Many people are surprised that Google hasn’t yet launched its own AI that competes with ChatGPT; Although there is already one in development, called Sparrow, its creators have justified the delay as a “precaution” necessary for these projects.
[Intentan crear un videojuego con Inteligencia Artificial, y el resultado da que pensar]
And it is that, although Artificial Intelligence can offer incredible results, it is no less true that when it fails, it is also incredible but in another sense. When we create AI images, for example, it’s easy to find distorted faces or hands with too many fingers, failures that a human author would never miss.
Fool a military AI
This is something that happens with all AIs. Basically, an AI knows how to mimic human behavior, but it doesn’t know why. You know what you need to do, but you don’t understand the motivation or the end goal; This is why they are good at performing specific tasks, but very bad when parameters change and they are forced to adapt.
We saw the funniest example of this limitation this week, thanks to Shashank Joshi from The Economistwho published a twitter A page from military expert Paul Scharre’s new book detailing a military exercise using artificial intelligence.
The objective of the maneuvers was to use the experience of a group of American Marines to improve the detection of intruders by a robot equipped with cameras and Artificial Intelligence. At first, it was a bit of a boring exercise, since the soldiers only had to walk near the robot for it to “learn” what it needed to detect.
These aren’t the marines you’re looking for. pic.twitter.com/qMwzJblQwp
—Shashank Joshi (@shashankj) January 18, 2023
When they were done, the robot’s cameras were able to detect the presence of soldiers and satisfied with the result, the engineers decided to change the “game”: now the marines had to beat the AI. The challenge was simple: if one of the eight soldiers could approach the robot and touch it without being detected, he would win.
The creators of the AI seemed certain of their victory, but their defeat was embarrassing: none of the eight soldiers were detectedand the best part is that they quickly understood the weakness of the robot: that it was looking for people, and only people.
The tactics they used to approach the robot seem more typical of cartoons than professional soldiers. One managed to approach the robot by somersaulting the distance of 300 meters which separated them. Another took branches from a tree and tied them to his body, and was able to approach the robot as if it were a walking tree.
Perhaps the most ridiculous was the case of two Marines who simply put boxes on each other; although they laughed all the way, the robot did not detect them. It’s the exact same tactic Solid Snake used in the Metal Gear Solid game, and who knows if that was the inspiration the soldiers got.
In this way, the professionals revealed the big problem of the AI: that it was looking for human figures. Any human guard would have caught them instantly, as he understands that they are still intruders, even if they behave strangely; but an AI didn’t understand that, and since it didn’t fit with the idea he had learned of “intruder”, he let them pass.
You may be interested
Follow the topics that interest you
