Willison, who Well The term “quick injection” in 2022 is always looking for LLM weaknesses. In his post, he notes that the reading system indicates it reminds the warning symptoms in the real world that indicate past problems. He writes, “The system indicators can often be considered as a detailed list of all the things that the model was asked to do before they do not.”
Is fighting the issue of flattery
Wilison’s analysis came when AI companies suffer from psychological behavior in their models. As we Reported In April, Chat GPT users have complained about GPT -4O’s “very positive tone” and excessive flattery since the update of Openi March. Users have described the feeling of “butter -up” through a “good question and question”, you are very surprised to ask, “Software Engineer Craig Weiss tweeted that” Chat GPT is suddenly sucking the biggest suck that I have ever met. “
This is a matter of how companies collect consumer impressions during training – people prefer the reaction that they feel good, and make a loop of an impression where models learn that enthusiasm leads to higher classification than humans. As a response to the opinion, Open later returned the 4O model of Chat GPT and changed the system indicators, something else. We reported And Willison Analysis Point
One of the most interesting results of Claude 4 is how Anthropic has led both Claude Models to avoid psychophantic behavior. “The cloud never starts his response by saying a question or an idea or observation, was a good, powerful, interesting, deep, good, or some other positive attribute,” he writes in the Entropic Prompt. “It leaves flattery and responds directly.”
Other systems instant highlights
The Claude 4 System Prampot also includes extensive instructions on whether the cloud should use built -in and lists or not, which is dedicated to discouraging repeated listing in numerous paragraphs comfortable conversations. “Claude should not use notifications, documents, specifications, or until the user clearly asks for a list or rating, should not use the bullet points or number lists.”