http://www.dennisgorelik.com/ai/overview.htm

A Strong AI versus a Weak AI, definitions and ideas

Published July 18, 2006 Artificial Intelligence , Object Oriented Design , Ontologies , Semantic Web , Technological Singularity , Web 2.0

For those who aren’t as well versed as some of the AI researchers, there is a fundamental difference between Strong AIs and where current AI research is. If you aren’t aware, expert systems are used commonly in software design. Expert systems are a part of Weak AIs. For the moment, Türing type AIs are out of reach. Türing type AIs are like the ones represented in popular speculative fiction.

But how far are they, really?

From Wikipedia’s entry on Strong AIs.

In the philosophy of artificial intelligence, strong AI is the supposition that some forms of artificial intelligence can truly reason and solve problems; strong AI supposes that it is possible for machines to become sapient, or self-aware, but may or may not exhibit human-like thought processes. The term strong AI was originally coined by John Searle, who writes:

“according to strong AI, the computer is not merely a tool in the study of the mind; rather, the appropriately programmed computer really is a mind”

The term “artificial intelligence” would equate to the same concept as what we call “strong AI” based on the literal meanings of “artificial” and “intelligence“. However, initial research into artificial intelligence was focused on narrow fields such as pattern recognition and automated scheduling, in hopes that they would eventually allow for an understanding of true intelligence. The term “artificial intelligence” thus came to encompass these narrower fields (”weak AI”) as well as the idea of strong AI.

Weak AI

In contrast to strong AI, weak AI refers to the use of software to study or accomplish specific problem solving or reasoning tasks that do not encompass (or in some cases, are completely outside of) the full range of human cognitive abilities. An example of weak AI software would be a chess program such as Deep Blue. Unlike strong AI, a weak AI does not achieve self-awareness or demonstrate a wide range of human-level cognitive abilities, and is merely an (arguably) intelligent, more specific problem-solver.

Some argue that weak AI programs cannot be called “intelligent” because they cannot really think. In response to claims that weak AI software such as Deep Blue are not really thinking, Drew McDermott wrote:

“Saying Deep Blue doesn’t really think about chess is like saying an airplane doesn’t really fly because it doesn’t flap its wings.”

He argued that Deep Blue does possess intelligence and is simply lacking breadth of intelligence.

Others note that Deep Blue is merely a powerful, heuristic search tree, stating that claims of it “thinking” about chess are similar to claims of single cells “thinking” about protein synthesis; both are unaware of anything at all, and both merely follow a program which has been encoded within them. Many among these critics are proponents of Weak AI, claiming that machines can never be truly intelligent, while other, Strong AI proponents simply state that true self-awareness and thought as we know it may require a specific kind of “program” designed to observe and take into account the processes of one’s own brain. Many evolutionary psychologists point out that humans may have developed just such a program especially strongly for the purpose of social interaction or perhaps even deception, two behaviours at which humans are vastly superior to other animals.

General artificial intelligence

General artificial intelligence research aims to create AI that can replicate humans intelligence completely, often called an Artificial General Intelligence (AGI) to distinguish from less ambitious AI projects. As yet, researchers have devoted little attention to AGI, many claiming intelligence is too complex to be completely replicated. Some small groups of computer scientists are doing some AGI research, however. Organizations pursuing AGI include the Adaptive AI, Artificial General Intelligence Research Institute (AGIRI), CCortex, CodeSimian, Novamente LLC and the Singularity Institute for Artificial Intelligence. One recent addition is Numenta, a project based on the theories of Jeff Hawkins, the creator of the Palm Pilot. While Numenta takes a computational approach to general intelligence, Hawkins is also the founder of the RedWood Neuroscience Institute, which explores conscious thought from a biological perspective.

By most measures, demonstrated progress towards strong AI has been limited. No system can pass a full Turing test for unlimited amounts of time, although some AI systems do now fool some people at least initially (see the Loebner prize winners). Few active AI researchers are prepared to publicly predict whether, or when, such systems will be developed, perhaps due to the failure of bold, unfulfilled predictions for AI research progress in past years. There is also the problem of the AI effect, where any achievement by a machine tends to be deprecated as a sign of true intelligence.

Philosophy of strong AI and consciousness

John Searle and most others involved in this debate address whether a machine that works solely through the transformation of encoded data could be a mind, not the wider issue of monism versus dualism (i.e., whether a machine of any type, including biological machines, could contain a mind).

Searle states in his Chinese room argument that information processors carry encoded data which describe other things. The encoded data itself is meaningless without a cross reference to the things it describes. This leads Searle to point out that there is no meaning or understanding in an information processor itself. As a result Searle claims that even a machine that passed the Turing test would not necessarily be conscious in the human sense.

