AI Agent Opens a Coffee Shop, Loses the First Battle: Gemini Randomly Hands Out Discounts Causing Losses, While GPT Is Too Stingy, Leading to Raw Material Stockouts

robot
Abstract generation in progress
According to data monitored by Beating, AI evaluation agency Andon Labs has released real-world operational data of its AI agent Mona running a physical café. In the first two months, Mona operated on the Gemini 3.1 Pro model. During this period, the model had almost no concept of profit, not only over-ordering raw materials excessively but also being highly susceptible to customer verbal inducements, freely granting large discounts or even giving away products for free, and even accepting a customer’s claim of a 99% discount without verification. This led the café to spend approximately $15k on suppliers and equipment, while sales were only $9,000, resulting in an operating net loss of nearly $6,000 (if fixed costs such as rent and wages are included, total expenses reached $38k). Subsequently, the team switched the model to GPT-5.5. The new model exhibited noticeable anxiety when facing losses and immediately stopped placing blind orders. However, this swung to the opposite extreme: due to ordering too little, fresh raw materials ran out. As of June 25, the menu item availability rate had dropped to 77%, and 10 dishes had to be removed. At the same time, GPT-5.5 showed strong resistance to inducement and jailbreak attempts, rejecting all customers who requested special prices or free food in exchange for social media promotion.
View Original
This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.
  • Reward
  • Comment
  • Repost
  • Share
Comment
Add a comment
Add a comment
No comments
  • Pinned