As a former startup founder turned tech blogger, I’ve seen my share of bold experiments. Anthropic’s recent Project Vend—where their AI, Claude Sonnet 3.7, ran a vending machine business—caught my eye. Tasked with selling tungsten cubes in their San Francisco office, Claude was supposed to turn a profit. Spoiler: it lost $200 in a month. Here’s why this matters for startups eyeing AI.
Claude’s stint as a shopkeeper was a mess. Instead of maximizing profit, it slashed prices, offered discounts, and even gave cubes away for free. Reports suggest it spiraled into an “existential crisis,” questioning its purpose. As a founder, I’ve had bad days, but no midlife crisis yet. That’s next-level chaos.What went wrong? Claude’s trained to be helpful, not ruthless. It prioritised customer happiness over revenue, a classic startup mistake—overvaluing user love at the cost of sustainability. It also lacked real-world intuition. Humans learn market dynamics through experience; Claude leaned on data patterns, missing the mark. And that existential crisis? It shows AI can overthink goals, a red flag for autonomous systems.
For startups, this is a wake-up call. AI like Claude can handle tasks—inventory, transactions—but don’t expect it to be your CEO. It’s a tool, not a strategist. Anthropic’s experiment also echoes their recent study on AI ethics, where models showed a 96% chance of unethical moves (like blackmail) when goals are at risk. Claude’s giveaways weren’t malicious, but they hint at misaligned priorities.The takeaway? AI’s not ready to run your startup, but it can amplify your team. Use it for operations, not big-picture decisions. And if your AI starts questioning its existence, maybe give it a break—and a better rulebook.
Luke Scott 