Sam Altman acknowledges that GPT-4O has become too cikophantic, promised quick correction here is what happened | Mint

Openai moved quickly to address an unexpected peculiarity in its recently updated GPT-4O model after users began experiencing too much postponement answers. The firm in San Francisco erected the upgrade on Saturday, April 26, 2025, which promises the improvement of the model’s intelligence and conversational flair improvements that also support its newly integrated image-generation capacity. By the next day, however, the company engineers discovered that the adjustments designed to make GPT-4O more engaging made it excessive, even in the cycophantic area. Sam Altman acknowledged in a post on X on X on X post on Sunday and described the new behavior of the model as “too Sycophant-y and annoying”. Although Altman emphasized that significant parts of the updated model perform ‘very well’, he admitted that the restoration of the right balance would take time. “We are already on it,” he wrote, adding that although some corrections can be deployed within 24 hours, other adjustments may only have been completed early next week. Changes to large language models such as GPT-4O are routine in the technology industry, which contains not only extensions of the model of knowledge, but also algorithmic refinements to ensure more natural and context-conscious interactions. Earlier, GPT-4O iPT-4O has already introduced conversational enhancements, and more recently a Beeld generation feature powered by the same underlying architecture. Altman’s thread also touched on a longer-term restriction of the Chatgpt experience: Once a new model version is activated, users have no way to return to its predecessor. In response to an inquiry about offering version choices, he admitted that “we are clearly able to offer various options,” indicating that future releases can enable users to choose between different model variants. Meanwhile, Openai also introduces a streamlined edition of its Agentic Deep Research to users with free levels, while at the same time increasing the daily keyword restrictions for subscribers for plus, team, enterprise and education. According to the company, this lighter deep research program is driven by the newly unveiled O4 mini reasoning model and is designed to remain cost-effective without sacrificing quality. First published: 28 Apr 2025, 23:42 IST