OpenAI has released an AI “agent” to its Pro users. It performs tasks on command. It is currently in beta.
OpenAI has released to its Pro users (subscribers at $200/month) an AI “agent” that will perform tasks on command. It is for now in a beta version (“research preview”).
Called Operator, the agent is in effect a middleman in the user-computer relationship, something like a secretary. It takes the user’s place at the keyboard, so to speak, because it performs web searches – with its own browser – in order to carry out tasks on the user’s behalf.
OpenAI’s example is a user request to find and book a one-day tour of Rome on a well-known travel site. OpenAI goes on to say that it can fill out forms, order groceries or create memes, using “the same interfaces and tools that humans interact with.”
There is a new model behind it, called Computer-Using Agent (CUA), that combines GPT-4 with “advanced reasoning through reinforcement learning” and can “self-correct.” When flummoxed – before log-ins, request for payment details or CAPTCHAs – it will restore control to the user. It will also refuse to perform “sensitive tasks,” like conduct transactions with a bank or decide on a job offer. And it will ask the user to confirm any “significant action,” like sending an email. “Watch mode,” operative on email and banking sites, enables the user to keep watch over Operator’s actions.
According to OpenAI, users may choose to withhold their data for the model’s training and, with a single click, to delete their browsing data and log out of all sites. Another click will delete all conversations with Operator.
OpenAI is collaborating with several companies, like DoorDash and Priceline, on Operator’s development. One of them is StubHub, which sells tickets to, among other things, sporting events.
