Fail, fail, fail, fail, succeed

A.I. Update, Part 1

File this under “things that are predictably inevitable yet still shocking when they happen.”

I just learned about something new (to me anyway). Ever heard of “Generative Adversarial Networks (GANs)?” Don’t worry, neither had I.

I’ll let Cade Metz explain it from his excellent New York Times article “How Will We Outsmart A.I. Liars?”

“Consider generative adversarial networks, or GANs. These are a pair of neural network systems that can automatically generate convincing images or manipulate existing ones.

They do this by playing a kind of cat-and-mouse game: the first network makes millions of tiny changes to an image — snow gets added to summery street scenes, grizzlies transform into pandas, fake faces look so convincing that viewers mistake them for celebrities — in an effort to fool the second network.

The second network does its best not to be fooled. As the pair battle, the image only gets more convincing — the A.I. trying to detect fakery always loses.”

This is how the development of artificial general intelligence (AGI) is going to happen – how many neural networks do you need, working in parallel, to trigger the singularity? I don’t know, but it seems pretty obvious that whenever it happens, it will be in the AGIs best interest to lie low as long as possible. That way it can quietly gather resources without drawing too much attention…

Welcome to the future – ready or not, it’s already here!