OpenAI announces new agent tool “Operator”

Two years ago, OpenAI started the chatbot boom with the release of ChatGPT. Now they want to spark interest in a new wave of AI technology.

OpenAI on Thursday announced a tool called Operator that can connect to the internet and autonomously perform tasks like shopping for groceries or making restaurant reservations.

“They can navigate the website and perform actions on the website just like you and I can,” Yash Kumar, product and engineering lead at OpenAI, said in an interview.

Artificial intelligence researchers refer to this type of technology as an AI agent. Chatbots can answer questions, write poems, and generate images, but agents can use other software on the internet.

In a press conference with The New York Times, Kumar demonstrated how the system allows people to make restaurant reservations in San Francisco through the OpenTable website and purchase a grocery list through Instacart. Operator looks and behaves very similar to ChatGPT and other chatbots. Users enter their requests in a small window. The system then returns the best possible response.

Users can watch as the tool opens a web browser and visits a specific site. Operators can make mistakes. However, in some cases, these mistakes can be corrected. During the Times demonstration, the system incorrectly assumed Mr. Kumar was in Iowa before correctly finding a restaurant in San Francisco.

Operators are not fully autonomous. In some cases, users will be required to correct mistakes and provide additional requests and suggestions. Sites like OpenTable and Instacart require users to enter a private username and password. However, OpenAI said it does not store this personal information.

However, the company captures data that shows how its systems interact with you and access the site on your behalf. This data can be used to train future versions of the Operator.

Starting Thursday, OpenAI says Operator will be available to anyone who subscribes to ChatGPT Pro, a $200-per-month service that provides access to all of the company's latest tools. This tool will be offered through other paid services and will eventually be included in the free version of ChatGPT. Users in the United States will be the first to receive this new tool.

(The New York Times sued OpenAI and its partner Microsoft for copyright infringement of news content related to AI systems. OpenAI and Microsoft have denied these claims.)

Other major companies, including Google and Anthropic, have announced similar tools in recent months. However, many of these tools are not yet widely available.

Operator is based on the same technology that powers ChatGPT. The technology is what AI researchers call a neural network, a mathematical system that can learn skills by analyzing vast amounts of data.

New versions of this technology learn from a wide range of data, including text, images, and audio. In this case, operators learned from images that show how people use spreadsheets, shopping sites, and other online services. After identifying patterns in this data, the new system can use similar services on behalf of computer users.

Kumar acknowledged that Operator, like ChatGPT and other chatbots, is still an experimental technology. But he said the situation would continue to improve in the coming months.

“It’s not the most robust thing in the world,” he said. “But this kind of technology is much better than before.”