AI: More Than Numbers

Here's a lovely video:



See that Nvidia logo? It's no accident. The main processing unit on a contemporary graphics card (GPU) is a big fat parallel processor -- compared to the relatively serial-based CPU -- and can therefore do a truckload of calculations at once. In some ways, this is more like a human brain, and so it makes sense that GPUs would be of interest to researchers looking at artificial intelligence. Indeed, about five years ago I chatted to David Kirk, Nvidia's 'chief scientist', who seemed to lament the fact that the capabilities of Nvidia's graphics cards were used for... graphics. What it could be used for was making things intelligent, artificially-speaking.

But this video isn't about AI.

InstinctTech is using a GPU to mimic the group behaviour of 4000 relatively 'dumb' entities. It's flock AI writ large; a problem that doesn't need solving and, in gameplay terms, is mostly just a visual effect. What happens next? 8000 entities? 8 million?

That's not to diminish InstinctTech's work -- this is beautiful, technically impressive, and puts a nice oily sheen on Nvidia's well-defined biceps. But I'll be more impressed with a single, computer-native construct that passes the Turing test. Or just comes close. Or, hell, is marginally better than marines in the original Half-Life.

And it doesn't even need graphics.