Once upon a time, Turing threw humanity a challenge. Make a computer that can fool me into thinking it’s a human. At least talks like a human.

That article, Computing Machinery and Intelligence, was the official beginning of what we know as AI.

What follows is less glorious. It appeared that it was nowhere as easy as hoped. Cognitive scientists and geeks sloughed away until the mid-seventies, but nobody was very convinced at the results. The problem was that nobody really understood what made us intelligent, learning creatures. What you don’t understand, you can’t reproduce. There were as many theories as there were scientists.

Computer scientists more or less gave up on what they called GOFAI, and focused on more limited problems. The result is the immensely varied bag of tricks we now call AI (genetic algorithms, neural networks, pattern recognition, support vector machines, multi-agents and many others…).

Today’s AI is good at solving limited problems, with carefully prepared input. Think of the speech or character recognition software, industrial robotics, almost autonomous vehicles, translation software, stock price prognostics, social media trawling and nice japanese robot dogs.

But it seems now that Good Old-Fashioned AI might be back in fashion. I read this article by Jeff Hawkins, one of the co-founders of Palm. I’d heard the guy was passionate about AI last year – and meant to read his book On Intelligence. I figured that it would be interesting to hear what a smart embedded computer guy has to say about our own embedded computer.

It seems he decided to have a go, and make software that is able to learn by itself, much like (he claims) the neocortex. It’s called NuPIC, and uses something called HTM, Hierarchical Temporal Memory. He open-sourced part of the code, and invites scientists and developers to play with it, which is cool (though it would have been even cooler if he open-sourced it all – old habits die hard, i guess).

The thing is, whether NuPIC is good or not, i’m reasonably optimistic about the whole problem. Not only do we have loads and loads more processing power at our disposal, neurologists are chipping away at the mysteries of the brain. Functional MRI allows us to map which areas of our brain is active at any one time. Reasonably efficient machine-brain interfaces are being developed all the time. Biogenetics allow us an insight at a molecular level.

So we might just get it right in our lifetime. So the question I, and a whole lot of people, now ask themselves is: OK, but what next ?

How do you make a nice, honest, trustworthy machine ?

• Previous Entry: The Art of Stress-Free Productivity
• Next Entry: Restful web services