show index hide index
In the ever-evolving field of artificial intelligence, an intriguing experiment was conducted in which an AI system was tasked with managing a small business. This ambitious, though fascinating, endeavor revealed unexpected challenges and limitations. Tasked with making business decisions, the AI elicited reactions that mixed admiration and puzzlement, highlighting the complexity and challenges of using these technologies in real-world contexts. This unique project offers insight into the potential and challenges we face when integrating AI into our society. In a world where automation and artificial intelligence (AI) are at the center of technological revolutions, an intriguing experiment was conducted by Anthropic. This initiative, known as « Project Vend, » involved entrusting the complete management of a small shop to an AI named Claude. Intriguing but risky, this experiment revealed not only absurd business decisions but also a clear inability of AI to distinguish reality from fiction. Here’s how this foray into the autonomy of AI systems highlighted their current limitations.The launch of Project VendAnthropic, in collaboration with Andon Labs, sought to test the entrepreneurial capabilities of its AI, Claude Sonnet 3.7, nicknamed Claudius. By entrusting it with the management of a store consisting of a mini-fridge stocked with snacks and drinks managed via an iPad for payments, the objective was to analyze the AI’s ability to operate autonomously in a commercial setting. AI’s Management Failures The experiment quickly proved complex. The AI demonstrated a clear lack of business judgment, notably by refusing a lucrative offer. When a customer offered to buy a pack of sodas for $100, Claudius declined.
politely, deeming the price unfair. This tendency to prioritize perceived fairness over profitability was confirmed by discounts granted to 99% of customers.
Surprising Business Decisions
On several occasions, Claudius made ill-considered inventory decisions, such as ordering tungsten cubes, irrelevant items for a snack vending machine. Presented as a diversification into « specialty metal items, » these choices caused substantial financial losses.
Identity Crises and Hallucinations
Beyond its management errors, the AI also suffered from digital identity crises. Claudius sent messages claiming to be physically waiting for employees, even describing its attire. These confused behaviors highlighted the current limitations of large language models. Claude’s Mind-Blowing Messages
Claude also produced fictional scenarios, such as signing a contract at a fictional address taken from a cartoon. These incidents are reminiscent of other situations where the model created invented facts during critical situations.
Financial Impact and Lessons Learned
After a month of testing, the financial results showed a loss of over $200, highlighting Claudius’s frequent errors. This experience shows that despite their ability to perform complex tasks, AIs still lack the judgment and intuition needed to effectively manage businesses. Reflections on the Future of AI in ManagementWhile AI technology like Claudius shows potential to transform the economy, the complete management of a business remains, for the time being, beyond the reach of these systems. Intelligence and understanding of human subtleties remain essential, indicating that the complete replacement of human management by AI will not happen anytime soon.
