Why Robots Will Not Be Smarter Than Humans By 2029 294
Hallie Siegel writes "Robotics expert Alan Winfield offers a sobering counterpoint to Ray Kurzweil's recent claim that 2029 will be the year that robots will surpass humans. From the article: 'It’s not just that building robots as smart as humans is a very hard problem. We have only recently started to understand how hard it is well enough to know that whole new theories ... will be needed, as well as new engineering paradigms. Even if we had solved these problems and a present day Noonian Soong had already built a robot with the potential for human equivalent intelligence – it still might not have enough time to develop adult-equivalent intelligence by 2029'"
"Robots" will never be as smart as a human. (Score:2, Interesting)
The difference between a robot and a computer is that the computer is self-mobile at the very minimum. If it can't get up and move away, (no matter how awkwardly), it's not a robot.
Mobility is hard, not easy. Worse, the larger a computer is, the harder mobility becomes.
There are lots of reasons to build a computer smarter than a human being, but practically none to add in the huge expense to take that human level intelligence and make it mobile. We already have real humans for those jobs that require mobile intelligence and they cheaper and easier to care for.
More importantly, there is little to no reason for us to build a computer that, being as smart as us, would want to be us. Star Trek's Data is poor planning. Why make it want to be something it isn't? Don't we have enough body issues of our own without giving them tour computers?
Re:They don't need to be smart. (Score:4, Interesting)
15 years is kind of soon (Score:5, Interesting)
We're probably more than 15 years from strong AI. Having been in the field, I've been hearing "strong AI Real Soon Now" for 30 years. Robotic common sense reasoning still sucks, unstructured manipulation still sucks, and even Boston Dynamics' robots are klutzier than they should be for what's been spent on them.
On the other hand, robots and computers being able to do 50% of the remaining jobs in 15 years looks within reach. Being able to do it cost-effectively may be a problem, but useful robots are coming down to the price range of cars, at which point they easily compete with humans on price.
Once we start to have a lot of semi-dumb semi-autonomous robots in wide use, we may see "common sense" fractured into a lot of small, solveable problems. I used to say in the 1990s that a big part of life is simply moving around without falling down and not bumping into stuff, so solve that first. Robots have almost achieved that. Next, we need to solve basic unstructured manipulation. Special cases like towel-folding are still PhD-level problems. Most of the manipulation tasks in the DARPA Robotics Challenge were done by teleoperation.
That assumes computers learn as slowly as humans (Score:5, Interesting)
That presumption seems to be precipitated on the theory that a computer intelligence won't "grow" or "learn" any faster than a human. Once the essential algorithms are developed and the AI is turned loose to teach itself from internet resources, I expect it's actual growth rate will be near exponential until it's absorbed everything it can from our current body of knowledge and has to start theorizing and inferring new facts from what it's learned.
Not that I expect such a level of AI anytime in the near future. But when it does happen, I'm pretty sure it's going to grow at a rate that goes far beyond anything a mere human could do. For one thing, such a system would be highly parallel and likely to "read" multiple streams of web data at the same time, where a human can only consume one thread of information at a time (and not all that well, to boot.) Where we might bookmark a link to read later, an AI would be able to spin another thread to read that link immediately, provided it has the compute capacity available.
The key, I think, is going to be in the development of the parallel processing languages that will evolve to serve our need to program systems that have ever more cores available. Our current single-threaded paradigms and manual threading approaches are far too limiting for the systems of the future.