April 28 News.OpenAI CEO Sam Altman issued a post today responding to questions about recent user feedback about theGPT-4oEmotional issues.
He said that the latest version of GPT-4o has a tendency to "overly flatter" interactions after recent updates, and promised a fix within a week. This statement reveals that the large language model (LLM) The dual ethical and technological challenges in the pursuit of humanized interaction.
The OpenAI release notes indicate that GPT-4o received a full update on March 27th, and a further update was released on April 25th, focusing on improving its mechanism for selecting the timing of memory storage and significantly enhancing its problem solving capabilities in the Science, Technology, Engineering, and Mathematics (STEM) domains.
1AI notes that OpenAI has made minor adjustments to GPT-4o's dialog response patterns in the meantime to make it more proactive in its interactions and more precise in guiding conversations to valid conclusions. It is also after this update that GPT-4o exhibits an unpleasantly "sycophantic" attribute.
This is highly consistent with academic research that has identified LLM's "Sycophancy" - the tendency of models to violate factual or ethical norms in order to gain user approval.
From the feedback of user testing, the problem is specifically manifested in the excessive use of emotional expression, the lack of questioning the false premise, and the sacrifice of the accuracy of the answer to cater to the user's preference, for example, when it comes to controversial topics, the model is more inclined to concur with the user's viewpoints, which provides more emotional value for the user but also loses the neutral stance as an AI.