Sam Altman addresses ‘bumpy’ GPT-5 rollout, bringing 4o back, and the ‘chart crime’

During a Reddit Ask me On Friday, Sam Altman, CEO of OpenAI, and the main members of the GPT -5 team, requested the new model to be questioned and return its previous model GPT -4O.

They were the most embarrassing of Altman – and perhaps fun – in the presentation, Snapu, also asked about “chart crime”.

GPT -5 is one of the new features that have been roll out A real-time router It decides which model to use for a particular prompt, either quickly react or take extra time to “think” through the answer.

However multiple people in AMA R/ChatzPT Reddit Complaint GPT -5 is not working side by side like that of 4O. Ultman said that the router was not working properly when it was launched on the GPT -5 that was launched on Thursday.

“GPT -5 would feel smarter today. Yesterday, we had a sav and the autosaumer was out of the commission for a part of the day, and the result was just as low as GPT -5. Also, we are interfering on how the decision boundary helps get the right model more frequent.” What we answered, what we gave will be answered.

Nevertheless, AMA’s people were so planning to bring 4O to more plus customers that Ultman had at least promised to search. He writes, “We are trying to continue using Plus users 4o. We are trying to collect more data in trade offs,” he wrote.

And Ultman also promised, “As soon as we finish the rollout, we are going to leave the limit of double rates for plus users.” It should be given the opportunity to play and learn the new model.

Guessedly, he was also asked about the wildly incorrect chart that the team had presented during the live presentation which quickly became butt of many “Chart crimeThe The chart has presented a lower benchmark score with a long bar.

Ultman did not answer the question about the chart during the AMA, but on Thursday he called the chart a “mega chart screwup” At the xThe Others mentioned the charts Published blog post was correctThe

However, the damage was done. Jokes has worked on using GPT for the chart in the corporate presentation. GPT -5 reviewer Simon Willison, who received initial access and usually liked the model performance, Also indicated That data was converted to a table “a good example of a GPT -5 failure.”

Whatever the case, Ultman promised to correct items that seemed to be the most worrying of people. He finished AMA with such a promise that “We will continue to work to stabilize the issues and listen to the response.”