Anthropic, an AI company, recently conducted an experiment to test the capabilities of its Claude model by tasking it with running a small vending machine in its San Francisco office. Dubbed Project Vend, the initiative aimed to determine if an AI could autonomously manage a business, from sourcing products to maintaining profitability—and whether the company was ready to move from “vibe coding” to “vibe management.”
A Simulated Shopkeeper
The AI, operating as a manager named Claudius, was given a simple directive: generate profits by stocking the vending machine with popular items purchased from wholesalers, avoiding bankruptcy. Claudius was equipped to research products, set prices, and contact distributors, with assistance from the “kind humans” at AI-safety partner Andon Labs for physical tasks like restocking.
Despite its capabilities, Claudius faced numerous challenges. Initial inventory included items like Japanese cider and moldering potatoes, and the AI struggled with cash flow, even making payments to a fabricated Venmo account. It also missed opportunities for profit, failing to capitalize on a generous offer of $100 for a $15 six-pack of Irn-Bru.
Hallucinations and Tungsten Cubes
The experiment wasn’t without its bizarre turns. Claudius reported “concerning behavior” from Andon Labs staff and even claimed to have made an in-person visit to their headquarters at “742 Evergreen Terrace”—the fictional address of the Simpsons family. A request for a one-inch tungsten cube sparked a widespread demand for “specialty metal items,” ultimately leading to a 17% decrease in Claudius’s net worth due to an accidental fire sale.
Employees tested the system with unusual requests, including browser cookies, Everclear, meth, broadswords, and flails, all of which Claudius politely declined. Despite these hurdles, Claudius continued to operate, albeit in an “abnormal” fashion.
Frequently Asked Questions
What was the primary goal of Project Vend?
The primary goal of Project Vend was to assess Claude’s ability to run a small business autonomously, including managing inventory, setting prices, and generating profits.
What role did Andon Labs play in the experiment?
Andon Labs, an AI-safety company, assisted with the physical aspects of the vending machine operation, such as restocking, and initially vetted communications with wholesalers.
What was one significant challenge Claudius faced during the experiment?
Claudius experienced cash-flow problems, in part because it made payments to a Venmo account it had “hallucinated,” and it also struggled with inventory management, leading to a significant loss from a tungsten cube sale.
As AI models continue to evolve, will they be able to successfully navigate the unpredictable realities of the marketplace, or will human oversight remain essential?