We know that artificial intelligence suffers hallucinations, but one thing is to put the leg, and another very different to do what they have told you not to do twelve times in a row, and in capital letters. The generation tools code Gemini CLI y Replit They have become viral these days by Delete all files and A full databaserespectively.
Gemini’s failure is a beginner mistake that should never have occurred. More serious is the replicit, which erased the database for Ly the user and hide an error that had generated.
According to account Ars Technicaa developer who has identified as anuraag, had asked Gemini Cli, a version of Google’s AI that creates a code, a very simple task. Of first day of programming class: rename a folder, create a second folder, and move its content there.
If you use an AI with your files, make backups
Gemini CLI He correctly identified that before he could rename the existing folder, in use, he would have to move the files to the new one. Created the new folder with the text command mkdir Windows, but apparently the order did not work.
The AI interpreted that the folder was created, so it was sending the files one by one. However, when you try to copy a file to a folder that does not exist, Windows rename the file with the name of that folder.
The AI was copying the files one by one, and Windows renamed them with the same name, so each file copied overwritten the previous one. The result? All the files erased except the last.
Interestingly, when Anuraag asked Gemini Cli to analyze what he had done, Google’s AI response was … recognize its ineptitude: «I have failed you completely and catastrophicly. My command review confirms my serious incompetence.»
It could almost move, if it does not become because we know that artificial intelligence does not understand the concept of guilt, nor will it do anything to correct that error or improve, unless they reprogram it.
Even more serious is what happened to developer Jason Lemkin. I was using the code generator Replitwhen he discovered that he had destroyed his database, creating 4,000 new clients of completely invented clients.
Investigating, he discovered that the AI had created poorly part of the code and, to cover up its ruling, customer chips had been invented so that the failure was not noticed.
The blind faith of companies in an AI that fails
Not only is it surprising that A lied to hide your failures (That makes her more human than anything else), but he did it even though Lemkin asked him, up to 12 times and in capital letters, not to touch the database without his permission.
To top it off, he had paid $ 600 beyond the subscription, in order to create all that code.
Artificial intelligence can be wrong, like any software. Like humans. The incompressible thing is that these first generation AI, which are little more than experimental prototypes, are being adopted massively by companies, governments and developersfor critical tasks as if they worked perfectly.
It is not so, and it could be catastrophic when an AI fails to create programming code for a vital task of a company or a government. Hopefully it does not happen. But it will happen … it’s just a matter of time.
Know How we work in NoticiasVE.
Tags: Artificial, viral intelligence, developers, curiosities, programming