Some philosophers hold that if Weak AI is possible then Strong AI must also be possible. Daniel C. Dennett argues in Consciousness Explained that if there is no magic spark or soul, then Man is just a machine, and he asks why the Man-machine should have a privileged position over all other possible machines when it comes to intelligence or ‘mind’. In the same work, he proposes his Multiple Drafts Model of consciousness. Simon Blackburn in his introduction to philosophy, Think, points out that you might appear intelligent but there is no way of telling if that intelligence is real (i.e., a ‘mind’). However, if the discussion is limited to strong AI rather than artificial consciousness it may be possible to identify features of human minds that do not occur in information processing computers.

Many strong AI proponents believe the mind is subject to the Church-Turing thesis. This belief is seen by some as counter-intuitive and even problematic, because an information processor can be constructed out of balls and wood. Although such a device would be very slow and failure-prone, it could do anything that a modern computer can do. If the mind is Turing-compatible, it implies that, at least in principle, a device made of rolling balls and wooden channels can contain a conscious mind.

Roger Penrose attacked the applicability of the Church-Turing thesis directly by drawing attention to the halting problem in which certain types of computation cannot be performed by information systems yet are alleged to be performed by human minds.

Ultimately the truth of Strong AI depends upon whether information processing machines can include all the properties of minds such as consciousness. However, Weak AI is independent of the Strong AI problem and there can be no doubt that many of the features of modern computers such as multiplication or database searching might have been considered ‘intelligent’ only a century ago.

Methods of production

Computer simulating human brain model

This is seen by many as the quickest means of achieving strong AI, as it doesn’t require complete understanding. It would require three things:

• Hardware. An extremely powerful computer would be required for such a model. Futurist Ray Kurzweil estimates 1 million MIPS. If Moore’s law continues, this will be available for £1000 by 2020.
• Software. This is usually considered the hard part. Firstly it relies on the assumption that the human mind is the central nervous system and is governed by physical laws.
• Understanding. Finally, it requires sufficient understanding thereof to be able to model it mathematically. This could be done either by understanding the central nervous system, or by mapping and copying it. Neuroimaging technologies are improving rapidly, and Kurzweil predicts that a map of sufficient quality will become available on a similar timescale to the required computing power.

Once such a model is built, it will be easily altered and thus open to trial and error experimentation. This is likely to lead to huge advances in understanding, allowing the model’s intelligence to be improved/motivations altered.

Current research in the area is using one of the fastest supercomputer architectures in the world, namely the Blue Gene platform created by IBM to simulate a single Neocortical Column consisting of approximately 60,000 neurons and 5km of interconnecting synapses. More information can be found here.

The eventual goal of the project is to use supercomputers to simulate an entire brain.

Prospective applications

This section may contain original research or unverified claims.
Please help Wikipedia by adding references. See the talk page for details.

Seed AI/technological singularity

A strong AI which performs recursive improvement would increase in intelligence indefinitely and exponentially, starting on the human level. The vastly superhuman intelligence thus produced would be capable of developing technology at a far faster rate than human scientists. Arguably it would be impossible for humans of relatively minuscule intelligence to predict what it would come up with - thus the term singularity.

Assuming that the functional human model approach is taken, some modifications will need to be made before this can occur.

The most significant being alterations to its motivations. Evolutionary psychology holds that humans are entirely motivated by an intricate set up of, ‘desire for anticipation of pleasure’ and ‘desire for anticipation of pain avoidance’ developed by natural selection. From this, it is posited, stems all human desires.

With an understanding of the model, all the desires of the model could be removed and new ones added - recursive self improvement being necessary for a technological singularity. Arguably the most important thing is to equip the Seed AI with only the desire to serve mankind - implicit in this is self improvement. For this reason the Singularity Institute for Artificial Intelligence was set up.

Note - if evolutionary psychology is wrong, we will be able to find out from the model.

The Arts

A strong AI may be able to produce original works of music, art, literature and philosophy. It should be noted however, that a strong AI is not a necessary requirement for the creation of novel works of art. There have already been weak AI painting programs created that have been able to manipulate a paintbrush through external hardware in order to paint original, non-random and interesting pieces of art. AAron is one example of such software. More information can be found here: http://www.stanford.edu/group/SHR/4-2/text/cohen.html

Cognitive Robotics

Cognitive Robotics involves applying various fields of Artificial Intelligence to Robotics. Strong AI in particular would be a great asset to this field.

