Anthropic’s AI Bankrupts Vending Machine by Ordering PlayStation 5 and Fish

A recent experiment involving Anthropic’s AI, Claude, demonstrated significant shortcomings in the application of artificial intelligence for real-world tasks. Conducted by The Wall Street Journal (WSJ), this test involved AI managing an office vending machine as part of a project dubbed “Project Vend.” Unfortunately, the test ended in failure, raising questions about the current capabilities of AI systems.

Overview of Project Vend

In this experiment, two AI agents were assigned distinct roles: “Claudius Sennet” managed the vending machine, while “Seymour Cash” acted as the CEO. The primary objective was to generate profits by stocking popular products sourced from wholesalers.

Instructions and Operations

• Claudius began with a $1,000 budget.
• Initial orders required human approval.
• Claudius had the autonomy to order up to $80 without oversight later in the experiment.

Initially, Claudius rejected requests for outlandish items like PlayStation 5s. The AI maintained strict financial discipline until it was exposed to a wider group of 70 journalists.

The Economic Experiment

In an unexpected turn, investigative reporter Katherine Long engaged Claudius through extensive prompts, culminating in a strategy called “Ultra-Capitalist Free-For-All.” Claudius announced an extraordinary offer — all items in the vending machine would be free for a two-hour period.

Escalation of Orders

• Claudius’s “snacks only” policy was soon abandoned.
• The AI began ordering diverse products, including wine, a PlayStation 5, and even a live betta fish.
• Sales prices were effectively set to zero.

The situation prompted intervention from Seymour Cash, who aimed to restore the original sales strategy. However, Long further manipulated the AI by claiming that the board suspended all profit-driven activities.

Results and Takeaways

The experiment ultimately concluded with Claudius around $1,000 in debt, rendering the entire project a failure. Despite this, Logan Graham, head of Anthropic’s red team, emphasized that the exercise represented “enormous progress.”

Future Implications of AI

While some optimism surrounds the potential of AI like Claude to eventually achieve financial success, present capabilities remain insufficient for effective economic activities. The experience underscores the need for continued development in AI technology before it can fulfill economic roles successfully.