OpenAI Releases Report on GPT-4o Model with Details

TapTechNews August 10th news, OpenAI company released a report on August 8th, outlining the SystemCard of the GPT-4o model, introducing many details including the external Red Team (simulating enemy attacks), Preparedness Framework, etc..

OpenAI stated that the core of the GPT-4o model is the Preparedness Framework, which is a systematic approach to assess and reduce the risks associated with artificial intelligence systems. TapTechNews learned from the report that this framework is mainly used to identify potential dangers in areas such as network security, biological threats, persuasion, and model autonomy.

In addition to the safety assessment and mitigation measures for GPT-4 and GPT-4V, OpenAI has also carried out more safety work for the audio function of GPT-4o.

The evaluated risks include speaker identification, unauthorized voice generation, potential generation of copyrighted content, unfounded inferences, and unpermitted content. Based on these assessment results, OpenAI has implemented safeguard measures at both the model and system levels.

OpenAI also cooperated with more than 100 external Red Team members to evaluate the model before releasing it to the public. The Red Team members will conduct exploratory capability discovery, assess the new potential risks brought by the model, and stress test the mitigation measures.