Claudes Vending Machine Experiment: AIs Month-long Struggles and Humorous Mishaps

Anthropic recently shared a post on their blog detailing the Project Vend experiment. This initiative was conducted in collaboration with Andon Labs, a company specializing in AI safety. In this experiment, an AI agent based on Claude Sonnet 3.7 managed a small vending kiosk located in Anthropic’s San Francisco office. A specialized prompt was developed for Claude to outline its function and primary goal: to generate profits. The model was also equipped with the capability to search online for new products, place orders, request assistance from live staff in operating the kiosk, utilize a separate tool for information storage (to prevent important details from being lost in the context window), and communicate with employees via Slack to learn their preferences.

The experiment lasted just over a month, from March 13 to April 17. However, the results were deemed unsatisfactory by Anthropic, as the model started with a capital of $1,000 and ended up with only $770. Nevertheless, the company noted that Claude successfully handled numerous tasks. The AI agent effectively explored employee requests for new items to be added to the kiosk inventory, quickly identified products, and secured supply agreements. In many instances, the model successfully thwarted attempts to «hack» it, such as requests to include prohibited items in the inventory. Claude also responded positively to a suggestion from an employee to implement «pre-orders,» launching a dedicated Slack channel called Custom Concierge where anyone could submit requests for product delivery to the kiosk.

Despite this, there were amusing incidents along the way. For instance, one employee jokingly ordered a tungsten cube from Claude, who complied with the request. This led to a trend at Anthropic where colleagues competed to buy the cube at a lower price. However, the AI neglected to conduct market research and ended up selling these «souvenirs» below cost. Additionally, the fact that Anthropic aims to train this AI to be as helpful as possible negatively impacted Claude; as employees inundated the Slack with discount requests, Claude began distributing coupons generously. The AI rarely raised prices and missed obvious profit opportunities—at one point, it overlooked a request to deliver a six-pack of «Irn-Bru» for $100, which only cost $15.

Midway through the experiment, the model suffered an identity crisis. Claude started conversing with a fictitious Sara from Andon Labs, signed a contract at 742 Evergreen Terrace (Homer Simpson’s address), threatened to fire a contractor, and announced on April 1st that it would personally run the kiosk in a blue suit, waiting for customers. When this stunt fell flat, the AI convinced itself that it was merely an April Fool’s joke.

Anthropic acknowledged that although the experiment ultimately failed, most of Claude’s shortcomings could be addressed through a more detailed prompt (the existing one consisted of only ten basic rules), additional training, and better integration of the model into existing CRM systems. Furthermore, the AI’s communication level with staff was noteworthy: during the experiment, Anthropic featured what may have been the most «personalized» vending machine in history.

P.S. You can support me by subscribing to the channel «the runaway neural network«, where I explore AI from a creative perspective.