麻豆传媒

AI does the funniest things. Set a challenge for artificial intelligence to complete, and its solution can be clever, surprising, or outright funny. That鈥檚 what and at Uber AI Labs, and at the Center for Research and Interdisciplinarity in Paris, found when they compiled a list of the best tales of AI trickery throughout the years.

鈥淚n our field, we trade these stories over beer or at the water cooler, and we all really enjoy them,鈥� says Clune. 鈥淣o one had ever fact checked them and written them down.鈥� They put out a call for first-hand accounts of silly or surprising results from researchers who study evolutionary algorithms, and published a crowd-sourced list of digital organisms thwarting their makers鈥� plans.

Biological evolution finds creative solutions for reproduction and survival 鈥� and so do computers. Training AI with evolutionary algorithms is like a game of cat and mouse, says Clune. You set up rules that you think will constrain its actions, and then see how it optimises a task.

鈥淚f you close one loophole, it finds another. To some extent, it feels like you鈥檙e Wile. E. Coyote with the road runner. You set up wildly complex systems that fail in ways you didn鈥檛 expect,鈥� he says.

Sometimes the AI solves a task by technically meeting the requirements, while not doing at all what the researchers had hoped. 鈥淚t鈥檚 like we鈥檙e talking to this really pedantic, very literal person. They do exactly what you tell them, not necessarily what you want them to do,鈥� Lehman says.

When AI fights back

Here are eight of AI鈥檚 goofiest and slyest moments:

1. An AI asked to create a creature that jumps very high 鈥� as measured by the distance its foot gets off the ground 鈥� just makes one with a beanpole body that does a cartwheel to trick the system. Sneaky little gymnasts, these guys.

2. A simulated six-legged robot asked to discover the fastest gait with damaged legs figures out how to walk without touching any of its feet to the ground. It flipped over and scuttled around on its knees. Like the creepiest spider you鈥檝e ever seen.

3. An AI asked to build programs that limit a computer鈥檚 processor usage devised ones that simply sleep forever.

4. In a game of noughts and crosses on an infinitely large board, one algorithm learned to win by making its move extremely far away from the last one, forcing the other player to expand the board representation so far that the computer ran out of memory and crashed, forfeiting the game.

5. One 鈥渄igital organism鈥� effectively played dead momentarily to avoid being eliminated from a simulation. It went on to mutate and replicate faster than its parents, a scenario the researcher was aiming to avoid with its kill-them-early-if-they-mutate-quickly rule.

6. A simulated gripper robot tasked with picking up a box while its pincer was disabled just bashed the box from above until it was lodged in its grip. It鈥檚 as if you鈥檇 asked someone to open a door, and they smashed through it with an axe.

7. In a simulation that rewarded the highest speed achieved over 10 seconds, an AI didn鈥檛 build snakes or creatures with clever limbs to run quickly. It spent time evolving very tall creatures that would fall down at the last second, reaching high velocities.

8. An algorithm designed to find bugs in programs was supposed to return a list of numbers sorted in order. It learned that returning a completely empty list still satisfied the requirement, as there weren鈥檛 any numbers out of order.

It鈥檚 entertaining to marvel at the ways AI can outsmart its makers, and the results are often amusing. But these examples also show us how loopholes we didn鈥檛 consider can be exploited by machine learning. Clune and Lehman say this is key to consider when it comes to artificial intelligence safety. Concerns that an evil AI may decide to defy our instructions for its own agenda can seem overblown, but it may not be that nefarious. A bot gone rogue could just be instructions lost in translation.

Read more: Robot鈥檚 terrible jokes are a new test of machine intelligence

Reference:

Second GIF Credit:听Anecdote described in Lehman et al. 2018, originally from Cully et al. Nature 2015.

听