Maybe We Will Finally Learn More About How A.I. Works

>I'm sorry but this whole tripe of, "We don't even know how they work" is the stupidest part of the current AI hype trend.

Well it's true.

>They pretend like we are clueless about what's going on inside our own algorithms and it's just silly at this point.

Nobody is pretending about anything. They're not "our own algorithms". The entire point of neural networks is that we don't know much or really anything about how to teach intelligence to a machine and so we figured out a way to get the machine to teach itself given data, structure and an objective. Nobody knows what algorithms, processes or constructs neural networks employ to perform their objective once trained.

>It would be like saying Jackson Pollock didn't understand how his paintings are even being made because he doesn't understand every step of the physics of the paint

No it would be like saying we understood how human intelligence worked because we know how neurons communicate.

No one, not even the creators of GPT-4, predicted that it would score well enough on the Bar Exam to get into a good law school if it were a person. They didn't have a clue. It was, "Let's build it and see what happens."

And when some lab somewhere starts up the AI that is going to kill us all, they're probably not even going to consider that outcome to be in the set of possibilities.

But yeah, AI researchers know their subject well enough that it wasn't just an accident that GPT-4 turned out more capable than any language model that had come before, and I'm sure AI researchers will continue to find ways to make AIs more capable, more powerful.

I wouldn't be surprised if there are 0 people at nytimes who understand how A.I. works. There is a difference between understanding how it works, and being able to backtrack through huge chains of statistics math that is involved in producing outputs. I think this is the point that is being misinterpreted and abused. It is more accurate to say we can't reverse engineer the exact calculations that are involved in an A.I. making specific determinations and producing certain results with a reasonable amount of time & effort. That is different from saying people "don't understand A.I."

This is more of a challenge with newer, larger technologies (deep learning and large language models). The operation of older, smaller scale A.I. systems (rule induction, tree induction, [smaller] neural networks) are much more clear.

I've been working in A.I. and analytics for over 30 years and agree that throwing up one's hands, complaining about how "we can't understand" what A.I. is doing is a bit theatrical.