US: OpenAI has unveiled one of its first AI agents called Operator, a system that can use its own browser to make travel reservations, fill out forms, order groceries and even create memes.
Operator, which is now available to ChatGPT Pro users in the United States at operator.chatgpt.com [a $200 monthly plan providing access to latest models], is designed to perform tasks autonomously, including looking at webpages and interacting with them by typing, clicking and scrolling.
OpenAI CEO Sam Altman has called the release “an early research preview” that currently has limitations and will evolve based on user feedback over the coming months. OpenAI also plans to launch more agents post-Operator launch, and it will be open to ChatGPT Plus, Team and Enterprise users across the world beyond that.
Powered by a new model called Computer-Using Agent [CUA] – built on GPT-40, Operator “sees” through screenshots and “interacts” via mouse and keyboard actions within a browser, enabling it to take action on the web without requiring custom API integrations.
Users can prompt Operator with requests, either by going through specific partner websites or going through a traditional search engine such as Google. If the AI agent encounters challenges or makes mistakes, it can use its reasoning capabilities to self-correct, although it is still in its early stages and has its own limitations, which leaves it some way behind humans’ ability to navigate the web currently.
The goal of Operator is to “transform AI from a passive tool to an active participant in the digital ecosystem” by streamlining tasks for users and creating innovative customer experiences to drive higher rates of conversion.
Operator’s ecosystem already includes “early contributors” from the travel and mobility sectors, such as Booking.com, Hipcamp, Tripadvisor, Uber and Priceline, to help make reservations and ensure Operator “addresses real-world needs while respecting established norms”. Other partner companies include DoorDash, Instacart, OpenTable, eBay, Etsy, StubHub, Thumbtack and Target.
Alyssa Ravasio, founder and CEO of Hipcamp, said: “Hipcamp has always focused on harnessing the best technology to increase access to the outdoors. We are proud to work with OpenAI on the early research preview of Operator and to have the opportunity to leverage the world’s best AI technology to get more people outside and under the stars. In a future where technology is becoming more omnipresent, we believe spending time outside in nature is more essential than ever.”
To get started on Operator, users can describe the task that they would like to be carried out and Operator can then take over. However, users can still choose to take over control of the remote browser at any point, and Operator is automatically trained to ask them to take over for tasks that require a login, payment details or CAPTCHA puzzle.
In addition, users can personalise their Operator workflows by adding custom interactions for all or specific websites, for example when repeated tasks are necessary. Similar to using multiple tabs on a browser, Operator can run multiple tasks at any given time.
OpenAI was keen to stress that safety for users is a top priority, and has implemented three layers of safeguards designed to prevent abuse and ensure that users are firmly in control.
These include:
• Training Operator to ensure that a person using it is always in control and can ask for input at important points
• Data privacy can be managed as easily as possible
• Defences have been built against “adversarial” websites that may try to mislead Operator through hidden prompts, malicious code, or phishing attempts
Looking ahead, OpenAI said that it plans to expose the model powering Operator [CUA] in the API [application programming interface] soon so that developers can utilise it to build their own computer-using agents. Operator will also be developed to handle longer and more complex workflows, and be made accessible to users of other plans in the future.
