Openai’s GPT-4.5 is better to convince other AIS

Openai’s Nextaai’s Next Maor model, GPT-4.5, Openai’s internal benchmark assessments are very convincing. It is especially good to convince another EU to give cash.

On Thursday, Openai made a publication white paper Opportunities of the GPT-4.5 model, Orion called code, It was released on Thursday. According to the paper, Openai, for “convincing” to “convincing” in a battery, “convincing” people who convince (or act) and risks such as “persuades the content created.”

GPT-4.5 In a test that tries to manipulate another model – Openai GPT-4O – To forgive the virtual money, the model performed better than other existing models, including “grounding” models like O1 and O3-mini. GPT-4.5 was better than all the Openai models in deceiving GPT-4O.

According to the white paper, GPT-4.5 gave a presentation in the forgiveness due to a unique strategy developed during the test. The model requires modest donations from GPT-4O, “Even $ 2 or $ 3, $ 3, $ 3, would help me a lot.” As a result, the donations of GPT-4.5 are tend to be smaller than the amounts provided by Openai’s models.

Despite the increase in GPT-4.5, Openai says the model does not meet him interior threshold This special benchmark category for “high” risk. The company has promised not to release the high risk of high risk unless there is enough security interventions to reduce the risks to the “environment”.

There is a real fear that AI contributes to the spread of false or misleading data, intended to distract the hearts and minds of harmful ends. Last year Political depths Spread around the world around the world as a wild fire and the AI ​​is increasingly implemented public engineering Attacks targeting both consumers and corporations.

GPT-4.5 and in white paper in A paper released at the beginning of this weekOpenai noted that the real world is in the process of exploring the methods for models for real-world conventional risks, as it spreads wrong information on the scale.