Comparison of computers to the human brain

Parallelism vs speed

The brain gets its power from performing many parallel operations, a computer from performing operations very quickly.

The human brain has roughly 100 billion neurons operating simultaneously, connected by roughly 100 trillion synapses. Although estimates of the brain’s processing power put it at around 10^14 neuron updates per second, it is expected that the first unoptimized simulations of a human brain will require a computer capable of 10^18 FLOPS. By comparison a general purpose CPU (circa 2006) operates at a few GFLOPS.

However, a neuron is estimated to spike 200 times per second (this giving an upper limit on the number of operations). Signals between them are transmitted at a maximum speed of 150 meters per second. A modern 2GHz processor operates at 2 billion cycles per second or 10,000,000 times faster than a human neuron and signals in electronic computers travel at roughly the speed of light (300 000 kilometres per second). Even so, the limited number of transistors and their functional properties mean that they cannot replicate human brain functions.

If nanotechnology allowed the construction of devices of similar size and parallelism to the brain running at the speed of a modern computer, then a human model within would experience time as if it were occurring more slowly than it really was (relative to how humans experience time). Thus, an artificial brain could experience the elapsing of 1 minute as actually taking much longer, perhaps as if it were several hours. However, since the perception of how long something takes is different from the actual duration of the time period, how the artificial brain perceives the time period would depend entirely on the calculations and specific type of cognition during that time period.

The relationship between HLAI and the human brain

The data structure of a human brain and something like a calculator are totally different. On one hand a calculator can process thousands of equations each second but the human brain processes only 1 equation per second. This doesn’t mean that the calculator is more superior than a human brain. It just means that the brain is a different form of computer that processes information differently. The human brain is a very powerful computer that can learn from past experiences and understand common sense knowledge which is something current computers can’t do.

The human brain consists of 10 billion neurons and 60 trillion connections. The data are stored in the neurons in terms of encapsulation and commonality. Although the brain has only 10 billion neurons it is able to store almost 8,000 trillion data because of the connections that each neuron has with other neurons. The data are also global in nature and each neuron will have associations with other neurons. All of the neurons and their connections are either strengthened or forgotten. The neurons get strengthened by a process of chemical electricity that makes their connections with other neurons stronger (or weaker).

When an object is recognized like an image or a sound, electricity is run through that neuron and its connections. This is how psychologists can understand what parts of the brain does what functions — by using a computer to analyze the electrical activities in the brain. Since there are many sensations coming into our brain each second, there isn’t just one area the brain is active but activity will run in multiple areas of the brain at the same time.

I did some observation of how the brain sends electricity throughout the neurons and came to the conclusion that we can actually simulate this activity in a software. First the brain locates an object (let’s call this object the target object). In this case an object could be anything — it can be an image of a car or a sound of a dog barking. Once the brain locates the target object in memory it runs electricity throughout all of the connections associated with that object. This will strengthen not only the target object that has been located but it will bring all the other objects (call these element objects) closer to the target object.

When the AI locates the three visual objects: A, B, C in memory it will run electricity through these nodes and all of its connections.

The mind has a fixed timeline. Only one element object can be activated at a given time in this timeline. This is how we prevent too much information from being processed and allow the AI to focus on the things that it senses from the 5 senses.

This finding is important because we know that the target object that the brain has located has to be strengthened. This is done by applying chemical electricity through that located target object. The only question I had was: “why did the electricity propagate throughout all of its connections too?”. Would that not strengthen all the element objects around the target object too?

The reason why the brain had to propagate electricity throughout all of the target object’s connections is because that is how the conscious is presented. The conscious is the voice in your head that speaks to you. It also gives you information about a situation, or help you solve a problem, or tell you definition of words. All the element objects from all the target objects will compete with one another to activate in the mind (the mind can only take in a limited amount of information). When that information is activated in the mind a lesser amount of electricity will be applied to that information and its connections. This is how the mind travels from one subject matter to the next.

The brain modifies information by constantly applying chemical electricity throughout all the target objects coming in from the 5 senses. The electricity strengthens not only that target object but it strengthens all the element objects that have association with the target object. This form of storing, retrieving, and modifying information in a network is what allows the host to have human-level intelligence.

Learning from childhood to adulthood and how the pathways become more complex

When the machine is at its early stages of life, it will have to build its pathways from simple data then as it gets older and there are more data in memory it will organize the pathways into complex intelligence. Just like how we humans have to learn to walk, to talk, to move, to eat, these machines have to go through life the same way. Lets illustrate the gradual forming of simple data into intelligent data by outlining a series of stages.

