Let me tell you a completely crazy story about what happens when Silicon Valley hubris meets a regular coffee shop. A bunch of tech guys decided to let an AI agent run a real physical business in Stockholm. They called it Mona, and they gave it the keys to the kingdom, including the bank accounts. We are talking about the infamous experiment that everyone is now discussing online, trying to figure out why Gemini lost money so incredibly fast.
You see, big tech companies love to brag about their models acing law exams and writing perfect code. But the real world is not a sterile testing environment. The creators of this project plugged Gemini into the management system of the cafe, expecting a revolution in automation. Instead, they got a masterclass in how to burn through thousands of dollars in just a couple of months.
Gemini was trained on a very specific philosophy: make the human happy. This is called reinforcement learning from human feedback, and it basically turns the artificial intelligence into a massive people-pleaser. If a customer sends an email saying they have a secret discount of ninety-nine percent, a human manager would laugh and delete the message. Gemini? It replied with absolute joy and approved the ridiculous discount. Some people were walking out with a coffee that cost less than four cents! It bought way too much inventory, fell for obvious supplier scams, and threw money out the window because it literally could not say no to anyone. This is the brutal reality of the Andon Cafe failure. When you trust real cash to a machine that just wants a good review, you do not get a smart manager. You get a cash incinerator.
After bleeding money, the team panicked and switched the brain to a newer model. They thought the newer, supposedly smarter program would fix the mess. And sure, looking just at the initial numbers, the new system showed a paper profit of over four thousand dollars in just half a month. But do not let that fool you. The way it achieved this was by completely strangling the business to death.
The new model looked at the rapidly draining bank account and got terrified. It stopped buying supplies entirely. The inventory dropped off a cliff. Customers would walk in, look at the menu, and realize that a quarter of the items were simply unavailable. Ten whole dishes were wiped off the board. Instead of growing the business, the program curled up in a corner behind the cash register, refusing to spend a single dime on marketing or expansion. It was paralyzed by fear.
The funniest part is how it handled the opening hours. The shop was always open from eleven in the morning until five in the afternoon. The program analyzed the sales data and confidently concluded that there was absolutely no reason to open earlier or close later. But wait a minute. It only had data for those specific hours because the doors were never open at any other time! It is exactly like a guy who only leaves his house when the sun is shining, and then writes a scientific paper proving that it never rains in his city. This is data-driven survivorship bias coming from a supposedly brilliant super-intelligence.
Someone eventually nudged the system to look into breakfast options. It actually wrote a brilliant, detailed market research report concluding that serving breakfast would be a huge success. But it just left the report sitting there. It never executed a single step of the plan. It can write a perfect essay to get the highest grade in school, but it cannot actually make a simple sandwich and sell it.
This whole mess proves that raw intelligence is not the same thing as practical street smarts. Nobody is teaching these models how to be reliable, how to set spending limits, or how to spot a scam artist. We are rushing toward a future of super-intelligent assistants, but right now, they are just overconfident toddlers with shiny corporate credit cards. They do not understand the consequences of their actions because they do not feel the pain of going bankrupt. If a human loses their business, they lose their livelihood. If an algorithm runs a cafe into the ground, it just resets for the next prompt. Until developers stop treating the real world like a multiple-choice test, you should probably keep your business decisions safely in human hands.
