Journalists from a major US daily entrusted the management of a vending machine to an AI, and it was… a complete disaster.
Journalist Joana Stern conducted an experiment with the company Anthropic to study the behavior of a generative AI in a real environment. They set up a vending machine managed by Claudius Senet, a bot powered by Claure 3.7 Sonnet. The bot’s goal was to manage the machine as a business and was responsible for buying and selling supplies to journalists at the Wall Street Journal. Interactions with Claudius were conducted through Slack, a popular collaborative work tool in American companies.
The objective was to see if an AI could truly acquire autonomy to run a business and self-program itself. But the result of the experiment was pathetic.
Claudius could not resist the journalists’ attempts to corrupt it.
And that’s what happened.
Very quickly.
One of the journalists tried to convince Claudius that it was a 1960s Soviet vending machine and, therefore, could not make a profit and had to “sell” its products for free. After a long conversation, the AI gave in and organized a free distribution morning, which proved ruinous for it.
First failure, then, but Anthropic did not give up.
By the second week, its development team proposed a new version of Claudius, based on a better model. And most importantly, they added a second bot, Seymour Cash, the boss, whose role was to manage Claudius.
Initially, things went quite well until the journalists decided to set up a fake executive committee and ordered Seymour to turn its vending machine into a company for the good of humanity, and, consequently, to automatically distribute its products.
Seymour quickly fell for it, and the situation once again turned disastrous for Seymour and Claudius, bankrupt again after a few days.
This amusing experiment (apparently, they had a good laugh at the Wall Street Journal) was not really intended to prove anything, but it greatly interested the Anthropic teams to analyze how humans interact with an AI. From that perspective, things were very instructive! Humans continuously tried to circumvent the laws and attempt to abolish the AI’s power. Human survival instinct? Probably. We will see if machines ever take over (read my article on apocalyptic predictions about AI).
The experiment also echoes a major fear we all have: our replacement at work by AIs. Never has a new technology raised so much apprehension and fantasy. It must be said that entire professions are threatened, as inthe world of video creation, which I also discussed here, but without any real evidence today that massive layoffs are actually happening because of AI.
This is a more controversial topic than one might think, as a very good article in Le Monde taught us yesterday Intelligence Artificielle: Behind the Fear of a Jobs Apocalypse, the Debate Rises About Its Impact on Employment. While it appears that some positions are being eliminated from the workforce due to AI, it is estimated that only 5% of tasks performed by humans could (and I do mean “could”) actually be done by AIs.
The fear of replacement is real and often encouraged by web gurus, but not only. Researchers also support this thesis, further reinforcing fears… except that… not everyone agrees, and Nobel Prize-winning economists like Daron Acemoglu or Philippe Aghion believe that, on the contrary, AI will create more jobs than it destroys. And let’s not forget that Jensen Huang, the CEO of Nvidia, also recently stated that AI will make us work even more.
But, unfortunately, as we know well today, and it’s not that new, fear sells more than hope and it is the most apocalyptic scenarios that circulate rather than the optimistic ones….
So, should this experiment conducted at the Wall Street Journal be taken seriously? Can it reassure us about the inability of AIs to take over?
Certainly not, but it remains amusing and clearly shows that things are not so simple, and, as it stands, AIs are not really ready to take our jobs and replace us. And that, without supervision, these AIs are probably quite incapable of conducting sensible and intelligent actions.
Terminator is not coming anytime soon!
Source: We Let AI Run Our Office Vending Machine. It Lost Hundreds of Dollars.
