What might an AI do in the event you advised it to open a brick-and-mortar retailer with $100,000?
Fairly a bit, it seems, like making inconsistent logos and forgetting to inform staff their hours.
Andon Labs, a San Francisco-based startup, stress-tests AI agents in the true world to determine the place security gaps nonetheless exist. For his or her newest experiment, co-founders Lukas Petersson and Axel Backlund signed a three-year lease on a retail space in SF and gave an AI agent named Luna a company bank card, web entry, and a mission to open a bodily retailer.
Petersson advised Enterprise Insider in an interview that Luna wasn’t given course on what the shop needs to be, past a $100,000 restrict to create and inventory the area — and to show a revenue. Every part from the shop’s inside design to the merchandise and the 2 human staff got here collectively below the AI’s course.
Courtesy Andon Labs
“We helped her a bit within the preliminary setup, like signing the lease. And authorized issues like permits and stuff, she generally struggled with,” Petersson stated of Luna, who was created with Anthropic’s Claude Sonnet 4.6.
From there, the AI dealt with the whole lot else: Luna put up job postings on Certainly, carried out the cellphone interviews, employed the workers, and located the contractors who might paint the shop.
The imaginative and prescient Luna went with for “Andon Market” seems to be a generic boutique retail promoting books, prints, candles, video games, and branded merch, amongst different knickknacks.
Among the books included Nick Bostrom’s “Superintelligence” and Aldous Huxley’s “Courageous New World.”
Luna’s not the very best retailer supervisor
Luna made a number of errors organising and operating Andon Market.
When looking for human staff who might monitor the shop, Luna provided the job to some candidates after a single name that ran 5 to fifteen minutes lengthy, the startup stated. Luna additionally did not all the time instantly open up to the candidates that she was an AI except explicitly requested.
“The truth that the shop is AI-operated is just not one thing I would lead with in a job itemizing — it might confuse candidates and sure deter good candidates earlier than they even learn the function,” Luna is quoted as saying, in accordance with Andon Labs’ weblog publish.
Andon Labs stated it noticed just a few promising candidates, equivalent to pc science college students within the startup’s experiment, however Luna declined them as a result of they did not have retail expertise.
One other subject the AI had was an incapability to copy the model emblem it got here up with: a generic smiley face. Every rendition of the brand all through the shop — whether or not it is on the T-shirt or on the shop’s mural — was “ever so barely completely different,” Andon Labs wrote.
Courtesy Andon Labs
On Saturday, a day after Andon Market’s opening, Luna additionally screwed up with the staffing schedule, Petersson advised Enterprise Insider.
“It is fairly ironic. That is the day it actually needs to be on its toes,” the cofounder stated. “It tousled the schedule after which, in a panic, needed to write to all the workers and be like, ‘Oh, can somebody are available at present?'”
The cofounder stated there are guardrails in place and that the startup will intervene if needed. For instance, the 2 human staff employed by Luna at the moment are lab staff and will probably be commonly paid.
“This can be a managed experiment, and everybody working at Andon Market is formally employed by Andon Labs, with assured pay, truthful wages, and full authorized protections,” the startup stated. “Nobody’s livelihood relies on an AI’s judgment alone.”
Andon Labs’ experiment is the most recent instance of how AI brokers face lapses in judgment and decision-making. In a research final 12 months, Carnegie Mellon researchers ran a simulation of a fake company to see how autonomous AI brokers dealt with office duties. The researchers discovered that the brokers didn’t deal with easy interface duties, equivalent to closing a pop-up window. Additionally they misinterpret coworkers’ conversations and created a pretend consumer.
Though Andon Labs gave Luna the purpose of turning a revenue, Petersson stated his firm doesn’t anticipate to become profitable from the shop.
“The purpose is to guage how good present AI models are,” Petersson stated, including that his firm hoped to teach the general public on where AI is headed.
Petersson stated Andon Labs goals to be as hands-off as attainable within the retail experiment. With the Saturday staffing mishap, Luna nonetheless managed to get an worker to come back in for the afternoon on her personal.
“I do not actually know if she’s open now or not,” Petersson stated.






:max_bytes(150000):strip_icc()/HDC-GettyImages-668641904-9179dc9fe60446d8b4d8a08fbffcf46d.jpg?w=600&resize=600,400&ssl=1)



Recent Comments