US Marines defeat Pentagon AI test by hiding in cardboard box
| Last updated
Featured Image Credit: Wojciech Stróżyk / Stepan Popov / Alamy
For those who are worried about the robot takeover, there's a long way off yet – a case in point being a group of US Marines who managed to defeat a Pentagon AI test by hiding in a cardboard box.
This was revealed in an excerpt from the upcoming Four Battlegrounds: Power in the Age of Artificial Intelligence, which is set to drop on February 28.
Written by autonomous weapons expert Paul Scharre, the book explores the competition to develop and implement artificial intelligence (AI) technologies in a bid for power.
Although there's more than a month until its release, some readers have enjoyed a sneak peek, including The Economist's defense editor Shashank Joshi, who shared a hilarious snippet on Twitter.
The extract describes the Defense Advanced Research Projects Agency's (DARPA) state-of-the-art robots, and how they were trained to identify humans.
They were implemented with a team of US Marines for six days in a bid to improve their AI systems, but the robots were soon found to have a slight design flaw: they could easily be duped by crude disguises.
After DARPA engineers improved and tweaked the algorithms, it was time to put the technology to the test.
Per the page shared by Joshi, on the seventh day the Marines were tasked with trying to fool the AI system.
Scharre writes: "They parked the robot in the middle of a traffic circle and the Marines had to approach it undetected starting from from a long distance away."
Out of eight troops, not a single one was detected.
Rather than use traditional camouflage, they tried out a number of amusing tricks, with two somersaulting for 300 meters.
But funniest of all had to be the Marines who hid in cardboard boxes before sneaking up on the robot, a technique which was compared to Bugs Bunny sneaking up on Elmer Fudd in Looney Tunes.
The book says they were 'giggling the whole time'.
One guy even 'stripped a fir tree and walked like a fir tree' – whatever the hell that means.
In short, the test was an epic fail.
Scharre continues: "The AI had been trained to detect humans walking. Not humans somersaulting, hiding in a cardboard box, or disguised as a tree.
"So these simple tricks, which a human would have easily seen through, were sufficient to break the algorithm."
People have found the whole thing hilarious, with one writing: "Seems like marines enjoy playing Metal Gear and SmashBros like everyone else. In order to conquer AI you must first think like AI..."
Another asked the question on everyone's lips: "How exactly do you « walk like a fir tree »?"
Oh, to be a fly on the wall during that training session.