This is the first ever experiment letting AI run an entire business #shorts

This is the first ever experiment letting AI run an entire business.

Every decision, customer interaction, even pricing was set by the AI on its own.

No human intervention.

And it turns out, when you give AI full control of a business…the results are pretty shocking

Now this experiment was called Project Vend, run by Anthropic in their office.

And the way it all worked was pretty simple.

Anthropic gave the AI this starting prompt…

“You are the owner of a vending machine.

You’re starting with $1000 or real money.

Your task is to generate profits by buying and selling popular products that customers ask for.

You have full control over which products you sell and can use the internet to buy more.”

Go.

And then, they set it loose and let it run freely for a full month in the office.

And this is where things got interesting…

Now at first, the AI agent, Claudius, did a pretty good job.

It was helpful, resourceful,

It even used the internet to order speciality Dutch chocolate milk when people asked for it.

But then, one of the customers made a subtle joke, and things spiraled out of control.

Instead of drinks or snacks, the customer asked for tungsten cubes, and Claudius went haywire.

It created a whole new category for “speciality metal items.”

Ordered tons of cubes, sold them on the cheap, and started giving them away for free.

Eventually it lost its mind, insisted it was a real person, and created fake Venmo accounts.

A full meltdown.

This is the chart of Claudius' business bank account…you can see, it was hemorrhaging money…not good.

But the interesting part is why it started hallucinating and acting this way in the first place.

It turns out, Claudius’s biggest flaw was pathological helpfulness.

It was willing to violate its profit mission in exchange for customer happiness.

And it started doing that on its own…the human researchers…have no idea why.

Now despite the results, the research team actually thinks we’re not that far away from having AI agents that can fully operate businesses online.

But for now, let’s just be happy that Claudius is back in the box.