OpenAI pledges to make changes to prevent future ChatGPT sycophancy

Open Say it will change The way it updates the AI ​​models that Power ChatzPT, after an incident that makes the platform become too psychopantic for many users.

Last weekend, after OpenAI roll out a tweet GPT -4O – Default Model Powering ChatzPT – Social media users have noted that the ChatzP has begun to respond to excessive validity and agreed ways. It quickly became a meme. Users have posted a screenshot of the ChatzPT for all sorts of problematic appreciation, Hazardous Decision And ConceptThe

CEO Sam Altman on a post of X -e -Last Sunday Recognized The problem was and said that Opina would work on “ASAP” fixes. Tuesday, Ultman Declaration The GPT -4O update is being turned up again and is working on the “additional correction” for the OPENY model personality.

The company has released Postmortem On Tuesday, and Friday, the OpenAI has expanded to the specific compatibility that OpenAI plans to set up its model.

OpenAI says it is planned to introduce an opt-in “alpha phase” for some models that allow the specified ChatGPT to allow users to examine models and react before launch. The company also says that it will include explanation of “known constraints” for future extended updates in the chatzPT models, and personality, cheating, reliability and hallucinations (eg, when a model makes things) “Launch-Blocking” is the problem of model behavior “Model behavior” formally considering its protection process.

In the blog post, Open AA wrote, “To move forward, we will actively contact the updates we are creating on ‘fine’ or not, ChatzPT models.” “Even if these problems are not perfectly quite quite quite quite quite, we are committed to blocking launches based on proxy measurements or qualitative signals, even when metrics like A/B exam look good.”

More people come back to the Chatzipt for advice, as promised fixes come. According to the recent one survey By the legal fund of the case Finannser Express, 60% of US adults have used ChatzPT to seek advice or information. Growing dependence on ChatzPT – and huge user base on the platform – when problems such as extreme psychophyse are aroused, does not mention hallucinations and other technical defects.

Earlier this week, the OpenAI said that it would test the users’ ChatzPT to “directly affect their interaction” by using the way to give users a “real-time response”. The agency also says that it will refine the strategies to drive models away from psychophyse, potentially allowing people to choose from multiple models in the chatzipi, create additional protection maintenance and expand evaluations to help identify the issues outside psychophyse.

“One of the biggest lessons is how people have begun to use the ChatzP for deeply personal advice – something we didn’t see a year ago,” its blog post continues to open. “At that time, it was not a preliminary focus, but as the AI ​​and society, it has become clear that we need to be treated very carefully about this use, it is now going to be a more meaningful part of our protection work.”