AI Research2

 

 

What keeps bottom-up reasoning from becoming a made up hallucination? In science, which deals in large numbers, it is falsifying H₀, the null hypothesis of the experiment which says there is no statistical difference between the result and the random Gaussian bell-shaped curve. The other alternative is the individual case, balancing the pros and cons and then reaching a (logical, common sense) decision. In the real world, both universal law and the individual case should converge upon the truth of the matter. But bottom-up AI currently has no null hypothesis or method to test a result against, and probably needs one, unless the database is curated beforehand for restricted tasks.

 

The above is the last step in a generative AI model, which is essentially determining final parameter weights to assess, “the truth of the matter, given all the evidence.”  On the simpler matter of determining, “generally held facts,” Google DeepMind has successfully fact-checked ChatGPT answers using internet searches. In business, “generally held facts” have to further become, “facts about my operation.” What this says is that, to be practical, there has to be a lot of cutting and fitting of generative AI to the specific case.

 

The fundamental problem with generative AI, why it will likely always require human agency, is that it cannot engage at all in top-down reasoning and therefore in its practical tradeoffs. The apparent top-down reasoning that it does engage, say in “proving” math theorems, is merely a simulation. About top-down reasoning:

 

1)    Top-down reasoning can be very efficient. Consider the millions of steps that bottom-up generative AI must take to reach a decision, as opposed to getting the principle right in the first place. It was a goal of Western thought to seek certain knowledge. Such a quest, which has now become open-ended, now includes the major factors in the context of a situation.

 

2)    Top-down decision making makes possible practical (to a point) tradeoffs in a particular application, for instance between costs and functionality. It has become a pastime to stymy generative AI with trick questions. We got it to demur (choke) when we asked it two specific engineering applications. It could not provide the circuit specifics for designing a 100 watt hi-fi amplifier, and it could not provide the specifics for designing a bridge (although it could provide alternate bridge designs). Finally, the chatbot pleaded, and we think accurately so, “Ultimately the best way to use me is to leverage my strengths. If you need help with broad ideas and information gathering, I’m a great resource. But for tasks requiring specialized knowledge (beginning with physics 101), complex calculations, or ensuring safety and functionality, it’s important to involve a human expert.”

 

A PBS television show captured the essence of generative AI. In the 19^th century, the steam engine materially transformed the U.S. continent, linking coast to coast. Generative AI has been called, “the steam engine of the mind.” New ideas will prevail because they are useful, and not hallucinations.