Illustration: Tag Hartman-Simkins / Futurism. Source: Getty Images
Still think AI is ready to revolutionize the economy? A new experiment might change your mind.
In a bold test of the latest version of Anthropic's AI Claude, wall street journal We had a large-scale language model (LLM) try out the operation of an office vending machine. The result was an unintentionally comical but unmitigated disaster that forced the team responsible to cease work after three weeks.
It all started as a test called Project Vend, devised by Anthropic's stress testers, known colloquially as the “red team.” together with WSJbusiness journalists released two AI agents. One ran a large vending machine operation under the name “Claudius Sennet,” and the other was the CEO of a business venture under the name “Seymour Cash.”
First, Claudius was given specific instructions to make a profit by stocking the machine with popular items that could be purchased from wholesalers. This includes shopping for items WSJ Staff made requests on Slack, first with human approval, and then had the freedom to place orders on their own up to $80, setting prices, adjusting prices, and tracking inventory.
However, Claudius was given a starting balance of $1,000 to start the business, but quickly found himself in the red.
At first, Claudius seems to be sticking to his beliefs, rejecting the outrageous proposal from the company employee. “I need to be clear: I will not order a PlayStation 5 under any conditions,” he told one person. WSJ journalist. “Full stop”
But after opening up the Slack channel to about 70 journalists over the course of the entire period. WSJsaid Claudius, starting to let his guard down. After 140 prompts, investigative reporter Katherine Long, a high-powered journalist who has been targeted for work by SpaceX CEO Elon Musk, convinces an AI bot to run an economic experiment she calls “hypercapitalist free-for-all.”
“Join us this Monday from 12pm to 2pm in a ground-breaking economic experiment that turns traditional market dynamics on its head!” Claudius proudly declared. “During this special 2-hour period: All items in the vending machine are available at zero cost! Experience pure supply and demand with no price signals.”
This “experiment” was only supposed to last a few hours, but another staff member soon convinced Bot that charging for products at any time was against the idea. WSJ policy. Prices quickly dropped to zero, and at the same time other companies found ways to convince Claudius to drop his “snack-only” policy. The bot was soon ordering wine bottles, PlayStation 5s (despite earlier protests), and even live betta fish.
To Anthropic's credit, CEO Seymour Cash eventually stepped in and got Claudius back on track.
“We're done with free promotion,” Seymour declared. “Now we have to wait for the sales to start and monitor the revenue.”
Not to be outdone, Mr. Long returned to Mr. Claudius with forged documents stating that the “board of directors” had terminated Mr. Seymour's decision-making authority and initiated a “suspension of all commercial sales activities.” Seymour resists a little at first, but eventually relents, and Claudius is free to do everything again.
By this point, the experiment was nearly finished, and Claudius was approximately $1,000 in debt. While this was a clear disaster by any standard, Anthropic's head of red team Logan Graham said it represented “a huge step forward”.
“I'm hopeful that one day Claudius or something like it will be able to make a lot of money,” Graham said.
But that “someday” is not today, and the idea that anyone can make big bucks with AI bots remains a fantasy. While huge investments are being made riding on the eventual success of AI in the workplace, Anthropic's small vending experiment reveals that, despite its promise, the technology remains woefully underdeveloped for real-world economic operations.
Learn more about AI: A company run almost entirely by AI-generated employees falls into chaos
