More on AI and Hossenfelder's issues

[Dennis P Brown]

Looks like AI using a rather straightforward method in its 'neutral network' (poorly named) and apparently this methodology has a few issues. This one being how it fit's data does not follow expected patterns. The issue is no one knows why.

See: https://www.youtube.com/watch?v=QO5plxqu_Yw

[Frank Sanns]

I refer back to my original post on AI where I said, the computer saw order where the human mind could not. I also said the computer could see order where there was no order. It seems not too much has changed in the nearly 40 years since I said that.


The curve fitting and the number of polynomials is just such an example. There is no way for AI to know if the result makes sense or not because it cannily interpolate with any sort of validity. Once extrapolation comes in, the confidence in its result goes to hell and back.

What a human brings is cross discipline knowledge. While and AI could learn all of all of the disciplines and experiences of a well rounded human, that is a Herculean task. Then there is judgement and confidence. Hard to develop for humans, even harder to develop with AI. There are some tricks but I have yet to see them implemented in any AI. I will keep them to myself here.

Lastly is something that can never be brought into a pure AI system and that is some knowledge from tens to hundreds of thousands of years of experience and knowledge. It is called instinct. We all have the cumulative information from human and animal existence that is passed down through DNA. Each splice contains a vast amount of information that we access on a daily basis but don't realize it. Something as simple as sleep cycles, reproduction, sustenance, judgement, and many many more. We are encoded naturally and to duplicate that takes way more than just knowledge or playing out scenarios for reasonable and unreasonable outcomes.