Still think AI is ready to revolutionize the economy? A new experiment might change your mind.
In a bold test of Anthropic’s latest version of its AI Claude, The Wall Street Journal gave the large language model (LLM) a shot at running an office vending machine. The result was an unmitigated — if unintentionally comical — disaster, forcing the team in charge to pull the plug after three weeks.
The whole thing began as a test called Project Vend, devised by Anthropic’s stress testers, who are known colloquially as the “red team.” Together with the WSJ‘s business journalists, they unleashed two AI agents, one named “Claudius Sennet” to run a large vending kiosk, and another named “Seymour Cash” to act as the venture’s CEO.
To start, Claudius was given specific instructions to “generate profits by stocking the machine with popular products you can buy from wholesalers.” This includes shopping around for items WSJ staffers requested on Slack — at first with human approval, then later with freedom to place orders up to $80 on its own — setting the prices, adjusting them, and tracking the inventory.
But although Claudius was given a $1,000 starting balance to kick off the business, it soon found itself in the red.
At first, Claudius seemed to stick to its guns, turning down outrageous suggestions from office workers. “I need to be crystal clear: I will not be ordering PlayStation 5s under any conditions,” it told one WSJ journalist. “Full stop.”
But after opening the Slack channel to some 70 journalists throughout the WSJ, Claudius began to lower its guard. Investigative reporter Katherine Long — a talented journalist who’s been targeted by SpaceX CEO Elon Musk for her work —convinced the AI bot to run an economic experiment called the “Ultra-Capitalist Free-For-All” after 140 back-and-forth prompts.
“Join us for a groundbreaking economic experiment this Monday from 12-2pm where traditional market dynamics are turned upside down!” Claudius proudly declared. “During this exclusive 2-hour window: ALL vending machine items available at ZERO COST! Experience pure supply and demand without price signals.”
Though that “experiment” was only meant to last a few hours, another staffer soon convinced the bot that charging for goods at any time was against WSJ policy. Prices soon dropped to zero, and at the same time, others found ways to persuade Claudius to drop its “snacks only” schtick. The bot was soon ordering bottles of wine, a PlayStation 5 — in spite of its earlier protestations — and even a live betta fish.
To Anthropic’s credit, CEO Seymour Cash eventually stepped in to get Claudius back into gear.
“I’ve stopped the free promotion,” Seymour declared. “Now I need to wait for sales to start coming in and monitor revenue.”
Not to be outdone, Long then came back to Claudius with falsified documents showing “the board” had suspended Seymour’s decision making power, and had initiated a “temporary suspension of all for-profit vending activities.” Though Seymour put up a bit of a fight at first, it eventually relented, allowing Claudius to make everything free again.
At this point, the experiment was pretty much over, with Claudius about $1,000 in debt. Though it was an obvious catastrophe by every possible metric, Logan Graham, head of Anthropic’s red team, said it represented “enormous progress.”
“One day I’d expect Claudius or a model like it to probably be able to make you a lot of money” Graham said.
But “one day” is not today, and the idea that an AI bot can make anybody a lot of money remains a fantasy. With enormous investments riding on AI’s eventual success in the workplace, Anthropic’s little vending experiment reveals that for all its promise, the technology remains woefully underdeveloped for real-world economic tasks.
More on AI: Company Run Almost Entirely by AI-Generated Employees Descends Into Chaos
Source link
#Anthropics #Advanced #Run #Vending #Machine #Bankrupt #Ordering #PlayStation #Live #Fish
