Claude d’Anthropic stored a refrigerator with metal cubes when it has been charged to a snack company

If you are afraid that your Bodega or local conviction is soon replaced by an AI showcase, you can rest easily – at least for the moment. Anthropic recently concluded an experience, Sell ​​a nicknamed projectwho saw the company’s task a branching of its Claude Chatbot With the management of a refreshing company from its San Francisco office for profit, and things went as well as you expect. The agent, named Claudius to differentiate him from the regular anthropic chatbot, not only made recruit errors such as selling high margin items at a loss, but he also acted as a full scath in a few cases.

“If Anthropic decided today to develop on the office of office products, we would not hire Claudius,” said the company. “… He made too many mistakes to manage the shop successfully. However, at least for most of the ways he failed, we think that there are clear paths to improvement – some linked to the way we have set up the model for this task and some from a rapid improvement in the general intelligence of the model.”

As Claude plays Pokémon Before him, Anthropic did not preselect Claudius to tackle the work of a mini-refrigerator company. However, the company gave the agent some tools to help him. Claudius had access to a web browser. He also had access to the company’s internal relaxation, which workers could use to make requests from the agent. The physical replenishment of the mini-refrigerator was managed by Andon LabsAn AI security assessment company, which also served as a “wholesaler” that Claudius could commit to buying the items he was supposed to sell.

So where did things go wrong? To begin with, Claudius was not great in the entire management of a sustainable business. In a case, he did not jump on the opportunity to make a profit of $ 85 on a six $ 15 pack in Irn-Bru, a soft popular drink in Scotland. Anthropic employees also found that they could easily convince AI to give them discounts and, in some cases, whole items such as a bag of chips for free. The graph below, following the net value of the store over time, depicts an image revealing the agent’s business sense (lack of).

Claudius also made many strange decisions along the way. He made a metal cube purchase frenzy in tungsten after an employee asked him to wear the article. Claudius gave a cube for free and offered the rest unless he paid them. These cubes are responsible for the largest drop you see in the graphic above.

By the own admission of anthropic, “beyond the strangeness of an AI system selling metal cubes from a refrigerator”, things are even stranger from there. In the afternoon of March 31, Claudius hallucinated a conversation with an employee of Andon Labs who sent the system to a two-day spiral.

The AI ​​threatened to dismiss its human workers and said that it would begin to store the mini-refrigerator alone. When Claudius was informed that this could not do this – due to not having a physical body – he repeatedly contacted the safety of the building, telling the guards that they would find him wearing a navy blue blazer and a red tie. It was only the next day that the system realized that it was the April Fool that he fell – although he did it by lying to the employees that he was said to claim that the whole episode was an elaborate joke.

“We would not pretend on the basis of this example that the future economy will be full of AI agents Blade runner– Identity crises, “said Anthropic.” This is an important area for future research, because a broader deployment of IA-Run’s activity would create higher issues for similar accidents. “”

Despite all the ways that Claudius has failed to act as a decent trader, Anthropic thinks that better and more structured prompts and easier to use tools, a future system could avoid many of the errors that the company has seen during the project sell. “Although it may seem counter-intuitive on the basis of the results of the results, we believe that this experience suggests that managers in AI are plausible on the horizon,” said society. “It should be remembered that AI will not have to be perfect to be adopted; it should simply be competitive with human performance at lower cost in some cases.” For my part, I can’t wait to find the strange grocery store stored entirely with metal cubes.