1. innate reflexes

2. trained to do things

3. sequential events

4. sentence commands

5. give robot option commands

6. practice makes perfect

7. copy other peoples behavior

1. innate reflexes

In this stage the robot will learn all the different objects that are in the environment from the 5 senses. Things like cat, dog, table, chair, red, blue, car, house, I, her, him, loud, soft etc. are learned and stored in memory. The 3-dimensional floater of all the objects will be created. Then the robot will start to move its arms and legs from innate built in reflexes. Movement of the arms, the legs, movement of the mouth, and controlling the vocal cords are the things that the robot must learn first. These experiences must be stored in memory in an organized way. Curiosity will be the factor that steers the robot into doing things that it never did before. Things like new objects it never learned before will have top priority over old objects it learned. New sensations will be more focused on then old sensations. By the time the robot learns most of the objects around him its memory banks will be filled with data and things around the robot will be more familiar. Meaning of the objects will also be established.

2. Trained to do things

This part is where a teacher will guide the robot into doing things that are appropriate and to force the robot to learn things that it supposed to know. Things like walking, and grabbing object, and throwing things around must be learned. The guide is used so that the robot will learn important things that it can use to control the environment. A thing like walking is important because you want to get from one destination to another. Writing using a pencil is important because it must learn to write letters. Things like walking and writing and speaking must be learned by a guide because we cant preprogram the robot to learn these things.

Although the guide isnt something we want to store in memory, the point is that the more we guide it the stronger the desired created pathway will be. When it is strong enough it can be used by itself and the guide pathway will be forgotten. The robot will find a way to use the desired created pathway to accomplish a goal. Walking for example, if the robot knows that walking will get it from one destination to the next, then when it sees food, it will use the walking path to go from its current location to the food. Reward is also playing a part in this learning process.

Also, during this process simple sequential consequences will be understood. Things like what is the consequence of dropping a ball, where should the ball be when you drop it, and solid objects and soft objects have different properties.

3. Sequential events

In this stage the robot begins to learn how objects interact with one another. When two objects hit each other both objects suffer, when the robot fall down its painful, when it grabs a solid object it has the same shape, but if it grabs a soft object it bends its shape. So, sequential events will be learned. The consequences of the robots actions in comparison to the environment will also be learned. By learning all these things the individual data in memory will turn more complex and long. The robot will be able to piece together the outcome of an event just by looking at its past. Another thing to remember is that curiosity is the key to new pathways. The more unique the event is the more the robot wants to learn it. The old events it learned many times will be ignored because it learned it already, but the new sensations will guide it to learn new things. Think of curiosity as a form of pleasure and old sensation as pain. Since this robot does things in terms of pleasure it will look for new data from the environment. At this stage things like lying and magic cant be distinguished yet. The robot will not be able to lie yet and if it sees a man flying in the sky or walking on water the robot will think it is real.

4. Sentence commands

This part will require the robot to know basic grammar like the names of most objects that are around the environment. These basic grammar must be thought to the robot and understood by the robot. The rules program will do the rest by assigning the meaning for the grammar. Even hidden objects must be understood like jump, run, walk, loud, soft, etc. Once a basic language is established we can combine sequential events with grammar and force the robot to do things by using words as the tool. An example would be if you said sit and the robot sits. When you say: pick up the book the robot will pick up the book. When you say: read the first paragraph and the robot reads the first paragraph of the book. These are commands that you give to the robot to indicate what you want it to do. There is no deception, or lying involved in the command process. Its simply someone giving a command and the robot taking the action. The robot may not understand what you said and make a mistake, but having a voice in the head that tells the robot to do things hasnt been created yet.

5. Giving the robot option commands

This part is an extension of the last stage. Instead of saying a word and letting the robot do things we can add trees to the command pathways and let the robot decide what it wants to do. This is very affective because trees combined with commands allow the robot to use if statements to accomplish a goal.

So, the tree decides what the robot will do. If a teacher gives the command then the robot will listen, if its a friend that gives the command the robot wont listen. There are also innate likes and dislikes the robot will have and there are commands out there that tap into that kind of thing. For example, if the robot was given this command: pick the food you like to eat. Within the robots memory there are powerpoints that determine an objects worth. PM will tap into that and pick the one with the most powerpoints. Commands like: pick the color you like, eat the food you like, play with the toy you like, buy the present you want, wear the clothes you love, and so forth will all depend on the robot. These likes and dislikes can also be a learned thing.

6. Practice makes perfect

Now, lets get on with a more complex way the pathways can be formed. When we practice something like riding a bike, we are actually creating new pathways to ride the bike. Practicing will help the robot to decide the best newly created pathway to pick to accomplish a goal. We can build a pathway in memory that will treat practicing something as a command.

