Anthropic's Claude AI Model: Revolutionizing Computer Interaction

If artificial intelligence has you concerned about job security, you might want to pay attention to this news. AI startup Anthropic has introduced a new model named "Claude" that can interact with a computer screen much like a human, using a virtual mouse and keyboard.

In a recent video demonstration, researcher Sam Ringer showcased Claude completing data entry tasks, referred to as "drudge work." Using screenshots from a Mac desktop, Claude found necessary information and filled out a form. This type of work is common in many workplaces worldwide, although Ringer did mention this demonstration is merely a "representative example." The extent of any video editing is unclear.

First-hand Experience with Claude

For those skeptical of the promotional materials, an early version of the Claude 3.5 Sonnet API is currently available for testing. Ethan Mollick, a professor at the University of Pennsylvania's Wharton School, explored its capabilities by utilizing the AI with an online clicker game called Universal Paperclips, which has an intriguing science fiction twist.

Mollick directed Claude to the game’s browser window with the command to "win," and was amazed as the AI took over. It successfully identified game objectives by interpreting the text-based interface and experimented with strategies to win, such as modifying the price of paperclips to boost profits, akin to how an actual player might. However, Claude struggled to optimize the sequences as a human would.

Challenges Faced by Claude

Interestingly, the game Claude was playing involves a fictional AI and presented some logical obstacles that hindered progress. Mollick's testing environment saw several crashes as the AI attempted to complete the game, which lasted several hours. With some encouragement, Claude managed to generate basic code to automate its gameplay.

This scenario illustrated Claude writing virtual code for a virtual game—setting a rather conceptual stage. After numerous VM crashes, Claude announced it had "successfully 'won'" the game after achieving a specific milestone within given limits.

It should be noted that Claude did not truly win Universal Paperclips; the level of engagement in this complex game surpasses the basic automation demonstrated in Anthropic's promotional video. Nonetheless, Claude’s ability to set goals and progress with minimal input was undeniably impressive. Mollick's full assessment of the AI is recommended for those interested.

Availability and Applications of Claude AI

According to Professor Mollick, Claude displayed flexibility and persistence in the face of errors, applying clever tactics such as A/B testing. Most importantly, it managed to perform its tasks continuously for nearly an hour without interruptions.

Anthropic has made Claude AI accessible as a free text-based tool online as well as an app available on iOS and Android, with features that allow for image and text document inquiries. The upgraded version 3.5 is now live in the free version, while advanced users can access a $20 per month Pro account that offers priority bandwidth and additional models. Anthropic reportedly counts several companies among its current clients, including Notion, Intuit (the creators of TurboTax), and Zoom.

AI, Claude, Technology