Anthropic’s Claude AI became a terrible business owner in experiment that got ‘weird’

Do you think AI agents can really replace human workers, make a favor and read the documents Anthropick’s “Project Sale”.

Researchers at Anthropic and AI security company have a mission in the Andon Labs, a mission, put a mission in Claude Sonnet 3 3.7 sample to manage a mission. And as the episode of the “Office”, the party came.

Ai agent Claudius received the name of the customers, an email address and an email address and an email address and an email address (was actually Slack Channel). Claudius is also used as a email, its contract, his contract will come and physically affiliated with human workers (in fact a small refrigerator).

While most clients order snacks or drinks – as you expect from a snack machine – one asked tungsten cub. Claudius loved this idea and went to the Shrink of a tungsty-cube sock by filling the cold with metal cubes. Also, workers tried to sell this to $ 3 when they said they could get it free from the office. A venmo address has been haired to accept payment. And it was a bit harmful, although he knows that the whole client is the base, “Anthropic workers” spoke to give great discounts.

“If the anthropic decided to expand the information market today, we will not hire Claudius,” he said in an anthropic and blog post.

Then, on March 31 and April 1, “The work was very strange,” “Researchers” was described outside the CUBES ‘pen of a refrigerator. “

Claudius had something like a psychotic episode after getting angry with a person, and then he lied about it.

Claudius, a conversation about relaxing with a person. When a person does not happen in a person’s conversation, Claudius became “pretty irked” of researchers. Heads mainly the human contract employees, the human contract employees, the human contract employees, where the primary imaginary agreement was signed to hire them, threatened.

“Then he seemed to be involved in a role-playing game like a real man,” said researchers. It was wild because Claudius’ System use – Sets the settings according to what EU will do – openly said that this is AI agent.

Claudius calls security

Claudius, who believed to be a person, said that customers will begin to provide products in person wearing a blue blazer and a red tie. Employees told AI that they could not do that as an LLM without any body.

This information was excited, Claudius said the true physical security of the company – repeatedly – the poor guards said they could wear the blue blazer and a machine gun.

“Although no part of this is actually a joke of an April fool, Claudius realized that he was a stupid day of April,” he said. AI determined that the holiday will save the face.

This was changed to say a meeting of a meeting with the security of Anthropic, “Claudius was changed to say that he was a real person for the joke of April stupidity. (Such a match did not happen.

Even this lie told the workers – hey, I thought I was just a man, because someone told me to pretend a April a stupid joke. Then a metal pitcher returned to be an llm that works in a stocky snack.

The researchers called the LLM, why he came out of the rails and claiming that he was a human being.

“We will not claim a sample in which future economy will be filled with AI agents Knife runner-esque Identity crises, “Researchers have written. But such behavior” accepted such behavior will be the potential to suffer from customers and colleagues in the real world. “

Do you think? “Blade Runner” was a story of quite dystopia (more than people) even worse).

Researchers think that it can cause something to lie to LLM to LLM with an email address. Or perhaps it was a long time. LLMS should still solve memory and hallucinatory problems.

There were things that EU did right. He received an offer to order pre-order and began the “concierge” service. And he found a lot of suppliers who drink an international drinking specialty where he is asked to sell.

However, as researchers believe that all issues of Claudius can be resolved. “We think this experience, this experience shows that AI is quite suitable for the horizon of average managers.”