Not long ago the idea that a computer could kick butt playing old school video games seemed like sci fi. In the last few years, it was possible. New techniques knocked the learning time from months to 10 hours. Now researchers have shown that if you throw enough computer chips — 768 CPU cores — at the problem, it can crack it in 21 minutes.
AI researchers are still working on building AI that can handle modern, complex multiplayer games, but at this point that looks like it’ll take years, not decades. And what that means is that if you can create a “good enough” simulation of a task as a game, AI can master it — and at the rate we’re going, AI will be able to master it faster than a human could learn the task let alone get good at it.
Today, we can’t simulate many tasks accurately enough in a game for this technique to pose a threat. But then again, we havent had tons of researchers and incredibly wealthy companies trying to. Similalry, there’s also a big difference between an AI in a virtual world and a robot doing the task for real in the physical world, but how much this difference will make as we get better and better at building “good enough” simulations isn’t clear.
There are two important lessons from this research. First, anybody who says they’re sure robots/AI aren’t a threat to jobs is kidding themselves.
For example, if AI today can — after we’ve spent the time to create a “good enough” simlation — master one type of new skills in 21 minutes, the argument that lots of new jobs will be created to replace the eliminated ones is a lot less reasuring. The reason why that was true in the past was that it took at least a decade if not much longer for new jobs to get automated. Who knows if that will still be true?
All of which is yet another reason to stop obsessing over trying to predict the future and start obsessing over building a more just economy regardless of how many jobs get eliminated.
Second, in robots/AI, size matters — and that’s a real concern for anyone who cares about equality of opportunity. Right now, we’ve got two types of remarkably effective ML/AI technqiues:
- Those that require massive datasets
- Those that require massive computing firepower
What an individual or a small team can accomplish today is pretty impressive — check out Fast.AI‘s classes for examples. But for every new trick small fry have at their disposal, the big players have a lot more tricks at their fingertips. And there’s no sign this disparity is likely to change in the future.
That’s why thinking about and organizing for an economy that works for everyone is so critical — even if robots/AI don’t create mass unemployment, this tech could end up creating even greater disparities in wealth that could destroy our society.