Openai updates the AI model operator, the AI agent who can navigate the web autonomously and use specific software in a virtual machine hosted in a cloud to fulfill users’ requests.
Soon, the operator will use an O3 -based model, one of the latest Openai models for “reasoning” models. Previously, the operator was based on a customized version of the GPT-4O.
With many reference points, the O3 is a much more advanced model, especially in mathematical and reasoning tasks.
“Replaces the existing GPT-4O-based model for the operator with a version based on Openai O3”, Openai I wrote in a blog post. ‘The API version [of Operator] will remain on the basis of 4o. ”
The operator is one of the many tools released by AI companies in recent months. Companies are struggling to make extremely sophisticated factors that can do reliable jobs more or less without supervision.
Google offers a “computer use” agent through API Gemini, which can also browse the web and take actions on behalf of users, as well as a more consumer offer called Mariner. Anthropic models are also able to perform computer tasks, including file opening and web navigation.
According to Openai, the new operator model, called O3 Operator, “was perfected with additional security data for use on the computer” including data sets aimed at “teaching the model [OpenAI’s] Decision limits for confirmations and denials. ”
Openai has published a technical report showing O3 operator performance in specific security ratings. Compared to the GPT-4O operator model, the O3 operator is less likely to refuse to perform “illegal” activities and seek sensitive personal data and less sensitive to an AI attack known as direct injection according to the technical report.
“The O3 Operator uses the same multi -layer approach for the safety we used to version 4o of the operator,” Openai writes in place on the blog. “Although the O3 operator inherits the O3 coding capabilities. It has no inherent access to a coding environment or terminal.”
