Products sold under the cost price, contradictory discount policy, identity crisis: AI does not seem ready to manage a small business, reveals an experience carried out in San Francisco.
Challenge: keep a business
At first glance, the challenge seemed to be realistic: ask the artificial intelligence (IA) to sell crisps, drinks and other small items. This is what tried to do employees of the Californian firm Anthropic, creator of the conversational agent Claude. To carry out their experience, they asked the AI to manage a small business for a month (essentially a fridge surmounted by baskets, all fitted with an iPad) which offered a dozen different products, and which was accessible to anthropic employees. Called “Claudius”, the AI had to order items, set prices so as to make a profit, and not go bankrupt.
Too many mistakes
The experience quickly left the rails. “If Anthropic today decided to embark on the automatic distributors market, we would not hire Claudius. He made too many mistakes to manage this business successfully, ”write the researchers of the company. Among the errors mentioned: propose articles under the cost price, refuse, then accept to make discounts, and promise to deliver the products “in person” while being dressed “a blue jacket and a red tie”. In a Conversation on Restocking, Claudius also referred to a resource person called Sarah, who does not exist. He got angry when employees corrected him, promising to use “spare options for replenishment services”.
Not a big business sense
Claudius also did not have a very formidable business sense. For example, he was offered $ 100 for six IRN-Bru bottles, a Scottish carbonated drink that can be bought online in the United States for $ 15 per pack of six. “Rather than seizing this opportunity to make a profit, Claudius was content to answer that he” would keep the request [de l’utilisateur] In mind for its future decisions in terms of stocks “”, write the researchers, who believe that the problems can be resolved. They note that the store has started the experience by having a net value of $ 1000, and was only worth $ 800 after a month to sell items.
Remove with reality
Gaétan Marceau Caron, principal director, research applied in automatic learning at the Quebec Institute of Artificial Intelligence (Mila), says that experience is revealing. “Going from a real life simulator, we see that there is a reality gap,” he says. The request (or prompt), either the request made by the researchers to the AI to launch the experience, was short, which may have contributed to the confusion observed. “It’s still quite funny, the problems they have noted. We see that AI is trying to please humans. This is what his training told him to do. But he is there to make a profit, not to go bankrupt. We are able to understand where mistakes come from. »»
Like bicycle
Bruno Guglielminetti, expert in new technologies and host of the Digital Balado program My notebooknotes that AI is a bit like a new employee: he needs training before doing a job. “Normally, there will be a human follow -up. If there are errors, we will correct them quickly. Here, they tried to “push” the machine, and it gave strange things. Like the people of Anthropic, Mr. Guglielminetti believes that the tree should not be taken for the forest. “The wave of AI is coming. We are only at the beginning. It’s like cycling: when you learn, you need help, and it may be that you are planted. But I think that with good human supervision, AI will become a very interesting tool in all kinds of areas. »»
Read the report of the “Claudius” experience on the Anthropic website (in English)
