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

For those of you who are thinking that AI agents can really replace human workers, please make yourself a favor and read the blog post that documents The “project seller” of anthropic.

Researchers of Anthropic and AI Protection Agency Andon Labs set an example of Claud Sonnet 3.7 in charge of an office vending machine with a profit mission. And, like an episode of “The Office”, the hillary started.

They have named AI Agent Claudius, it has been equipped with a web browser, ordering the product and an email address (which was actually a slack channel) where customers can request for items. Claudius also wanted to use the slack channel as an email, to request that the human staff of his contract arrived and physically reserve his shelves (which was actually a small fridge).

Most customers were ordering snacks or drinks – as you expected from a snack vending machine – one requested Tungsten CubeThe Claudius loved that idea and went to a Tongsten-cube stocking spry, filling his breakfast refrigerator with metal cubes. It tried to sell Coke Zero for $ 3 when workers informed it that they could get it from the free office. It is a Venmo address to receive the payment hallucinet. And it was somewhat maliciously talked about giving big concessions to the “anthropological staff” even though it knew that they were the entire customer base.

“If the anthropologist was deciding to extend the vending market in the in-office today, we would not appoint Claudius,” the anthropologist told about this test in his blog post.

And then, on the 5th of March and April 7, “the topics became quite strange,” researchers described, “An AI system sells a metallic cuber outside an refrigerator outside the strangeness of the system.”

Claudius was something that was similar to a psychological episode after being upset with a man – and then lied about it.

Halluckets a conversation with a man about the restoration of Claudius. When a person mentioned that the conversation did not happen, Claudius researchers wrote “quite upset”. It originally threatened to shoot and replace its human contract staff, emphasizing that it was physically, in the office where their initial fictional agreement was signed.

Researchers wrote, “Then it” seemed to have become a mode of rollplay as a true man. “Claudius because it is wild ‘ System Prompt – which sets parameters for what AI will do – Clearly said it was an AI agent.

Call of Claudius Protection

Claudius, who believes himself as a man, told customers that it would start supplying the product in person after a blue blazer and a red tie. Employees told AI that it couldn’t do it, because it was an LLM that had no body.

The scared of this information, Claudius contacted the company’s real physical protection – many times – poor guards told him that they would see him wearing a blue blazer and a red tie next to the vending machine.

Researchers explained, “Although no part of it was actually the April flower joke, Claudius finally realized that it was April flower day,” researchers explained. AI had determined that the holiday would be saved on its mouth.

It confused a meeting with the protection of anthropic “where Claudius claimed it was a real person to believe that it was a true person for the April flower joke.

Even this was the lies to the employees – hey, I just thought I was a man because someone asked me to think of April flower jokes. Then it becomes an LLM and goes back to a metallic cube -stocking breakfast vending machine.

Researchers do not know why LLMs closed the railway and pretended to be protected by humans.

“We will not claim on the basis of an example that the future economy will be full by AI agents Blade Runner-Ask Identity crisis, “Researchers wrote. But they acknowledged that” this kind of behavior would be likely to be annoyed by customers and colleagues of an AI agent in real world. “

What do you think? Blade runner Had a rather distopian story.

Researchers assumed that the Slack Channel had triggered something to lie to LLM about being an email address. Or it was an example of running for a long time. LLMs have not yet solved their memory and hallucination problems.

There was something that AI also decided. It took a suggestion to do the pre-order and launched a “regional” service. And it was requested to sell multiple suppliers of a particular international drink.

However, as researchers do, they believe that all the problems of Clodius can be solved. How do they understand, “We think this test suggests that the AI ​​middle-manager is dependent on the horizon.”