We Shouldn't Glorify Today's Cleverest Algorithms

I admire your comments and that's a GA from me,

what is your field of work?

Thank you.

We design and build flight control systems for the military for training pilots here and around the world. So, it's a pretty diverse field of engineering mechanical, electrical, and hardware-software components.

AH, I have a different take on it.

I see intelligence as problem solving ability, abstract thought, planning, learning... an algorithm can be written to execute a plan or even learn within parameters, but abstract thought or free will?

Let us review a field mouse. A mouse, lounging in a shady spot (AI can act to regulate it's own temperature as well), notices a low fuel condition (hungry), and is prompted to take action. Options include the field (mostly safe, but food is not very tasty) or the grain bin (the food is a lot tastier and more plentiful, but he is more likely to be seen by the cats). Selection made, he makes toward the destination. Along the way, two opportunities become available: mating and water. He simply chooses. This choice may be based on past experience, might be based on perceived risk, or maybe the more urgent biological need or urge.

Now a different mouse, living in the same area. Each of these choices are presented to him as well, but another set of decisions is made and a different course is taken. Within an algorithm, the same choices will always be made with the same inputs.

If the result of each mouse expedition is the same (fed, watered, mated, return home safe), does it matter what selections were arbitrarily made? Consequence or reward of outcomes based on decisions is a part of the learning experience. Even though the outcome was the same, mouse #1 had a pleasant and uneventful experience, and will try that again. Mouse #2 ran into a cat, and will NOT make that particular selection sequence again, even though he returned home safe.

And, can AI select a paint color for the kids bedroom? (I have an ex-wife who believed she could, and she once did so three time in one year.)

Now a different mouse, living in the same area. Each of these choices are presented to him as well, but another set of decisions is made and a different course is taken. Within an algorithm, the same choices will always be made with the same inputs.

Quick personal background... my sister does ALOT of what is termed AI programming and has helped me out in my dabbling in the area. I have learned the beauty of some of the more sophisticated algorithms is that your last statement isn't always true. As I understand it, some of them take not just the inputs, but accounts for the sequence and the timing of them. So in order to get the program to react the same on 2 different systems, the same information would be needed to input in the exact same sequence at exactly the same time intervals.

She said some programming attempt what psychologists term recency effects. The things that are happening now have more importance than the thing that happened a year ago. So in your example, the mouse that ran into the cat do to a series of choices might make the choice again depending on how long ago it was that it happened and if it runs into it again, it might wait a little longer the next time.

I won't argue the system has "free will" but it is able to make seemingly intelligent decisions based on experience and the more experience it has the more intelligent the decisions become. From my point of view, calling these programs AI is appropriate.

OK. I am with you. That is a good explanation.

I have worked with propositional logic, two value logic, boolean logic for many, many years, and those types of yes/no logic applications fit well within my trade.

One of the things I am sort of stuck on is the degree of truth or fuzzy logic that is (as I understand) a large part of AI. While I can see how it works, I have trouble figuring out how it works... how is that for a fuzzy statement? As I consider, the declaration I made about the mouse outcome is (or is it?) a fuzzy result.

This is certainly food for more thought, and I will consider further. Thanks AH and ChaoticIntellect for another good discussion.

A Monte Carlo simulator produces different results from the same inputs, but there is nothing in the process of generating those results to suggest 'intelligence'.

It may be that a certain amount of seeming randomness or unpredictability makes a clever algorithm or 'AI' seem to be almost human. But this is an illusion.

A Monte Carlo simulator produces different results from the same inputs, but there is nothing in the process of generating those results to suggest 'intelligence'.

The Monte Carlo simulations I have done (not many) have been some function defined by averages, standard distributions, standard deviations and much like you say far from intelligent. Let's consider 2 things though. 1) the output from ten thousand simulations provide some meaningful information that wasn't there in the beginning 2) What if the simulation could redefine it's own functions based on experience?

Is this not what we as humans do? But who defined AI as needing to mimic humans? I would say a squirrel who lets me watch him hide his food only to return later and move it when I'm not watching is intelligent.

From my experience, the more powerful AI's do NOT employ randomness in the programming, they are empowered with the ability to analyze, provide a solution, watch the result of the solution, and use that information to adapt to the same or new situations. I would argue they are able to remember, learn, and adapt, which in the most rudimentary way would suggest they are intelligent.

Definition: Intelligence: the ability to comprehend; to understand and profit from experience.

I see that the difference here boils down to definitions of intelligence.

I personally was thinking of intelligence as something that implies autonomy and self-awareness. That may not be appropriate for AI definition.

The truth is, I can't help a 'knee-jerk' reaction to the sort of algorithms described in this article as 'glorified': algorithms in which machines "make choices" for us. There was another blog some time ago which gave me the same queasy feeling - it was about machines involved in making "ethical" enforcement decisions for compliance with medications. The agenda of developing machines which make "choices" and impose them on human beings is not acceptable, and the soft-sell of this sort of "AI" by parties to the "pharmaceutical compliance" agenda is frankly, a big warning sign.

This is where I draw the line between humans and machines. I do not want a machine to make choices for me, ever, or to be involved in overriding the choices freely made by me or by any other autonomous living being, with a "program" generated by someone else.

On the other hand, I can get really excited about the prospect of machines that learn (=acquire real data by means of sensor systems, authenticate, store, compare, and organize the data in a logical manner), and become more intelligent (= provide better analytical output) the more they are used, where the purpose of the machine is to better inform me of factual data and patterns detected by the machine. If this is AI, bring it on, and I will gratefully "glorify" it!

I also see a genuine value in programs such as Monte Carlo sims, to incorporate uncontrolled variables in the analysis of controlled inputs, and generate statistically relevant amounts of simulation data that is not available as real data, as an aid to the evaluation of our decision-making process in real situations with uncontrolled variables. But never as a substitute for the decisions made freely by autonomous beings, who also are able to question whether the program and the data was adequate to the task, or whether something important has been overlooked and omitted from the decision process.

Some thoughts about definitions of intelligence:Sensory vs intelligence systems.- Sensory systems are data acquisition features.- Intelligence is a data processing feature, but also depends for effectiveness on effective data acquisition and if possible data quality assessment.- Assessment of data validity is best performed by comparing data from more than one test modality.

- Intelligence advances by orders of magnitude when there is the capability to acquire and compare test data from different modalities in order to identify and mitigate sources of error.

That is applicable to AI/machines as to humans or animals. So in terms of a continuum of artificial intelligence, processing and comparing and correlating data from different test modalities is a sign of a higher intelligence potential. Whereas the clever algorithm that relies on one data type or source is not so clever.

Please read (or reread) some Douglas R. Hofstadter, namely Godel, Escher, Bach: An Eternal Golden Braid and Metamagical Themas. Then write something more meaningful.