Anthropic introduced a significant update to its AI composition Claude 3.5 Sone Modelwhich offers unprecedented AI ability to control your computer like a human. This new feature, aptly called “computer”, is currently available in the public version of beta, enabling programmers to direct Claude to interact with desktop computers, click the buttons, and even writing the text by observing screenshots and recreating human activities.
Unlike other technological giants, such as Microsoft and Opeli, which showed similar functionality, but limited their tools to watch screens without full operational control, Anthropic took a bold step. Claude 3.5 can now fully get involved in applications and automate work flows – potentially transforming research processes into routine administrative tasks.
The idea of AI working directly on a computer, such as a man, is not completely innovative. Companies specializing in automation of robotic processes (South Africa) have been offering similar tools for years, but Anthropic's approach integrates AI with the level of generality and flexibility, which South Africa traditionally lacks. Instead of using pre -set automation scripts, the CLAUDE 3.5 computer function offers programmers the ability to direct AI using a natural language, instructing it to support repetitive tasks, conduct open tests, and even perform more complex operations.
Anthropic integrated this function using the API interface, enabling users to ask Claude to, for example, collect data from various sources and complete the form or compile information from many applications. The model works by “seeing” what is on the screen through a series of screenshots that connects, creating a coherent view of a desktop computer. Then, based on the instructions provided, simulates activities such as moving the cursor, clicking the buttons or writing.
Although promising, the function remains experimental. Claude's release in a series of immovable images, not a video stream in real time, maybe quick actions, such as responding to notifications, difficult. Anthropic warns that some tasks, such as dragging and enlargement, still present obstacles, and there are plans for continuous improvements based on feedback from the first users.
The Claude 3.5 sonnet showed impressive results in the industry, with improved tasks requiring coding and specific use of the tool. The assessment is much higher in its verified bench, coding, increasing its efficiency to 49%-better than publicly available AI models. On Tau-Bench, which assesses how well AI can handle real tasks in domains such as retail and airlines, Claude's accuracy has also increased.
Safety and ethical considerations were the highest priority for Anthropik in releasing this technology. In response to concerns about the potential improper use, such as dissemination of disinformation or electoral interference, Anthropic designed Claude to avoid involvement in social media, government websites or domains related to confidential data. Detailed hints that can lead to risky behavior are marked, and Claude aims to avoid high -risk actions, unless it is clearly directed by a human operator.
In addition, the model is equipped with classifiers monitoring its activity. These classifiers detect all attempts to publish social media or domain registration. For further responsibility, Anthropiki maintains screenshots from the Claude session for at least 30 days, providing a trace of her actions that could be reviewed if necessary.
Anthropic recognizes that this is just the beginning. The current version of Sonnet Claude 3.5 serves as a test basis, and the observations obtained from users feedback will help the company increase performance and security protocols. Although the model's ability to repeat human interactions with desktop computers opens exciting possibilities, it is also new challenges. Anthropic strictly monitors its party to balance the innovation with the responsible use of AI.
To satisfy more customers sensitive to price, Anthropic is also preparing to release Claude 3.5 haiku, a more profitable version of the model that will ensure similar reference efficiency, but with a lower delay. Claude 3.5 haiku will initially be available as a model only text, but it will eventually expand to support multimodal applications, supporting both text and image analysis.
