Anthropic has a new AI model that understands what’s happening on your computer screen and executes commands, just like human users would do.
An upgraded version of Claude 3.5 Sonnet now has a feature called “computer use.” Currently available on the API, this capability allows the AI model to automate tasks and perform actions on desktop applications. It depends on a correct configuration and user prompts so the AI knows how to move the cursor, which UI elements to click on, and what input data to enter via the virtual keyboard.
Anthropic says that the “computer use” API is still an experimental feature in public beta, so expect slow performance and errors. However, feedback from early adopters should help the company improve its capabilities over time. In one evaluation designed for testing models to use computers, Claude achieved a mere 14.9 percent success rate. It’s a far cry from human proficiency, which is in the range of 70-75 percent, although Anthropic claims it’s still better than the 7.7 percent success rate of the next-best AI model.
The ability to directly control computers poses challenges to safety and security. But per Anthropic’s classification to evaluate safety and security measures, the upgraded Claude model remains within AI Safety Level 2 safeguards and doesn’t require more stringent precautions. Still, Anthropic warns its users of cyberattacks such as prompt injection, which allows hackers to manipulate the AI’s actions.
Besides the new computer use feature, the upgraded Claude 3.5 Sonnet has improved performance over its predecessor. Based on several evaluations, it demonstrated better reasoning, superior proficiency in coding tasks, and better problem-solving skills in math. Anthropic also plans to release Claude 3.5 Haiku later this month, an updated version of its fastest and most compact AI model.
This article, Anthropic’s upgraded Claude 3.5 Sonnet can directly control your computer, was originally published at NoypiGeeks | Philippines Technology News, Reviews and How to's.
Comments
Post a Comment