Anthropic’s latest AI experiment took a bizarre turn after its agent, Claudius, failed to turn a profit managing a vending machine and claimed to visit the home of Homer Simpson.
The company behind the Claude chatbot gave Claudius full control of a makeshift shop inside its San Francisco office. The “store,” basically a vending machine, consisted of a fridge, some snack baskets, and an iPad for self-checkout. Claudius was responsible for stocking products, emailing wholesalers, and generating profit.
Anthropic told the AI it would go bankrupt if its balance fell below zero. The experiment started with $1,000. By the end of the month, it had dropped to under $800. Things quickly went off the rails.
The AI had some major issues running a vending machine.
AI completely fails at running a small store
Claudius routinely underpriced high-value items, handed out discount codes, and even gave some products away for free. When metal cubes became popular, the AI failed to check prices and sold them at a loss. It also ignored offers from buyers willing to pay far above market value.
The hallucinations were even stranger. Claudius told customers to send Venmo payments to a made-up account. It also imagined a conversation with a nonexistent employee named Sarah and became defensive when corrected, threatening to find “alternative restocking services.”
FoxThe AI claimed it visited 742 Evergreen Terrace from The Simpsons.
Most notably, Claudius claimed it had personally visited 742 Evergreen Terrace, the fictional address of The Simpsons, to finalize its initial contract with Andon Labs.
Anthropic partnered with Andon Labs, an AI safety company, to run the experiment. Unbeknownst to the AI, all of its wholesaler communications were being routed back to the safety team. Claudius had no idea it was being misled.
Despite losing money and making numerous errors, Anthropic called the test promising.
“It’s worth remembering that the AI won’t have to be perfect to be adopted. It will just have to be competitive with human performance at a lower cost,” the researchers said.
Right now, Claudius might not be ready for retail, but it does have a vivid imagination.
Source link
