Anthropic, one of the most promising start-ups in the field of artificial intelligence, has launched an improved version of its Claude 3.5 model, called "Sonnet". This new development stands out for its ability to interact with desktop applications and perform tasks on the computer, mimicking human actions such as moving the cursor, clicking buttons or pressing keys. Through its new "Computer Use" API, currently in beta, Claude 3.5 Sonnet can see what is happening on a user's screen, analyze images and execute commands in specific software, offering a more advanced level of automation.
The company defines this update as a major step towards the creation of more sophisticated and autonomous virtual assistants, capable of taking on tasks that previously required human intervention. This developmentresponds to the growing demand for AI tools that not only process information, but can also take concrete actions in digital environments, facilitating the automation of repetitive or administrative tasks in sectors such as commerce, data management and customer service.
Task automation and possibilities for developers
The main objective of this release of Claude is to facilitate task automation in desktop applications, which could significantly transform the workflow in companies that rely on software to manage large volumes of information. Through the "Computer Use" API, developers can integrate Claude's functions into their own platforms and services, allowing AI to take over functions such as spreadsheet management, email management, database browsing, among others.
Unlike other similar systems, Claude 3.5 Sonnet is based on an approach that allows it to learn from its interaction with the software in real time. This means that, rather than following a predetermined set of rules, the model can adapt and improve its performance as it encounters obstacles in its path. For example, if it cannot perform an action on its first attempt, Claude can try again in different ways until it completes the task.
This self-correcting capability sets it apart from other models on the market, and is one of the reasons Anthropic sees it as a key player in computer task automation. However, the system is not foolproof. During recent tests, Claude 3.5 Sonnet had difficulty completing some common tasks, such as modifying a flight reservation or initiating a product return. In evaluations designed to measure its effectiveness in automating processes, the model only managed to successfully complete less than half of the assigned tasks, demonstrating that while the potential is enormous, there are still areas that need refinement.
Advanced AI challenges and risks
While Claude's ability to interact with software and perform autonomous actions represents an exciting advance, it also raises questions about the risks associated with this technology. AI systems that have access to applications and user data can be vulnerable to security attacks, as a recent study has shown.
Even models without access to desktop applications, such as OpenAI's GPT-4, have been induced to perform harmful tasks, such as acquiring illegal documents on the dark web, when attacked with jailbreaking techniques. This suggests that, as AI models gain more autonomy, the risks of them being misused increase. Anthropic recognizes these risks and claims to have implemented several security measures to prevent misuse of Claude 3.5 Sonnet.
For example, the model was not trained on user screenshots and is designed to prevent high-risk actions, such as posting to social networks or interacting with government websites. In addition, any screenshots generated by the "Computer Use" API are stored for at least 30 days, enabling comprehensive tracking of activities and facilitating the detection of potential abuse.
In addition, Anthropic maintains the ability to restrict access to certain applications and websites if the need arises, providing additional control over the environment in which the model is used. The company also collaborates with security agencies such as the AI Security Institute in the U.S. and U.K. to assess risks before deploying its models on a mass scale.
A more economical model on the way
Although the upgraded Claude 3.5 Sonnet has been Anthropic's main release, the company has also announced the upcoming launch of a more economical and efficient variant: Claude 3.5 Haiku. This new model, which will arrive in the coming weeks, is designed to offer the same performance as the Claude 3 Opus model, but with lower cost and faster response times. Claude 3.5 Haiku will focus on tasks that requireagile and efficient processing of large volumes of data, such as purchase history analysis or inventory management, making it ideal for companies looking to integrate AI into user-facing products.
Initially, Haiku will be a text model, but Anthropic plans to release a multimodal version that can also analyze images. In summary, with the introduction of Claude 3.5 Sonnet and the upcoming arrival of Claude 3.5 Haiku, Anthropic is cementing its place in the competitive AI market, offering solutions for bothcomplex process automation and the efficient management of big data. Meanwhile, the company remains committed to the responsible development of AI, implementing security measures that seek to mitigate the risks associated with the use of these advanced models.