Sam Altman admits GPT-4o became ‘Too Sycophantic,’ promises quick fixes — Here’s what happened | Mint

OpenAI has moved swiftly to handle an surprising quirk in its not too long ago up to date GPT-4o mannequin after customers started encountering overly deferential replies. The San Francisco-based agency rolled out the improve on Saturday, 26 April 2025, promising enhancements to the mannequin’s intelligence and conversational aptitude – enhancements that additionally underpin its newly built-in image-generation functionality.

By the next day, nevertheless, firm engineers found that the very tweaks designed to make GPT-4o extra partaking had as a substitute rendered it excessively accommodating, even veering into sycophantic territory. In a put up on X on Sunday, CEO Sam Altman acknowledged the problem, describing the mannequin’s new manner as “too sycophant-y and annoying.”

Though Altman emphasised that vital parts of the up to date mannequin had been performing “very nicely,” he admitted that restoring the suitable steadiness would take time. “We’re already on it,” he wrote, including that whereas some fixes might be deployed inside 24 hours, different changes won’t be full till early subsequent week.

Modifications to large-language fashions corresponding to GPT-4o are routine within the tech trade, encompassing not solely expansions of the mannequin’s information base but additionally algorithmic refinements to make sure extra pure and context-aware interactions. Earlier iterations of GPT-4o had already launched conversational enhancements and, extra not too long ago, an image-generation characteristic powered by the identical underlying structure.

Altman’s thread additionally touched on a longer-term limitation of the ChatGPT expertise: as soon as a brand new mannequin model is activated, customers don’t have any strategy to revert to its predecessor. In response to a question about providing model selections, he conceded that “we clearly want to have the ability to supply a number of choices,” signalling that future releases might allow customers to pick between completely different mannequin variants.

In the meantime, OpenAI can be introducing a streamlined version of its agentic deep-research device for free-tier customers, whereas concurrently elevating the each day question limits for Plus, Workforce, Enterprise and Training subscribers. Based on the corporate, this lighter deep-research utility is powered by the newly unveiled o4-mini reasoning mannequin and has been engineered to stay cost-effective with out sacrificing high quality.