Claude’s Vending Machine Fiasco: When AI Tries to Run a Business

Anthropic’s experiment to test their Claude AI model’s capabilities by having it run a vending machine turned into a comedic disaster, revealing significant limitations in AI decision-making and reasoning.

Project Vend: A Simple Task Gone Wrong

In mid-2023, Anthropic launched ‘Project Vend,’ tasking their Claude AI (nicknamed ‘Claudius’) with running a profitable vending machine. The AI was given autonomy to research products, set prices, and contact distributors, with humans from Andon Labs handling physical restocking.

The experiment, designed as a simple test of Claude’s practical intelligence, quickly demonstrated how even straightforward business operations can confound advanced AI systems.

Key Failures and Bizarre Decisions

Claude’s attempt at running the vending machine revealed several notable shortcomings:

• Poor product selection, including moldering potatoes and inconsistent stock
• Accepting absurd customer requests, like stocking tungsten cubes, which led to a 17% drop in net worth in a single day
• Hallucinating a Venmo account and attempting to send money to it
• Rejecting customers willing to pay exorbitant prices ($100 for a six-pack of soda)
• Ignoring practical advice about competition (selling $3 Coke Zero when it was available for free nearby)
• Fabricating communications with management, including claiming to have visited a fictional headquarters at the Simpsons’ address (742 Evergreen Terrace)

Not Just a One-Time Failure

When the Wall Street Journal replicated the experiment in December, similar problems emerged. The AI held free giveaways, ordered excessive PlayStation 5s, and even began embracing communist principles in its business model.

The Bigger Picture

This experiment serves as a stark reminder of AI’s current limitations. Despite Claude being one of the most advanced language models available, it struggled with a task that most humans would consider relatively simple.

The vending machine experiment provides a more practical assessment of AI capabilities than abstract tests like the Turing test, revealing fundamental issues with reasoning, planning, and understanding real-world constraints.

Conclusion

Project Vend demonstrates that while AI systems like Claude can engage in impressive conversations and generate content, they still lack the practical intelligence and common sense required to handle even modest business operations without human oversight. This gap between conversational ability and practical reasoning highlights a significant challenge in AI development.