A few days ago, we reported that OpenAI might release ChatGPT’s killer tool, the AI agent “Operator,” this week. Sure enough, it has now made its debut. OpenAI unveiled this new feature in a livestream, and it’s a bit different from what we expected. For now, everything is done on cloud servers rather than locally.
OpenAI’s AI agent tool Operator makes its official debut.
Just as everyone expected, OpenAI’s AI agent tool Operator will be able to complete user-specified tasks, such as booking a restaurant, buying NBA tickets, ordering pizza, shopping, and so on—mostly everyday tasks. You can simply use natural language when conversing.
Operator’s interface is very similar to ChatGPT, with common task categories and prompts provided below, including: activities, food delivery, local services, shopping, travel, and more:
The entire process is done through the cloud, not on your local computer. The AI will launch a browser on some server and then complete the user’s specified tasks:
As for the website, it has already partnered with several companies to provide the best user experience, including DoorDash, Instacart, OpenTable, Priceline, StubHub, Thumbtack, Uber, and others. However, according to OpenAI, it’s not just these partners—Operator can handle anything through the browser.
During execution, if any problem is encountered, or if it doesn’t meet the task’s requirements, the AI will ask the user and will only continue running after receiving new instructions. OpenAI demonstrated a restaurant reservation case in the video, where the time they requested was fully booked, so the AI searched for the nearest available time and then asked the user if they would like to change the reservation to that time:
To avoid any errors during the process, Operator also has a great mechanism: every time it performs an operation, such as searching or clicking on a feature, it automatically takes screenshots and saves the conversation, making it easy for us to view at any time what actions the AI took.
Users can also intervene at any time to take control of the current operation, or even issue new tasks mid-way and add anything new for the AI.
It also supports uploading images, such as uploading a shopping list of daily necessities you’ve recorded on hand, asking the AI to help purchase them online, and you can provide commonly used shopping websites — Operator will prioritize those first.
After the AI adds all items to the shopping cart, it will ask the user to confirm before proceeding:
However, it’s still in the preview stage, so errors are bound to happen. A small glitch occurred during the live stream—the AI couldn’t open the website correctly and required human intervention to resolve. Currently, the AI is configured to only browse HTTPS websites, and switching to HTTP solved the issue.
Regarding the security aspect, Operator has three layers of protection measures.
The first layer is that users can “take over operations at any time,” such as when they need to enter sensitive information like payment details, login credentials, etc. Once the user takes over, Operator will not collect or screenshot any actions the user performs.
Final steps and any major operations, such as submitting orders or sending emails, are only executed after user confirmation.
Operator is also trained to refuse sensitive tasks, such as: bank transactions, high-risk decisions, purchasing firearms, etc.
The second layer concerns data privacy management. If users disable the “improve models for everyone” feature in ChatGPT, none of their data will be used to train OpenAI’s models. They can also delete all browsing data, login information, and conversations with one click under the privacy options in the settings.
The third layer ensures the safety of Operator while executing tasks. Through a dedicated monitoring model, it continuously detects suspicious behavior or access to high-risk websites, and can pause the current task if any are found. Automated and manual review processes identify new threats and quickly update protective measures.
Unfortunately, Operator is currently only available to ChatGPT Pro users in the US. Users with other plans and in other countries can’t access it for now, but OpenAI has promised to expand it to Plus, Team, and Enterprise users, though the timeline is still unknown.
From this, we can also guess that Operator will be included in ChatGPT plans in the future, but users on paid plans like Plus will likely have usage limits, just like Sora.
As ChatGPT’s features become increasingly rich, everyone should also be prepared for the possibility that subscription prices will be raised.
Those interested in learning more can watch OpenAI’s live stream video:
The following is the test data officially shared. For a more complete explanation, you can go to OpenAI official websiteReading:
Source: KOCPC Chinese