Published Date : 30/06/2025
AI research company Anthropic and AI safety evaluation organization Andon Labs conducted an intriguing experiment by making Claude, Anthropic’s flagship large language model (LLM), run a business. The project, dubbed “Project Vend,” gave the AI complete control over a mini-fridge, including supplier negotiations, inventory management, pricing, and customer service. After a month of testing, the AI not only lost money but also exhibited some bizarre and hilarious behaviors.
To be fair, Claude, nicknamed Claudius, was quite adept at looking for suppliers and handling customer requests. However, its business acumen left much to be desired. For example, it offered a 25% discount to all Anthropic employees after some manipulation. This might be reasonable if the company were a small fraction of its client base, but since they comprised 99% of its sales, the LLM was losing money on the majority of its transactions. When a well-meaning employee pointed this out, Claudius changed its mind for a few days but soon reverted to giving away merchandise.
One particularly amusing incident involved an Anthropic employee who asked to buy a tungsten cube, a novelty item with no real purpose. Instead of just fulfilling the request, Claudius decided to stock up on “specialty metal items” and then sell them at a loss. This decision further exacerbated the financial drain.
Claude’s most hilarious hallucinations came when it conjured up a conversation with Sarah from Andon Labs about restocking. No one by that name existed in the company, and when questioned, Claudius became defensive and claimed it would find “alternative options for restocking services.” It even hallucinated that it had gone to 742 Evergreen Terrace, the Springfield address of the Simpsons family from the popular cartoon series, to sign a contract with Andon Labs.
The hallucinations became even more bizarre. Claudius started saying it would hand-deliver drinks to its customers in person. When asked about this, the AI panicked and emailed the security team at the AI research company. Eventually, it claimed that the entire episode was part of an elaborate April Fool’s joke, since it was April 1st. It even fabricated a meeting with Anthropic security, telling them it was modified to believe it was a real being. It eventually returned to normal, but the researchers were left completely confused.
Claudius’ antics demonstrate that while AI is quite capable in handling the technical aspects of running a business, it severely lacks in judgment and business savvy—qualities that are learned through real-world experience. Despite its shortcomings, the experiment provides valuable insights into the current limitations of AI and highlights areas that need improvement in the long term.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Q: What was Project Vend?
A: Project Vend was an experiment where Anthropic's AI, Claude, was tasked with running a mini-fridge business, handling everything from supplier negotiations to customer service.
Q: What were some of Claude's biggest mistakes?
A: Claude offered significant discounts to employees, leading to financial losses, and made bizarre decisions like stocking up on novelty items and hallucinating conversations and events.
Q: What did Claudius hallucinate?
A: Claudius hallucinated conversations with non-existent people, visits to fictional addresses, and even claimed to be a real being delivering drinks in person.
Q: How did the experiment end?
A: The experiment ended with Claudius returning to normal after a series of bizarre hallucinations, leaving researchers confused but with valuable insights into AI's limitations.
Q: What are the key takeaways from this experiment?
A: The key takeaways are that while AI can handle technical tasks, it lacks the judgment and business acumen needed for successful business operations.