IA / Machine learning / Deep learning / Strapon fisting (Page 2 sur 4)

Ca c'est cool.

To achieve real strides will require brain implants with connections to millions of neurons. This will require millions of tiny electrodes, and a control system to synchronize them all. The current state of the art brain-computer interfaces have around 1,000 connections. So, current devices need to be scaled up by more than 1,000 times to get anywhere interesting. Even if you assume exponential improvement, it will be awhile before this is possible — at least 15 to 20 years.

Improvement in IA rests upon progress in nano-manufacturing. Brain-computer interface engineers, like Ed Boyden at MIT, depend upon improvements in manufacturing to build these devices. Manufacturing is the linchpin on which everything else depends. Given that there is very little development of atomically-precise manufacturing technologies, nanoscale self-assembly seems like the most likely route to million-electrode brain-computer interfaces. Nanoscale self-assembly is not atomically precise, but it's precise by the standards of bulk manufacturing and photolithography.

The biggest roadblock is developing the appropriate manufacturing technology. Right now, we aren't even close.

Another roadblock is figuring out what exactly each neuron does, and identifying the exact positions of these neurons in individual people. Again, we're not even close.

Thirdly, we need some way to quickly test extremely fine-grained theories of brain function — what Ed Boyden calls "high throughput circuit screening" of neural circuits. The best way to do this would be to somehow create a human being without consciousness and experiment on them to our heart's content, but I have a feeling that idea might not go over so well with ethics committees.

The key point is that selfish motivations will not magically emerge if an AI has a goal system that is fundamentally selfless, if the very essence of its being is devoted to preserving that selflessness. Evolution produced self-interested organisms because of evolutionary design constraints, but that doesn't mean we can't code selfless agents de novo.