OpenAI up to date the GPT-4o synthetic intelligence (AI) mannequin on Saturday. The replace was aimed toward enhancing the intelligence and character facets of the massive language mannequin (LLM), which powers the default expertise for these on the free tier of ChatGPT. Nevertheless, on Sunday, the corporate discovered that the replace had sudden outcomes, and the AI mannequin’s responses grew to become too agreeable. The San Francisco-based AI agency is now engaged on fixing the difficulty, and stated it would take until subsequent week earlier than the issue might be fastened.
OpenAI Replace Introduces Unpredictable Glitch in GPT-4o
In a submit on X (previously often known as Twitter), OpenAI CEO Sam Altman acknowledged that the GPT-4o AI mannequin was up to date to enhance the intelligence and character of the AI mannequin. Notably, all ChatGPT customers have entry to this explicit mannequin, and the corporate’s new built-in picture technology characteristic can be powered by the identical AI mannequin.
Updating an AI mannequin is a standard apply for tech corporations. Normally, this contains updating the data base and tweaking algorithms to enhance sure facets of the mannequin. Even earlier than including picture technology by way of an analogous replace, OpenAI had additionally up to date GPT-4o to make its responses extra pure and conversational. Notably, Google additionally up to date the Gemini 2.0 Flash to enhance its conversational functionality.
Nevertheless, simply days after rolling out the replace, Altman took to X to inform customers that the newer model of GPT-4o unexpectedly grew to become too agreeable. The CEO stated, “The final couple of GPT-4o updates have made the character too sycophant-y and annoying.”
He added that the corporate started engaged on fixing the difficulty, and whereas a few of it will be fastened in a day, others would possibly take a complete week. Altman additionally added that some components of the brand new GPT-4o mannequin are “superb.”
In the identical thread, a commenter requested Altman if the corporate can add a means for ChatGPT customers to tell apart between the brand new and the outdated variations. At present, as soon as a mannequin is up to date, customers can’t revert to an older model. Acknowledging the limitation, the CEO acknowledged, “Ultimately we clearly want to have the ability to provide a number of choices.”