AI Design Philosophy

Realistic Solutions

• Methods of creating an AI must be described in terms of the tools (programming structures) that will be used to create the AI.
• Just talking of general concepts or vague ideas is futile.
• Why create an AI strictly in terms of the way a human thinks?
• Humans were created by very specific biological selection using a very limited set of tools.
• These deficiencies in our design force us to think in ways that don’t apply to silicon based intelligence.
• The worst “micro computer” programs I have ever seen were written by programmers that had trained and built their skills on “mainframe” computers.
• On “micro computers” their “mainframe” techniques were absolutely horrible.
• The “serial” computer versus “parallel” brain debate in AI.
• The axiom “When in Rome, do as the Romans do!” seems to be a good fit here. (When using silicon, use silicon techniques and when using biology, use biology techniques)
• Our current computers are very fast “serial” devices.
• Our brains are massively parallel devices because of the inherent restriction of a very slow cycle time in biological brains.
• Why give up all the benefits of using the vast serial speed of current general purpose computers by adopting parallel techniques that a serial computer does poorly?
• If you believe that the human approach to intelligence is the only way to create an AI then only designs for AI on massively parallel computers have any chance of success.
• This hypothesis might be true.
• The idea of emulating a 10 billion neuron brain with a single (or small number of) very fast serial computers is not now (or in the foreseeable future) possible.
• If this theory is correct then:
1. All AI research must be put on hold until the correct hardware is designed.
2. The software tools and techniques that can make this new hardware productive must be invented.
3. Sufficient programmers would have to trained to use these techniques.
4. This scenario would set back AI research by at least a generation. (25 years)
• I believe AI can be implemented on today’s high speed serial computers.
• I choose to believe that the advantages of very fast serial computers of today has not been realized.
• I believe AI can be achieved on present day computers with the right set of solutions.
• I don’t believe that emulating the human brain will be the solution to AI, based on high speed serial processors.
• The above 3 beliefs are not facts or set in stone. I have used my knowledge of the computer field for the past 30 years to come to these beliefs, but they are still just beliefs.
• I leave open the prospect that a neural net or other hardware modules could be used by a fast serial computer for certain tasks that would benefit from highly parallel processors.
• Humans use specialized tools to do many things that we can’t do with our bare hands.
• Why couldn’t an AI use all kinds of specialized hardware as it becomes available.
• This specialized hardware is not necessary to build an AI.
• Conclusion:
• Let’s create the artificial intelligence using the advantages that silicon can give us while trying to avoid the mistakes that make the human design truly terrible.

Self Built Internal Structures

• A “chatbot” does not an AI make. (intentionally bad grammar.)
• Words and concepts must be connected to something more than just themselves. (other AI researchers call this the Semantic Grounding Problem”.)
• AI must build it's own internal structures as in building them it would know the way of retrieving them when needed and could enforce their connection to the necessary game.(model).
• The above statement means that an AI must have a rudimentary language communication system from the very beginning.
• We talk to babies in simple word expressions and think nothing of it.
• Why can't an AI have only limited language to start with and then build not only the number of words in it's vocabulary but also it's syntax complexity with time?
• The ability of an AI to eventually be able to read and learn from written text makes it imperative that the AI have a language capability.
• It just doesn’t need to start with a University level of understanding language!
• Teaching an AI about words and concepts should be quicker and more natural if the AI builds it’s database from natural language interactions rather than direct database manipulation.
• This doesn't imply that generating complex sentences is required.
• Children can comprehend much more than they can articulate themselves.
• Concepts will be accompanied by virtual models (games)that could be displayed to a teacher to see what kind of comprehension the AI has on a particular topic.

