Anthropic has introduced a significant upgrade of its AI range with the Sonnet Claude 3.5 modelwhich has an unprecedented capacity for an AI to control a computer like a human. This new feature, well -named “use of computers”, is currently available in public beta, allowing developers to interact to interact with office computers, click on buttons and even type text by observing screenshots and reproducing human actions.
Unlike other technology giants, such as Microsoft and Openai, who presented similar features but limited their tools to the visualization of screens without complete operational control, Anthropic has taken a daring step. Claude 3.5 can now fully engage with applications and automate workflows – potentially transforming research processes with routine administrative tasks.
The idea of an AI working directly on a computer like a human is not entirely new. Companies specializing in the automation of robotic processes (RPA) offer similar tools for years, but the approach of anthropic has integrated AI with a level of generality and flexibility which traditionally lacks by RPA. Rather than using predefined automation scripts, the user -use function of Claude 3.5 offers developers the possibility of directing AI using natural language, asking it to manage repetitive tasks, carry out open searches and even carry out more complex operations.
Anthropic has integrated this functionality via an API, allowing users to ask Claude, for example, to collect data from various sources and fill out a form or compile information from several applications. The model works by “seeing” what is on a screen through a series of screenshots that writes together to form a coherent view of the office. Then, on the basis of the instructions provided, he simulates actions such as moving a cursor, click buttons or entry.
Although promising, the functionality remains experimental. Claude's dependence on a series of fixed images rather than a real -time video flow can make rapid actions, as reacting to notifications, difficult. Anthropic warns that certain tasks, such as trail and zoom, are always present, and there are plans for continuous improvements based on the comments of the first adopters.
Claude 3.5 Sonnet has demonstrated impressive results on industry landmarks, with improved scores on tasks requiring coding and specific use of tools. It marks significantly more on Swe -Bench verified, a coding reference, increasing its performance to 49% – better than the AI models accessible to the public. On Tau-Bench, which assesses how AI can manage real world tasks in fields such as retail and airlines, Claude's precision has also increased significantly.
Safety and ethical considerations have been an absolute priority for Anthropic in the release of this technology. In response to concerns concerning potential improper use, such as the spread of disinformation or electoral interference, Anthropic has designed Claude to avoid engaging with social media, government websites or areas associated with sensitive data. Specific prompts that could lead to risk behaviors are reported, and Claude is designed to avoid high -risk actions, except explicitly led by a human operator.
In addition, the model is equipped with classifiers who monitor its activity. These classifiers detect any attempt to publish social networks or registration of the field. For more responsibility, Anthropic retains screenshots of Claude's sessions for at least 30 days, guaranteeing a trace of his actions which could be examined if necessary.
Anthropic recognizes that this is only the beginning. The current version of Claude 3.5 Sonnet serves as a test field, and the ideas acquired by user comments will help the company improve its performance and security protocols. Although the model's ability to reproduce human interaction with office computers opens up exciting possibilities, it also presents new challenges. Anthropic closely monitors its adoption to balance innovation with responsible use of AI.
To respond to customers more sensitive to prices, Anthropic is also preparing to publish Claude 3.5 Haiku, a more profitable version of the model, which will offer similar reference performance but to lower latency. Claude 3.5 Haiku will initially be available as a text model only, but will eventually develop to support multimodal applications, managing both text and image analysis.
