic

Openi has published by postmortem on recent questions of sycophacy With a default AI model by powertling chatgpt, Gpt-4o – Questions that forced the company to restore updating the model published last week.

During the weekend, after updating the GPT-4O model, users on social networks noted that Chatgpt began to respond to an over-confirm and acceptable way. He quickly became meme. Users have released Chatgptic screen shots that applaud all kinds of problematic, dangerous decisions and ideas.

In the post on X on Sunday, the executive Director is Altman admit The problem and said Openi would work to repair “as soon as possible.” Two days later, Altman announced The GPT-4O update was returned back and opened Openii on the “additional repairs” of the model of the model.

According to the OpeniThe update, which was intended for the default model “to be” felt more intuitive and effective “, was too informed by” short-term feedback “and” did not fully explain that the interactions of users with Chatgpt develop with time. “

“As a result, GPT -4o distorted the answers that were overly supporting but unwavering,” Openii wrote in a blog post. “Sycophantic interactions can be uncomfortable, disturbing and cause trouble. We have reduced and we are working to correct it.”

Openii says it implementing several corrections, including the training of its training techniques for the basic model and system instructions to explicitly deprive the GPT-4o away from Sikofan. (System inquiries are initial instructions that direct comprehensive behavior and tone of models in interactions.) The company also builds more security protection to “increase [the model’s] Honesty and transparency, “and continuing to expand their assessments to” help recognizing the problem outside Sikofance, “it says.

Openii also says that it explores ways to allow customers to “real time” in real time to “directly affect their interactions” with Chatgpt and select more Chatgpt personality.

“[W]Exploring new ways to include broader, democratic feedback in the default behaviors of Chatgpt, “the company wrote in her blog post.” We also believe that users should have more control over the behavior of the chatgpta behavior and, to the extent that it is safe and feasible, adjust if they do not agree with the default behavior. “