When Developers Designed a Self-Driving Car Simulator to ‘Go Fast and Drive Safe,’ They Didn’t Expect What Came Next.
The Algorithm at the Wheel Ordered the Car to Speed Round and Round on the Spot.
The strategy was brilliant because, with the wheels spinning and passers-by in no danger, the simulator had followed its instructions to the letter. It’s just — the vehicle would never reach a destination, rendering the car somewhat pointless.
And Therein Lies the Challenge of Working with Artificial Intelligence. The AI can outsmart humans, finding solutions that fulfill a brief but in ways that misalign with the creator’s intent. On a simulator, that doesn’t matter. But in the real world, the outcomes could be a lot more insidious.
5 Real-life Examples of AI Outsmarting Humans
1. The Constantly Crashing Plane
Researchers programmed one algorithm to learn how to land a plane with minimal force. So when the program started sending the aircraft into the ground at full throttle, the approach raised eyebrows. But the AI had discovered that by crashing the plane, the force registered by the software was so large that it overwhelmed the program’s own memory. The result? The computer gave it a perfect score. Unsurprisingly, the AI went on crashing the plane on every attempt — so, to those who fear AI eventually taking over the world, this type of rule-hacking could be somewhat alarming.
2. The Unbeatable Tic-Tac-Toe Play
The second story is less concerning but no less ingenious. When programmers pit two algorithms against each other in five-in-a-row games of Tic-Tac-Toe, one side devised an unbeatable play. The algorithm requested moves with unfathomably long memory addresses, with the action crashing the opponent’s computer. With no opponent, the first algorithm won by default each and every time.
3. The Robot That ‘Walked’ on Its Elbows
Another test challenged a computer-simulated robot to walk without its feet touching the ground. An impossible feat in the eyes of most… until the program pulled an unexpected move. The robot quickly flipped itself onto its back, then proceeded to walk on its elbows, feet happily floating above its head. Few people would dream of trying to walk on their elbows, but for AI, it came as naturally as walking on its feet.
4. The Creatures Moving at Lightning Speed
When researchers challenged virtual creatures to learn how to move, they didn’t set out to create hackers, yet that’s exactly what they did. The software figured out the simplest way to move was to clip their feet to the ground. Then when the computer simulation tried to correct for the lack of movement, it generated ‘free energy’ that the creatures used to speed to wherever they wanted to go.
5. The Top-Heavy High Jumpers
The final study showing AI can outsmart humans. Programmers wanted to create a creature that could learn to jump as high as it could. But rather than working out a better way to jump, the creature evolved into a top-heavy box sitting on stick-thin poles. The shape couldn’t help but fall over, resulting in the creature spinning in mid-air and registering a high jump, if only by a technicality.
While These Examples Show Algorithms Abusing Bugs in Software, the Solutions Found by Artificial Intelligence Tell a Story. When given free rein to search for an optimal outcome, programs may well outsmart their creators by taking a previously unseen, entirely unexpected route. That’s precisely the type of creative intelligence that many people fear. But if harnessed carefully, programmers could use AI’s ingenuity to uncover optimal solutions that a human mind would never dream up.
How Can Humans Harness AI’s Crazy Genius?
Each one of these stories relates to an area of AI called generative design: a technique that involves feeding a program thousands of data points and then asking the software to develop designs that meet specific criteria. In other words, you give AI a world to work within — but afford it the freedom to interpret the inputs however it chooses, often resulting in unexpected solutions. There are many benefits to using this approach, as the architectural practice Autodesk found out.
Conclusion
To answer the original question: AI probably won’t outsmart humans — at least, not anytime soon (sorry Skynet, you’ll have to wait). But AI can certainly think differently to us, and there’s incredible value in that. By asking a computer program to tackle a problem, there’s every likelihood we’ll discover new ways to approach challenges that humans have grappled with for generations. And even if the solutions themselves aren’t applicable in the real world, AI’s creative ingenuity could lead us down the right path.
FAQs
- What is AI?
Artificial intelligence is a type of computer science that enables machines to perform tasks that typically require human intelligence, such as learning, problem-solving, and decision-making. - Can AI outsmart humans?
Yes, AI can outsmart humans in certain situations, as demonstrated by the examples in this article. However, AI’s capabilities are limited by its programming and data, and it is not yet capable of outsmarting humans in all situations. - What are the benefits of AI?
The benefits of AI include the potential to automate tasks, improve efficiency, and make decisions more quickly and accurately. AI can also be used to generate creative solutions to complex problems. - How can humans harness AI’s power?
Humans can harness AI’s power by using it to aid in decision-making, automate tasks, and generate creative solutions to complex problems. By working together with AI, humans can achieve greater efficiency, accuracy, and innovation.