No Intelligence by Default

• I don't believe that intelligence can be created by just moving around a relatively dumb group of software agents. (It is possible that some agent work is useful in some areas but it is not the basis of intelligence.)
• Some emergent behavior has been found in ant and termite colonies but this emergent intelligent behavior is very limited. (The cooling structure of termite colonies is quite ingenious even though no termite has any idea about building cooling systems.)
• Some agent systems try to emulate the reward and punishment that seems to work in our neurons.
1. A large number of agents would be 1,000 or 10,000 but our brains have 10,000,000,000 neurons. (1 million times as many)
2. Many good ideas are lost to our brains because of this mechanism.
3. How do you keep the good ideas from decaying away and still allow all the junk to leave as it should?
4. Our brains are very bad at recalling very specific detail from memory which is probably a result of the way our brains are made. (I think this is a very bad human attribute.)
• I don't believe that just a Neural Net is the answer to intelligence. (Could be useful in some problem areas.)
• We don’t have good hardware that can create very large neural nets.
• We don’t have software or programmers that can program massively parallel computers.
• If a neural net produces even a correct answer, there is no trace back as to why this conclusion was reached.
• Neural net results would be closer to intuition than to reasoning.
• Can you imagine an AI who makes decisions “because they just feel good”?
• People are intelligent because we have a built-in ability to learn and we are taught for many (20+) years by many other humans.
• Most of our intelligence does not come from experimentation or hands-on exploring.
• (Making experiments is a good thing but not necessary to gain intelligence.)
• It comes from being taught how to do things and model things in our heads by others.
• If we are taught about the way the sun works but have no experimental data or hands-on experience touching the sun, do we really understand the sun or are we just manipulating symbols (words)?
• I say we do have some understanding of the sun. Touching the sun might confuse the important aspects of the sun and make us understand less.
• I have seen this with Chemistry experiments at University. The theory was coherent and useful but the physical experiment just confused the issues.
• (Someone had to do the experiment to come up with the theory but it didn’t have to be me!!)
• Human intelligence does not come quickly or easily in humans.
• Why should we expect to produce a University grad AI in 1, 5 or even 10 years of effort?
• I have read that a baby that doesn’t get talked to, doesn’t develop speech. It seems that we are given the tools for language but that external teaching of language is absolutely required for speech to occur.

Models from the Beginning

• If I said "Do you understand death?" and you answered "yes", I might not believe you.
• However, if you described a picture of what death meant to you and that picture was plausible (if not the one I might think of) then I would probably assume that you know the meaning of death.
• The check is when you are just saying the words "I understand what death is" or if you have built some kind of a model of death.
• The act of translating the words to the model is what most people would call understanding.
• If the picture your words painted in my head were not very plausible then I would assume that you don't have a correct understanding.
• The above analysis is the reason why all concepts in this AI will have a model.
• I don't call it a model but rather a game.
• A game is a special kind of model and I like the idea of a goal, rules and players like the word game implies.

AI Brittleness

• The goal of reducing the "brittleness" of the AI is a good one.
• The adaptability of the AI program has nothing to do with the program using discreet structures or just having pointers to everything.
• A network architecture or a relational database design doesn’t affect the “brittleness” at all.
• The type of database has no bearing on how flexibly the data can be used.
• If you want “fuzzy” values, you can easily program them with an absolute number and a “fuss” factor.
• If these values are interpreted as the first number plus/minus the second number, you have an absolutely fuzzy implementation using absolute numbers.
• If you wanted to add a probability factor, you could easily add that as another absolute number multiplied by a number from 0-1..
• Conclusion: “Brittleness” is a result of how data is used not how it is stored.
• Relational databases are much more efficient that either hierarchical or network databases and can be accessed without discrete code.
• If a link is broken, then an enter branch of a list can be lost. (Unless there are other links that can be used to salvage the first link.)
• If a relational table has a corrupted index, the index can just be remade from the data.
• The data is always stored in a flat, easily retrievable place, and indexes can always be remade if needed.

Real World Required?

• Is a direct connection through sensors and actuators required for intelligence?
• Some researchers say that without a grounding in the real world, AI is nothing but a sham.
• I disagree.
• The fact that Helen Keller couldn’t see detracted from her intelligence?
• If an intelligent human is deprived of some or even all of their senses, are they less intelligent?
• Without senses of any kind, it is probable that humans couldn’t learn and therefore wouldn’t be considered intelligent but once they were intelligent, the lack of sensors would not make them any dumber.
• If an AI has an opportunity to learn (in some way) then whether that connection is directly to the real world or just to a developer, I see no reason the AI couldn’t be considered intelligent in both cases.
• I do believe that an AI can’t be totally useful if it doesn’t have a presence in the real world that is totally under it’s control but I don’t think that this is necessary to create intelligence.
• If you want your AI to be an expert in dealing with moving around or picking things up, then you must have some kind of connection to the real world.
• However, most of what humans find important are not mundane things that can be gotten directly from reading sensors attached to the real world.

Mathematics Required?

• I see no reason to involve Mathematics in an AI design at all.
• All persons less than 10 years old, know nothing of Mathematics and we consider at least some of them to be intelligent.
• Most adults have only a very limited knowledge of Mathematics and they seem to get along just fine.
• I am not saying that Mathematics is totally useless but it could be learned quite late in the AI’s development cycle without any problems.
• Most people seem to get along without any formal logic or probability theory as well.
• These areas are studied to death even when the developers don’t have an AI at all and I think this is one of the reasons no AI exists today. (The universe is NOT all Mathematics!)