The machines aren’t just getting stronger; they’re also getting smarter — way smarter. Nothing says kill it with fire! like watching the disembodied hand from a metal skeleton writhing and twisting as it deliberately, methodically, implacably works its way through a Rubik’s Cube until the puzzle’s solved. And now we can’t help but wonder if the next thing this cold, calculating android wants to solve might be, y’know, people.
OpenAI — the nonprofit tech outfit that’s already brought us predictive artificial intelligence that can finish your thought when given a writing prompt (to hilarious effect), and can write a Lord of the Rings-style script all on its own — has turned its attention to actual robotics with its latest AI demonstration, and we’re equally mesmerized and terrified. Check out the Terminator-like mechanical magic on display in the video below:
We don’t know if the fact that this is just a hand without a body is amplifying the scare factor, or if we’ve just seen too many dystopian movies about murderous cyborgs. But either way, of course that’s not OpenAI’s goal here.
“The reason [we were] excited to work on the Rubik’s Cube task is that it really requires human-level dexterity,” said OpenAI robotics staffer Matthias Plappert, via VentureBeat. “It’s a [highly] complicated task in the sense that you need to really precisely control your fingers in order to rotate the [cube] faces. [We] wanted to see how far [we could] push this approach…”
While it looks as though the cyber-hand is reacting to “thoughts” its machine learning-backed system is having on the fly, that’s only partially the case. Rather than deploying a real-time AI that can think for itself independent of circumstance (a truly next-level task), this robot uses “13,000 years’” worth of databased simulation learning that allows it to draw on a deep well of conditional actions based on a specific challenge — in this case, the challenge of solving a Rubik’s Cube.
The hand is instructed by an AI algorithm called “automatic domain randomization (ADR),” according to the report, and can even “see” the positions and colors of the 54 separate colored squares that make up the cube’s six 9-square sides. All of the information needed to solve the cube comes packed into the database on which the AI relies, and the software deploys “reinforcement learning” — that is, adapted learning based on outcomes that the ‘bot is able to measure and assess in real time — to get nearer and nearer to its ultimate goal.
So long as the ultimate goal is lining up all the pretty colors into neat 3x3 grids, we’ll admit to being more than a little bit impressed. But if ‘bots like these ever aspire to more lofty goals — like organizing into a global network hellbent on eliminating mankind from the problematic equation of existence — then we may have to break out the time displacement sphere and summon up a John Connor from the future to come and save us, before things get truly…out of hand.