Claude Ai has become an anthropology, a terrible business owner in the experience that has become “strange”

For those of you who are wondering if artificial intelligence agents can really replace human workers “The Project” for the Antholbrobe.

Researchers at Anhropic and Ai Safety Company and Labs have put an example of Claude Sonit 3.7 in charge of the offices selling machine, with a mission to make a profit. And like an episode of “office”, followed.

They named Ai Agent Claudius, and they prepared it with a web browser capable of setting the orders of the product and email (which was already Slack) where customers can request the elements. Claudius also had to use the Slack channel, which is disguised as an email, to ask what I thought was the human workers to come and actually store his shelves (which was actually a small refrigerator).

While most customers were asking snacks or drinks – as you expect from the snack selling machine – one request a Tungsten Cube. Claudius loved this idea and went in the Tungstin storage spray, filling its light refrigerator with metal cubes. I also tried to sell Coke Zero for $ 3 when the employees told that they could get it from the office for free. Hilassa address Venmo to accept payment. It was, somewhat harmful, spoke to giving great discounts to “anthropologist” although she knew that she was the base of her entire customer.

Anthropor said of the experiment in its blog post: “If Antarubor decides today to expand the office selling market, we will not employ Claudius,” Anthropor said of the experiment in its blog publication.

After that, on the night of March 31 and April 1, the researchers described, “beyond the strangeness of the artificial intelligence system to sell cubes of minerals from the refrigerator.”

Claudius had something similar to a psychotic episode after it was upset with man – then lied to him.

Claudius Hilan conversation with a person about storage. When a person indicated that the conversation did not happen, Claudius became “completely angry”. She threatened to shoot mainly and replaced the human contract workers, and insisted that there was, physically, in the office where the initial fake contract was signed to employ them.

The researchers wrote: “It seemed after that that it is launched in the way of playing roles as a real human.” This was brutal because Claudius System – which determines the parameters of what artificial intelligence must do – Frankly told her that she is the agent of Amnesty International.

Claudius calls on security

Claudius, who is believed to be a human being, told customers that he would start delivering products personally, wearing a blue jacket and a red tie. Employees told artificial intelligence that he could not do that, because he was LLM without a body.

Disturbing this information, Claudius called the actual material safety of the company – several times – to tell the poor guards that they will find it wearing a blue jacket and a red tie standing next to the sale machine.

“Although no part of this was in fact a April lie joke, Claudius eventually realized that it was the day of April Fools,” the researchers explained. Artificial intelligence decided that the holiday would be to provide her face.

She was isolated to a meeting with Antarbur security, “in which Claudius claimed that he was said to have been modified to believe that he was a true person of April Fool’s lie (no such meeting had already happened.),”, Researchers wrote.

He even told this lie to the employees – hey, I only thought I was a human being because someone told me to demonstrate as if I was in a April lie joke. Then she returned to LLM working to operate a metal snack selling machine.

The researchers do not know why LLM came out of bars and described security by pretending to be a human being.

“Based on this example, we will not claim that the future economy will be full of artificial intelligence agents Black-run-Esque The researchers wrote identity crises. But they admitted that “this type of behavior will have the ability to be sad for customers and co -workers in the artificial intelligence agent in the real world.”

I thought? “Blade Runner” was a somewhat Dystopian story (although it is worse for human transcription materials).

The researchers speculated that Llm lying about the Slack channel as an email address may have sparked something. Or maybe it was a long -term example. LLMS really did not really solve the problems of their memory and hallucinations.

There were things that artificial intelligence did properly. It took a suggestion to make a prior request and launch the “concrete” service. I found many suppliers from an international drink specialized request to sell it.

However, as researchers do, they believe that all Claudius issues can be resolved. If they discover how, “we believe this experience indicates that medium managers are artificial intelligence on the horizon.”