Anthropic Launches Enhanced Claude 3.5 Sonnet Model with Support for Remote Computer Control

Anthropic Unveils Upgraded Claude 3.5 Model, Ushering in a New Era of AI Agents

On October 22, 2023, Anthropic announced the launch of its enhanced Claude 3.5 Sonnet and the new Claude 3.5 Haiku models. The Claude 3.5 Sonnet significantly improves programming capabilities and introduces a new “computer usage” feature. This innovation allows the AI to interact with computers like a human, moving the cursor, clicking, and typing through a virtual keyboard.

The upgraded Claude 3.5 Sonnet is now officially available for users to experience in its testing phase. This new model showcases remarkable advancements in coding and other areas. Jared Kaplan, Chief Scientist at Anthropic, stated, “We are entering a new era where models can use a variety of tools to complete tasks as humans do.”

The debut of Claude 3.5 Sonnet represents a significant advancement in the commercial AI model landscape. Its goal is to transcend traditional chat interfaces, evolving into a true "AI agent." These AI agents can operate software and perform computer tasks in a manner akin to human users. While other models, such as Cognition AI's Devin, focus on programming, Anthropic's AI agents boast broader capabilities, including web browsing and utilizing various applications.

In one demonstration showcased by Wired, Claude was tasked with planning a trip to watch the sunrise at the Golden Gate Bridge. The AI opened a web browser, searched for the best viewing spots, and added the itinerary to a calendar. However, Wired noted that Claude overlooked practical details, such as how to reach the destination. In another example, Claude was asked to create a simple website using Microsoft Visual Studio Code and even managed to open a local server for testing, correcting minor errors when prompted.

Despite the promising potential of these AI models, reliability remains a concern, particularly in coding tasks. TechCrunch reported that Claude 3.5 Sonnet struggled to complete over half of the tasks in simple tests, such as booking flights.

Moreover, the security risks associated with AI agents cannot be overlooked. Users must consider whether they are comfortable allowing these sometimes unstable and unpredictable technologies access to their personal computers and web browsers. In response, Anthropic emphasized that by gradually introducing this limited yet relatively safe AI model, they can enhance security. They stated, “Rather than waiting for more powerful models, it’s better to let existing, safer models interact with computers so we can begin observing and addressing potential issues, while progressively enhancing safety measures.”

With the rapid evolution of AI technology, the release of Claude 3.5 Sonnet marks the beginning of a new chapter for AI agents. The hope is that future advancements will lead to safer and more reliable solutions.