Neural Networks

[HDRKID]

Defeating a neural network is easy. Researchers are seeing that it is as simple as attaching a sticker to a stop sign. For example, if a stop sign is slightly modified, then the neural network finds it cannot see one. Well, this is a major issue. After all, it affects driverless cars. Sadly, they just keep on going. Yeah, autonomous vehicles need to see all stop signs... like even those that are slightly different from our norm.

Adversarial Traffic Signs – Self-Driving Cars – Medium

How to hide backdoor in AI software

 

[Orpheus Rex]

This pinpoints, generally, the huge chasm between what people think AI is capable of and what it actually is. The current best AI is more stupid than an amoeba. The singularity is nowhere near. This also pinpoints the danger of AI. Current AI isn't dangerous because it's intelligent, but because it's mindboggingly stupid and used to control powerful systems. An AI has no access to platonic forms, as demonstrated by the example above. Yet, if the AI is taught the rule "Do not harm humans" it cannot understand human to encompass all that we understand human to be and to differentiate from things that are like human. If someone is missing a limb, it's likely that an AI would not recognize the person as human - hence the danger.