AI is neither A nor I, and inhumane

from an email threada about self-driving cars, I wrote a couple of short pieces on AI.

On the main point: a custom road designed for cars that can navigate it is called a railroad.

Now about AI:

AI today is neither A nor I

Sunday, October 9, 2022

12:40 PM

AI has been just 10 years into the future for my entire life.

Artificial means an artificer created the thing. A skyscraper had an architect.

We tried and failed to construct intelligent systems, with LISP, knowledge bases, rules and inferencing engines.

So now we think neural networks might work based on a vague analogy to how we think the human brain might work.

We don’t design and construct a neural network. We grow it out of the dirt that is the contaminated data that we think characterizes the problem domain. A neural network is a lossy compressed representation of its training data.

While these networks can do remarkable things, (deepfake porn!), it is in no way intelligence.

In talking to phone company customer service I learn that those people have had their capability of intelligent action removed by the corporation. They cannot get the information needed, and they are denied the authority to act.

Human intelligence means knowing when and how to break the rules.

When a human being [1] interacts with a system (whether it be a macbook, or the IRS, or (shudder) health care) they need to have some understanding of how it operates so they can know what to do.

If you put an AI in between the human and the system, then when it fails (it will), the human’s problem has just become at least twice as big. Instead of “how do I print this document” it becomes “how do I tell this moron to print my document”. I need to know how the system works,  and I need to know how the AI works and fails.

[1] we are not users; we are not consumers; we are agents in what we call society.