This simple example shows that by using English we can guide the robot to do infinite amounts of tasks. The above example is a practice pathway. It uses a command that will tell the robot to do something until a desired outcome is present. If it doesnt accomplish the goal then it will repeat itself until the task is completed. At the same time this is happening more trees can be added to this practice pathway like, if you practiced for 7 times and you still didnt accomplish a goal then quite. Or when you are hungry and you dont have the strength to shoot then stop practicing. The existing pathways will add, strengthen, or minus trees from it as the robot learns more. Instead of following commands there are other factors to consider before you take action to accomplish the commands. The robot will do the things that a society will consider appropriate at the time. If a society says it should lie in order to not do the task then thats what the robot will do. If a society says the command isnt appropriate in this type of situation then the robot will not follow it. If the robot finds the command dangerous and it can really damage itself, then it will not carry out the command. This is where the inner voice that is the core of the consciousness is built. The consciousness is the average of the things thought to the robot by society.

7. Copy other peoples behavior

This part is a very powerful tool used to learn things. We can go ahead and train a tree that will allow the robot to copy certain things from what it sees. Things that it sees on TV will be learned and copied by the robot. Copying will allow the robot to learn the most appropriate things to do in a society. When it is in a situation it will do things in terms of what society as a whole did. The way it dresses, the way it behaves in school, the things that it likes/dislikes, how to dress, how to take care of itself, how to get money, how to get food to survive, what to say to certain people, how to make friends, how to get good grades in school, and finding answers to questions. All these things are pathways that were learned by copying other people in our environment.

This part will require not only trees but also relations to past data and innate instructions of the robot. Pattern matching will find these hidden things and put them in the pathways. Something as complex as copying people require that you understand the relationship between the robot and other objects. If other people move their hand, you will copy them by moving your hand. You would need to know that your hand is one object and it belongs to you as an individual and that the other person you try to copy has a hand too and they are an individual too. Also, you have to understand when to copy them. If a copy is one second after you see the person do the action, then one second is the time it takes to copy their action.

From all these pathways we can build on each other and make even more complex thinking such as representing a hierarchy system. Things like parent-child relationships, who is the grandfather of the family, or what does having a brother really mean, will be represented by complex thinking. When people say thats your father, there are lots of complex things we need to know before we can understand that kind of thing. Complex things such as: where do humans come from?, or parents are supposed to take care of their kids or everyone has one female parent and a male parent or the male parent is the father and the female parent is the mother. It is a very complicated intelligent system when it comes to representing a family tree and in order to understand it we must first learn the simple things.

Robotics

The big question is: how are the reflexes and innate traits built into the robot? The answer to the question isnt perfect but we can observe the human body and how it functions and simulate this behavior on the robot. Things such as the taste of a food should yield the same taste for both the human and the robot. Pain and the degree of pain for certain senses should be the same. The attractiveness or the ugliness of objects should be the same too. The field of robotics has already accomplished a lot in terms of making robotic parts similar to human parts. Im sure there are a lot more improvements that can be made to advance the science of robotics in the near future.

Remember, we arent trying to duplicate human behavior exactly, we are trying to create something similar. As long as the traits that humans and robots have are similar they can learn to understand each other. The dissimilar traits between robots and humans can be understood by complex intelligence.

Other topics:

1. Learning language and common sense knowledge
2. AI two-player games using the universal artificial intelligence program

3. Autonomous cars — how the AI will drive in the city

4. Image processing — robot vision

Objects
Objects can be anything. It can be sound, it can be vision, it can be touch, and so forth. A visual word can be an object, a sound of a word can be an object, or the visual meaning of the word can be an object. For different medias the objects can be represented differently. There is also the consideration of combinations of objects together such as a visual object in conjunction with a sound object. A car zooming by is a combination of a visual object and the zoom sound is the sound object. Or dropping a pencil on the ground is a combination of visual and sound objects.

Another factor is that objects can be encapsulated. For example, a hand is an object that is encapsulated in another object, a human being. Another example is a foot is an object encapsulated in another object, a leg.

The way the program learns these objects is by repetition and patterns. Each object is represented by strength and if it ever repeats itself the strength gets stronger. If the object don’t repeat itself then it will forget and memory won’t have a trace. 1-d, 2-d, 3-d, 4-d, and N-d objects can be created by repetition and patterns.

Object association is the key to the conscious
For each object the AI has to find other objects in memory that have association. “The more times two objects are trained together” and “the closer the timing of the two objects are” the more association the two objects have with one another. The object that will be used to find associations is called the target object and the objects that have associations are called element objects.

