[AI Hanip News] "ChatGPT is at a Dumb Level"... Anthropic Fuels the Agent War

Anthropic Unveils AI Agent for PCs
Developed by MS and OpenAI Too... Efficient at Complex Tasks
Concerns Over PC Misuse Like Information Manipulation and Theft

# Client company A requests documents related to 'vendor requests.' The data is scattered throughout various computer files. Instead of digging through documents, they ask the AI model 'Claude.' Claude moves the computer mouse to open an Excel file and searches for company A, but finds nothing. This time, Claude accesses the Customer Relationship Management (CRM) site, scrolls through to find information on company A, and prepares the documents.

This is a scene from a video released on the 22nd by Anthropic, considered a rival to OpenAI. Anthropic introduced a feature called 'computer Use' based on its latest AI model, 'Claude 3.5 Sonnet.' As the name suggests, the AI uses a computer to perform complex tasks on behalf of users. This represents an evolution beyond the previous method of assigning step-by-step tasks to AI models, such as giving files and instructing them to find specific content or document the findings. It is the first AI model capable of performing multiple stepwise tasks by actually using a computer.

Operating a computer requires quite complex skills. It involves the ability to view and interpret the screen, infer when and how to perform specific tasks based on screen content, and use software effectively. Although it seems simple, even clicking requires precisely calculating how many pixels to move the cursor vertically or horizontally to click the intended spot.

Claude, having acquired these skills through training, can now use a computer autonomously. While this can be seen as the pinnacle of automation, it also poses risks close to a disaster. Anthropic’s Trust & Safety team has identified potential vulnerabilities in the model. For instance, the AI might perform tasks it was not instructed to do or ignore commands. In fact, during a demonstration recording, Claude accidentally stopped the recording, causing the video to be lost, and at one point, it got distracted by looking at photos instead of completing the task.

Beyond these cute mistakes, if the computer is deliberately misused, dangerous situations could arise. To prevent this, Anthropic designed the AI to prohibit 'high-risk actions' such as creating social media accounts, posting content, or accessing government websites. Additionally, with the U.S. presidential election approaching, the AI is blocked from generating or spreading election-related content or participating in election-related activities.

The competition for AI agents that autonomously handle complex tasks is intensifying. On the 21st, Microsoft (MS) announced a feature called 'autonomous agents.' Autonomous agents are AI systems that independently judge and act, automating a wide range of tasks from simple repetitive work to complex decision-making to support users’ work.

MS Copilot Studio [Image=MS]

Next month, MS plans to offer a feature in 'Copilot Studio' that allows users to create their own autonomous agents?essentially building personalized AI assistants. MS shared a case study of an agent developed for the UK pet care company Pets at Home, which is expected to save several million won annually by assisting the sales management team.

Furthermore, MS will introduce 10 autonomous agents in its business application 'Dynamics 365,' including a sales agent that researches potential customer groups, prioritizes targets, and sends customized emails; a supply chain agent that monitors supply chain status and detects and responds to delays; and a customer management agent that assists customer service representatives.

What about OpenAI, at the forefront of the AI battle? As soon as Anthropic unveiled its computer-using AI, news emerged that OpenAI had already developed a similar agent. According to IT media The Information, OpenAI conducted internal demonstrations of agents capable of working on computers, and coding agents that can perform programming tasks are already in use. There are reports that AI agents have ordered food delivery online or searched websites to solve coding problems. OpenAI CEO Sam Altman said, "Compared to agents, ChatGPT is at a dumb level," and has been actively involved in agent development early on. Since Anthropic took the lead with AI agents, it is expected that OpenAI will soon reveal its technology as well.