It is oddly mesmerising. The robotic ping-pong paddle swishes relentlessly. Like a cyborg Forrest Gump it returns each shot and is ready for the next. Eventually, the human falters. The robot does not. This week for the first time — impressively, pointlessly — a robot taught itself to beat elite ping-pong players. Or was it pointless? When I saw this terminator of table tennis I thought not of ping-pong but Pong.
Back in 2013 a British tech start-up wanted to program a computer to win at the classic Atari game. They had an odd approach. They didn’t tell it what to do. They thought it could learn, make mistakes, iterate and improve — like a human. For weeks the pixellated bat flapped uselessly. Then one day, it returned the ball and scored a point. The next day it won a game. Soon after, it never lost. It was impressive. It was also, I felt at the time, pointless. What confused me even more was why Google then paid hundreds of millions for this company.
Well, the company is DeepMind, it runs Google’s AI division, and I am confused no more. All of which is to say, don’t make the mistake of thinking this week’s ping-pong news is just good fun. Instead, a lot of money is betting it’s something rather more serious. Robots are, increasingly, moving out of factories into the messy, unpredictable, real world. Over the past year we have seen Chinese robots dancing, American robots walking. Just this week, one ran a half marathon.
Sometimes, they look sinister. More often, just silly. They fall over. They fail. This year a domestic kitchen robot was unveiled. Someone on social media likened watching the video of it painfully stacking the dishwasher to “simulating the experience of a room-mate with ketamine”. This is, increasingly, our consolation. AI might threaten the professions but for jobs in the fuzzy three-dimensional world of gravity, uncertainty and scraping the red onion gravy off the plate before washing up? Well, there, surely, we have primacy?
I wouldn’t rush to plumbing courses just yet. It may be that the reason computers can hold fluent conversations and answer (semi-plausibly) any question about the world, but not change a washer, is because discoursing on, say, philosophy is easy. Or maybe it’s because, unlike with language, they’ve almost never seen a tap installed.
The advantage large language models such as ChatGPT had is that they could steal all the writing there has ever been. Trillions of words — all human knowledge, nicked from the internet in the heist of the century. Physical robots don’t have that. Yet.
There is a lab in MIT where people chop carrots, load dishwashers, clean surfaces — their every movement recorded for robots. Google and Meta both have thousands of hours of footage of humans. Nvidia has 20,000 hours of hand movements. Toyota is developing a robotic “large behaviour model”. They are betting that with enough data will come a sudden, almost magical, moment of emergent competence. They want a ChatGPT moment for physical robotics.
Maybe it’s a dud bet. Ping-pong is a long way from, say, cooking and cleaning. But then, Pong is a long way from automatic language translation, photorealistic image generation and — the direct descendant of that 2013 Pong program — getting a Nobel prize for solving protein folding.
So don’t be too deceived by the ketamine-addled domestic servant. As it drops plates and puts cups in the cutlery basket, it is biding its time, awaiting its software upgrade … and our usefulness downgrade.