Anthropic, the Amazon-backed competitor to OpenAI, has just rolled out a powerful new update to its Claude 3.5 Sonnet model, allowing AI to control personal computers completely.
This release pushes the boundaries of what AI can do by allowing it to handle fundamental desktop tasks such as executing keystrokes, controlling the mouse, and accessing applications installed on the system.
This update marks Anthropic’s entry into the race to transform AI models from simple chatbots into more capable “AI agents.” These agents aim to perform computer-based tasks just like a human would, from browsing the web to using software for programming or even trip planning.
According to its chief science officer, Jared Kaplan, Anthropic’s vision is for AI to wield all the tools humans rely on to get things done, moving beyond conversation into more practical, hands-on assistance.
Anthropic’s AI agent stands out by being marketed as a versatile tool rather than limited to niche tasks. While some AI agents, like Cognition AI’s Devin, are tailored specifically for programming, Claude can handle a variety of scenarios.
In a demo shared by Wired, the model successfully planned a trip to the Golden Gate Bridge, opening a web browser to find the best viewing spots and adding the details to a calendar. However, the AI didn’t think to include critical travel information, like how to get there, highlighting its limitations.
In another demo, Claude tackled a more technical task by setting up an essential website using Microsoft’s Visual Studio Code. The AI even launched a local server to test the site, though it made a minor coding error, which it corrected upon request. These demos demonstrate the technology’s potential but expose its quirks and occasional clumsiness.
Despite these advancements, there are lingering concerns about the reliability and security of such robust AI systems.
For example, Claude struggled with some seemingly straightforward tasks, like booking flights and modifying reservations, successfully completing fewer than half of them during a test reported by TechCrunch.
These AI agents also raise significant security questions. With models like Claude gaining full access to files and applications on a computer, users must weigh the risks of giving such experimental technology so much control.
Anthropic, however, believes that releasing these tools now is the best way to make them safer. It argues that allowing today’s limited models to operate at this level will enable developers to spot potential issues early and gradually improve safety measures.
While promising, Claude’s evolution into an AI agent leaves users with a tricky question: are the benefits of handing over control to AI worth the potential risks?
