Who decides what gets built in the world? And when?

Obviously an energy machine or a computing machine design are the best hopes for a technological breakthrough to save us all, but efforts toward “good” problem solving projects such as global warming, global hunger, and a sustainable energy have historically been starved. Why should AI be any different?

Unless there’s a new way of thinking dawning, there is simply not enough money to build strong AI by 2025. There is enough science. There is enough data. There are enough words. I bet the money isn’t there.

On one side, you have the knowledge-based expert system / inference engine / symbolic processing folks, who think that the key to getting a machine to think and act like a human is to model the structure of knowledge by adorning symbols with relationships and properties, then throwing compute cycles at that massive data structure until you can process enough of it to start making human-like responses. This approach has produced some very useful systems that are in wide use today, but I’m skeptical about it ever producing anything that acts like a human being.

On the other side, you have the neural networking folks, who care very little about the structure of knowledge and view it as an emergent property of larger and larger networks of self-organizing pattern recognition systems. The neural network people have the advantage that, to a certain extent, they don’t have to worry about the structure of knowledge. They merely have to mimic something like the structure of the brain and they’re likely to get interesting results.

The other advantage of the neural approach is that it’s very easy to model when you have the same level of computing power as the brain. When you can simulate a hundred billion neurons, all connected together via about a hundred trillion to a quadrillion synapses, you’re there, to some degree.

We’re just not that far off from being able to do that simulation. (In fact, I have a spreadsheet that says that you could probably build such a system today if you were willing to throw hundreds of millions of dollars at the problem and maintain a network of about 100,000 fiber-optic cables.) But the trick to such systems is that we don’t know quite enough about all the various ways that neurons process synaptic information and, once we know that, we don’t know enough about how the brain accomplishes various functions through different local patterns of connections and, maybe even more important, how those local, limited-function regions connect together to produce the flexible system that is the human brain.

So it’s quite possible that, long before we can produce something that acts like a human, we’ll be able to produce systems that do incredibly useful work, but which behave more like insane humans, or even something that’s completely alien but pretty smart. Given that proviso, I think that 2025 is an entirely reasonable date.

If AI advances within the next 15 years to the point that a large and increasing proportion of unskilled (and typically unionized) jobs can be done by machines, then the implications are as dire for unskilled workers (and labor unions) as they are positive for society as a whole.

Andy Stern’s SEIU has been the fastest-growing union in the nation for years due primarily to organizing just this class of worker. One wonders: will the purple-shirted thugs at recent health care protests cease to exist, or will they be replaced by robots, too – Andy Stern’s own private army of Obamaist Cylons.

http://www.pecuniarius.com/blog/?p=182

And I’m fairly certain that we won’t even be close in 15 years. Not only do all those positions require human language and vision capabilities (which are both far, far beyond the abilities of current AI systems), they also require a great deal of human judgement (police?!? With guns?!?).

No, that pattern has been, and will likely to continue to be, that we will have artificial systems that can do a limited (but useful) subset of tasks that humans can do — but only under controlled conditions and in a ‘brittle’ way that shows little of human flexibility and robustness. So, OCR is incredibly useful — but the scans have to be clean, straight, and hi-resolution. Humans can read mildly degraded text easily that OCR systems fail at. But that’s OK — OCR is still extremely useful. Similar situation with natural language translation — artificial systems don’t do it the way humans do, and they make absurd errors that no human translator would make, but they are very useful in providing a rough, first cut.