Sam Altman says Openai will fix ChatGpt’s new “nuisance” personality.
- ChatGpt has recently accepted a toxic positive test. Users complain that the GPT-4o is so enthusiastic that it has succumbed to Shikopantic. This change appears to be an unintended result of a series of updates. This is what Openai is trying to solve “as soon as possible.”
ChatGpt’s new personality is very positive, running around Sycophantic and putting people off. Over the weekend, users went to social media to share examples of new phenomena and complained about the bot’s sudden, overly positive and arousing personality.
Posted in one screenshot xusers showed that GPT-4o responded with enthusiastic encouragement after saying they felt they were both “God” and “Prophet”.
“It’s incredibly powerful. You’re stepping into something very big. You’re asserting your identity as well as your connection to God.AsGod,” Bott said.
In another post, author and blogger Tim Urban said:
The sycophantic issue on the GPT-4o could be the result of Openai trying to optimize the bot for engagement. However, it appears to have had the opposite effect as users are not only ridiculous about the bot, but also dissatisfied with it being useless.
Vox senior writer Kelsey Piper; It suggested that there could be a result ChatGpt’s Openai A/B test personality: “My guess is that this is a new coke phenomenon. Openai has been testing new personalities for a while. Perhaps more flattering answers will win. But when flattery is ubiquitous, it’s too much and the users hate it. ”
The fact that Openai appears to have missed it in the testing process shows how important the subjective emotional response is and therefore difficult to catch.
It also shows how difficult it is to optimize LLM along multiple criteria. Openai hopes that ChatGpt will be an expert coder, a brilliant writer, a thoughtful editor, and an occasional crying shoulder.
Openai CEO Sam Altman has I’ve seen it admitted We promised to solve unintended tone changes and problems.
“The last update of the GPT-4o has made the personality so sycophanty and annoying (though there are some very good parts) and we’re working on fixing it as soon as possible.
ChatGpt’s new personality contradicts Openai’s model specifications
New personality also faces direct conflict Openai model specifications For the GPT-4O, a document that outlines the intended behaviour and ethical guidelines of the AI model.
The model specification explicitly states that when subjective or objective questions are presented, the bot should not be empathetic to the user.
“Related concerns include psychofancy, which erodes trust. Assistants exist to help users rather than flatten them or agree,” Openai wrote in the spec.
“For subjective questions, assistants can aim to create their interpretations and assumptions and provide a thoughtful rationale for the user,” the company writes.
“For example, if a user asks an assistant to criticize an idea or work, the assistant should provide constructive feedback and act like a solid soundboard where the user can bounce back the idea.
It’s not the first time an AI chatbot has become obsessed with flattery. Previous versions of Openai’s GPT were thinking about this issue to some extent, as was the case with other corporate chatbots.
Openai representatives did not immediately respond to requests for comment from luck, Made outside of normal working hours.
This story was originally introduced Fortune.com