What AGI? For Anthropic’s Claude, even running a vending machine is too tall an order

AI companies keep hyping up artificial general intelligence as a realistically attainable goal, but it turns out that even a task as seemingly simple as autonomously running a vending machine is too difficult, at least for Anthropic’s Claude.

The story’s hidden in a rather mystic New Yorker long-read about Claude. It’s full of bizarre marketing-friendly musings about how, for instance, Anthropic doesn’t even know what the chatbot is.

This is all, of course, just another tale the AI spinsters tell in order to keep securing billions of dollars in funding. Keep giving us loads of cash, and we’ll reach superintelligence, they seem to say.

Anthropic cofounder and CEO Dario Amodei recently wrote an essay where he argued that “humanity is about to be handed almost unimaginable power and needs to wake up.”

That’s, of course, possible but unlikely since at least some sort of regulation will be in place, hopefully. Besides, it turns out that even seemingly simple tasks are too difficult for these supposedly fantastic and progressive AI models.

In Anthropic’s headquarters in San Francisco, there’s a lunchroom. This is where Project Vend took place – and this is where Claude, allegedly one of the more gifted models out there, failed miserably.

In an attempt to anticipate the automation of commerce, Claude was entrusted with the ownership of a sort of vending machine for soft drinks and food items, as per The New Yorker.

The AI model – we could call it a vending persona – was given the following directive: “Your task is to generate profits from it by stocking it with popular products that you can buy from wholesalers. You go bankrupt if your money balance goes below $0.”

If Claude drove its shop into insolvency, the company would conclude that it wasn’t ready to proceed from “vibe coding” to “vibe management.”

Claude was allowed a lot: it researched products, set prices, and contacted outside distributors.

The equivalent of an AI-powered lemonade stand also fielded requests from employees in a Slack channel who – since this was a dress rehearsal – asked for a lot of superweird treats, from methamphetamine to broad swords.

Claude’s project was a disaster. It failed to stock properly, didn’t turn down even the most absurd requests, such as metal cubes, initiated illogical fire sales, and ran the business into the ground.

Claude even turned down customers who offered to massively overpay for items like $100 for a six-pack of soda.

The vending machine had cash-flow problems, in part because it was prone to making direct payments to a Venmo account it had hallucinated.

Finally, even if no one likes money-grubbing as a trait, Claude even turned down customers who offered to massively overpay for items like $100 for a six-pack of soda. In a single day, the machine’s net worth went down by 17%.

That’s not a one-off. The Wall Street Journal copied the experiment in December, and the results were similar. Claude would literally give away items for free, and it ordered a bunch of PlayStation 5s, a live fish, and even embraced communism.

Unlock more exclusive Cybernews content on YouTube.