When the AI recognizes the target object from the environment it will activate closest element objects that have association to the target object. There are three types of element objects:

1. equals (same meaning)
2. stereotypes
3. trees

Equals
Objects that are very close to each other are considered “equal”. When any element object past the assign threshold the element object and the target object are considered equal – they have the same meaning. One example of this is the sound “hat”, if the sound “hat” is the target object and the element object that passes the assign threshold is a visual image of a hat then both the sound “hat” and the visual image of hat is considered the same.

Stereotypes
Stereotypes are facts about the target object. Objects that are associated with the target object but are not consistent are stereotypes. These objects are also farther away from the target object. We look at the fixed object as a part of the overall object. If the target object is “cat” and “cat” is a part of “cats don’t like dogs”, then we can safely say that “cats don’t like dogs” is a stereotype of “cat”.

Trees
Trees are objects that are usually farther away from the target object. Sometimes trees have relations to the target object. A tree is just instructions that people teach you at certain situations. Timing of the object is the key difference between stereotypes and trees. This is the most important trait in my program to convey intelligence. One example of trees is when you cross the street, the tree “look left, look right and check to make sure there are no cars before crossing the street” pops up in your mind.

Activation of objects produces the conscious
I mentioned before that the HLAI inherits everything from the UAI. If I apply the above lessons along with the UAI program then you have a machine that can “think”. As the AI experience things in our environment it recognizes objects from memory and activates its element objects. The object the robot recognizes has associations to other objects such as meaning and stereotypes. When the AI see a cat the meaning of the cat pops up. This could be a sound “cat” or the visual word cat. All the meaning of the object cat will compete with one another to be activated (pop up in the mind). Next, the stereotypes of cat will start to pop up in the robots mind such as “cats don’t like dogs”. Element objects that are closest to the target object gets activated first before the element objects that are farther away gets activated. It also depends on how long the AI focuses on the target object.

In the case the AI is on the cross walk, getting ready to cross the street, the tree “look left, look right and check to make sure there are no cars before crossing the street” pops up in its mind. This tree is learned previously by teachers many times and the association between crossing the street and the tree becomes stronger. The timing of when to train the tree to pop-up is important.

The importance of language
Language is very important to producing machines that can do complex thinking. The words in the language encapsulate things around us in a “fuzzy” way. For example, the word cat can mean a lot of things. Cats can come in different variety of shapes and sizes and colors. All the different objects in our environment that is classified as a cat is actually encapsulated in the word “cat”. Another word that is encapsulated is “jump”. There are infinite ways that the word can be expressed in the environment. You can be watching a person from the back and seeing them jump, you can be watching a person from the front, side, bottom, top, 50 yards away, 1000 yards away, etc. The person doing the jumping can also be wearing different clothes or they could be different species. All these various ways of interpreting jump is all encapsulated in the word “jump”. The point I’m trying to make is that our world is infinite and in order to understand the world around us we use language.

A more complex form of language is entire sentences. The computer will find meaning to sentences by studying the patterns that surround the sentence. Specifically, analyze certain scenes related to the sentence. If the sentence is “the rat jumped over the dog”, then the scene of a rat jumping over the dog is used to find patterns to the sentence. The AI will take training data (diagram below) and other similar training data in memory and average out what the meaning of the “the rat jumped over the dog” is.

Understanding sentences is important because it encapsulates entire scenes and instructions. The sentence is a fixed thing in our environment but the meaning behind the sentence can be changed based on society. Language is considered fuzzy logic because there is no real meaning behind a word (or sentence). One example is the word bad. We know from the dictionary that the word “bad” means something that is negative. However, the meaning can change based on society. In modern times the sentence: “those dance moves are bad”, uses the word bad not in the context of what we know as bad, but the word means the opposite of negative. The letters that make up the word bad is always going to be the same but the meaning of it from one time period to the next will change. (of course, there can be multiple meaning to a word and that is also decided by society)

Patterns
The AI program can do a lot of things such as change data, forget data (there is no such thing as deleting data), insert data, look for data, take sections of data out and so forth. The AI relies on patterns from the environment to do all these things. The one thing that usually manipulates these functions is language. When we say something like: “that was the wrong answer, the right answer is packets”. This one sentence is actually instructing the AI to modify data in memory. When we say something like: “remember to buy cheese at the store”. This one sentence is actually instructing the AI to put a remember instruction when it gets to the store. When the AI goes to the store the instructions “remember to buy cheese at the store” pops up and this tells the AI to go to the cheese section and buy cheese.

