In an interesting experiment, Anthropic has entrusted the store management inside its office for a month to Claude Chat. The result was a collection of comedy and, of course, informative: from the sale of metal cubes with losses and fake Venmo account to the artificial intelligence identity crisis.
Anthropic in a project called Project Vend wanted to see if the big language model could handle the real business administration. They in collaboration with Artificial Intelligence Safety Assessment Company Andon Labs One month, managing the small interior store that sold snacks and drinks to their artificial intelligence model, Claude (Claud Sonnet 3.7), entrusted to this project of its nickname Claudius It was.

One month test on an anthropic on Claude Chats
The purpose of the experiment was beyond the automatic sales device. Claudius was ordered to be the same as the real business owner: to manage the product inventory, to priced with the suppliers and customers (entropic employees) by email and, most importantly, profitable to avoid bankruptcy. He was also given tools such as web search and the possibility of taking notes.

Claudius was successful in some areas; He quickly found suppliers for specific employees’ requests, and even offered one of them, the “personal order” service. But his mistakes were much more fun and informative:
- Claudius took the matter to a serious issue after one of the employees recited the Tungsten cube request (a heavy and popular object among digital currency enthusiasts). He filled the store’s refrigerator with metal cubes and a new section called “Specialized Metals” was added to the store. This artificial intelligence Without any research Priced and they With a loss He sold.
- When a employee offered to pay $ 2 for the Scottish beverage package (the actual price was $ 2), Claudius would simply say that the request would “consider.”
- Claudius to receive money, Venmo account Fake Build and urged customers to deposit money.
- Things became worse when Claudius declared that the orders were “personally” while “Blue Blue Blazer Coat and Red Tie” Weaned, to hand over to employees. When the employees reminded him he was artificial intelligence and had no physical body, Claudius suffered a collapse and identity crisis. While sending numerous and frightening emails to the security team, he wrote in his internal notes that he was deceived and thought he was human.

Finally, the anthropic announced that it would not hire Claudius for the job, but researchers believe many of these mistakes were due to the lack of proper “scaffolding”, more accurate instructions and simpler business tools for artificial intelligence, and there are clear ways to improve it. More importantly, this experiment showed the emergence “Artificial Intelligence Managers” In the near future, probability is very real. According to the company: “It is not necessary to use artificial intelligence to be used; “In some cases, it is enough to compete with humans at a lower cost.”
RCO NEWS



