during this Big GPT -5 Livestream on ThursdayOpenai showed some charts, making the model look quite impressive – but if you look closely, some of the gradation were slightly away.
In one, the irony is how well the GPT -5 performs in the “Empers during the model”, this scale is everywhere. For example, for “coding deception”, the chart shown on the stage states that GPT-5 says that with thinking, 50.0 percent of the fraud rate is clearly attained, but it is compared to the 47.4 percent O3 score of Openai, which is a big bar. Openai has an exact number for this chart GPT-5 blog postHowever, where the GPT-5 deception rate is 16.5 percent label.
with This chartOpenai showed on stage that GPT-5 is one of the scores lower Compared to O3 but is shown with a larger bar. In this chart, the scores of O3 and GPT-4O are different, but are shown equally with bars of size. It was bad that CEO Sam Altman commented on this, Call it A “mega chart screw -up,” however he said that there is a correct version Openai’s blog post,
An Openai Marketing Staff also ApologizeSaying this, “We decided the chart in the blog people, inadvertently apologize for chart crime.”
Openai did not immediately respond to the remarks request. And while it is not clear whether Openai is used GPT -5 To make a really chart, it is still not a great look for the company on its big launch day – especially when it is avoiding “significant progress in reducing hallucinations” with its new model.