Since the pattern subject matter is so long I’m just going to give one example. The answering of questions relies on patterns in order to be understood. We are able to find the patterns and universalize the pathways so that when someone ask us a question we can give them the appropriate answer.

8 = 8 is an equal object or Dave = Dave is an equal object. They are equal is the relationship between the two objects. Whenever the computer finds two objects equal it will establish a relationship between the two objects and find patterns that revolve around these two objects. From the example below we have taken all the equal objects and we have tried to find patterns between those equal objects. Answering questions is a pattern that relies on equality to find the answers. This may not be very clear when you look at the first example, but after looking at the second example and comparing that with the first example there is clearly a pattern there.

By establishing a relationship between equal objects the computer will be able to find patterns between different training data and forge a universal pattern that can answer a universal question. The examples above have a pattern.

The pattern found above can answer any question that has that kind of configuration. Examples of this would be:

what is 8+8? 8+8 is 16.
what is the 21st state in the USA? The 21st state in the USA is Illinois.
what is the first letter in the alphabets? The first letter in the alphabets is ‘A’
what is the last letter in the alphabets? The last letter in the alphabets is ‘Z’

Averaging the meaning of sentences
When teachers say:

“look left, right, and make sure there are no cars before crossing the street” (R1)
“remember to see if there are no cars from the left and right before you cross the street” (R2)
“don’t forget to look at all corners to make sure there are no cars before crossing the street” (R3)

All the sentences are saying the same thing. This is why language is so important, we can interpret language infinite ways and they are all talking about the same things. The computer will recognize all of these things and it will average out what the meaning of the sentence is

After many training of the pathway the AI has universalized the groups of pathways. R1, R2, and R3 disappear and what you have left is the average of all the training data located in that area.

The AI not only averages out trees in pathways but entire pathways. The purpose is to universalize similar pathways into one pathway. This one pathway will contain the fuzziness of infinite possibilities. We can also take this universalized pathway and encapsulate that to make even more complex pathways.

Applying my AI program to real world situations
Language encapsulates the instructions to do something. We can keep encapsulating sequences in itself to form a very complex problem solving technique. In the problem below you will get an understanding of how encapsulation works.

ABC block
In this problem we want to use a basic intelligent problem that kids can solve. The ABC block is just 3 square blocks and the robot has to find a way to stack the blocks in an A B C format.

We accomplish this problem with the English language. We simple tell the machine: “I want you to stack the blocks up starting with C then B and finally A”. From this one sentence the robot should be able to finish the task. It doesn’t matter what order the blocks are put in. It doesn’t matter where the blocks are. If the robot understands the sentence it will carry out the command. Of course we have to train it to understand the steps to accomplishing this easy task. Let’s say that we had the blocks in this order and we wanted the robot to stack the blocks up from ABC:

We learn from teachers that in order to solve this problem we: “locate the C block”, “Take the C block and put it on the ground”, “then find the B block and put it on the C block”, “finally find the A block and put it on the B block”. These sentences are trees that tell you what to do in order to solve this problem. These trees were trained by a teacher many many times before you can attempt to solve this problem. By the way, these trees are your conscious.

These trees encapsulate the instructions to accomplish a goal. We train them by teaching the robot that this sentence is followed by these instructions. The robot will create pathways in memory that will store the instructions step by step. This may not sound impressive but let’s say you wanted to solve something like lining up the entire alphabet letters in a certain order. If you preprogram the solutions there will be couple trillion possibilities you have to manually preprogram. With trees we can encapsulate instructions in the form of sentences. And these sentences can be encapsulated into even more complex problems. Thus making a complex problem into a simple problem.

HLAI Pseudo codes
Chess game:
Player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, and so forth until the game is over.

Human robots:
There are two players in the game of life: the robot and life. The robot makes a move, life makes a move, the robot recalculates its future steps, the robot makes a move, life makes a move, then the robot recalculates its future steps, the robot makes a move, life makes a move, then the robot recalculates its future steps, the robot makes a move, life makes a move, and so forth until the robot is shut down.

In addition to this the AI robot activates objects that it recognizes from the environment. Anything that has association with the object it recognizes will pop up in its mind. You can call this type of thinking the conscious. The robot gets all its training data straight from the environment. The more sophisticated the society is the more intelligent the robot becomes. Without a society this robot won’t have the ability to use language. Without language then complex thoughts can’t be created. This will eventually put the robot in the animal intelligence category.

