Like Anthropic’s Laptop Use and Google DeepMind’s Mariner, Operator takes screenshots of a pc display screen and scans the pixels to determine what actions it might probably take. CUA, the mannequin behind it, is educated to work together with the identical graphical consumer interfaces—buttons, textual content bins, menus—that individuals use after they do issues on-line. It scans the display screen, takes an motion, scans the display screen once more, takes one other motion, and so forth. That lets the mannequin perform duties on most web sites that an individual can use.
“Historically the way in which fashions have used software program is thru specialised APIs,” says Reiichiro Nakano, a scientist at OpenAI. (An API, or software programming interface, is a chunk of code that acts as a form of connector, permitting completely different bits of software program to be hooked as much as each other.) That places loads of apps and most web sites off limits, he says: “However in case you create a mannequin that may use the identical interface that people use each day, it opens up a complete new vary of software program that was beforehand inaccessible.”
CUA additionally breaks duties down into smaller steps and tries to work by means of them one after the other, backtracking when it will get caught. OpenAI says CUA was educated with strategies much like these used for its so-called reasoning fashions, o1 and o3.
OPENAI
OpenAI has examined CUA in opposition to a lot of trade benchmarks designed to evaluate the flexibility of an agent to hold out duties on a pc. The corporate claims that its mannequin beats Laptop Use and Mariner in all of them.
For instance, on OSWorld, which assessments how effectively an agent performs duties reminiscent of merging PDF information or manipulating a picture, CUA scores 38.1% to Laptop Use’s 22.0% As compared, people rating 72.4%. On a benchmark known as WebVoyager, which assessments how effectively an agent performs duties in a browser, CUA scores 87%, Mariner 83.5%, and Laptop Use 56%. (Mariner can solely perform duties in a browser and due to this fact doesn’t rating on OSWorld.)
For now, Operator may solely perform duties in a browser. OpenAI plans to make CUA’s wider skills accessible sooner or later through an API that different builders can use to construct their very own apps. That is how Anthropic launched Laptop Use in December.
OpenAI says it has examined CUA’s security, utilizing pink groups to discover what occurs when customers ask it to do unacceptable duties (reminiscent of analysis how one can make a bioweapon), when web sites comprise hidden directions designed to derail it, and when the mannequin itself breaks down. “We’ve educated the mannequin to cease and ask the consumer for data earlier than doing something with exterior unwanted effects,” says Casey Chu, one other researcher on the crew.
Look! No arms
To make use of Operator, you merely sort directions right into a textual content field. However as an alternative of calling up the browser in your laptop, Operator sends your directions to a distant browser working on an OpenAI server. OpenAI claims that this makes the system extra environment friendly. It’s one other key distinction between Operator, Laptop Use and Mariner (which runs inside Google’s Chrome browser by yourself laptop).