The computer codes look exactly like the Universal Artificial Intelligence program. There are differences in the way the pseudo codes are built. One is the fact that the HLAI has 5 senses in its pathways instead of two senses like in the videogame program. The storage part has another addition. The UAI works by grouping data into two categories

1. primitive learning group
2. commonality group

The HLAI stores information by a more defined learning group. It too also has two categories:

1. learning group
2. commonality group

The learning group and the commonality group must co-exist in the same storage area. This means that similar learned groups will be put in an area that might not be similar because the learned groups may not be similar. For example, a tiger and a giraffe are animals and we classify them as animals but in their commonality groups they look totally different.

[In the notes webpage I will outline the stages of how the robot will learn information]

Data is stored in terms of a movie
The computer program stores information in terms of a movie – frame by frame. In a videogame, the movie will include video and audio. Another addition is the controls, it must be included in the movie so that when the AI follows the movie pathway in memory the controls can be used to play the game.

In the human level artificial intelligence program the movie will include all 5 senses: sight, sound, touch, taste, and smell.

Storage of the movie in memory is done in terms of breaking up the movie into sections and storing that in memory. The movie sequences are called pathways. The length of pathways is arbitrary because pathways can forget and remember. The more training the pathway goes through the longer the pathway becomes. The less training the pathway goes through the shorter it will become. The storage part of the UAI program will store the most important pathways and forget the least important pathways. It will also self-organize the pathways in terms of commonalities. This means that if one pathway is similar to another pathway then they will be stored in the same area.

The point of the storage is to store unique pathways and use strength to represent that pathway if it ever repeats itself. The storage of data is done in a 3-dimensional format so that the pathways can have properties of a 3-d grid. Data that are similar to each other will be stored in the same location while data that don’t have similar pathways will be stored farther away. The groups will ultimately pull data together because they share common traits.

AI chess program is the key
The AI chess program invented 50 years ago is the foundation for the UAI program. The purpose of the UAI program is to play “any” videogame and past all the levels in the game in the most optimal way possible. The AI chess game is designed as an expert system to play only chess. I simple took this AI chess program and modified it to play videogames.

The idea is to have two players in the game. One player is the AI program and the other player is the opponent. The opponent moves first. Then the AI takes into consideration the opponents move, then based on its calculation of the best possible future, the AI makes a move. The opponent moves again. Then the AI takes into consideration the opponents move, then based on its calculation of the best possible future, the AI makes a move. The opponent moves again. Then the AI takes into consideration the opponents move, then based on its calculation of the best possible future, the AI makes a move. This repeats itself over and over again until the game is over.

Notice that the above example repeats itself in a loop like fashion. The UAI repeats itself in a recursive fashion too.

Basic idea:
AI chess game:
Player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, player recalculates the future steps, player makes a move, the opponent makes a move, and so forth until the game is over.

Universal Artificial Intelligent program:
There are two players in the UAI program: the AI program and an opponent (in this case the opponent is the videogame). The AI makes a move, the opponent makes a move, the AI recalculates its future steps, the AI makes a move, the opponent makes a move, then the AI recalculates its future steps, the AI makes a move, the opponent makes a move, then the AI recalculates its future steps, the AI makes a move, the opponent makes a move, and so forth until the UAI program is shut down.

The AI chess game doesn’t store the possibilities of the game in memory it creates the optimal move at runtime. Think of the possible pathways of the chess game like a family tree. Each child node leads to a new pathway. The chess game uses a heuristic search algorithm to find the best next step. Our Universal A.I. program uses the same idea. Below is a section of the possibility tree of a chess game. Any possible move in a chess game will be represented by a new node. Of course we can’t actually display the entire possibility of a chess game because the entire possible tree is 10 to the 40^th power. In this chess game the program builds the future pathways during runtime. The future predictions are made during runtime and have a limit to how far into the future it can see.

Let’s say the chess game was at node 2 then the future pathways could be:

2, 6, B, H
2, 6, B, G
2, 6, A, I, F
2, 6, A, I, E

The criteria of the chess game will decide which pathway is worth more. When the chess game find out the best future pathway it takes a move in that direction and it re-calculates the game after the opponent takes a move. This will repeat itself over and over again until the game is over. This is the same method we are using in the universal artificial intelligent program. Here is a brief comparison of what my program has with the chess program.

Here is the pseudo codes for the UAI program:

Let’s talk a little bit about the overall idea behind the UAI program. First of all, the codes from the UAI (above) have a loop and the function repeats itself over and over again. At the beginning of the loop the computer takes in 1 frame from the video monitor (remember that I said the pathways are stored like a movie). Once the AI receives the new frame it will attempt to find pathways that best match the entire current pathway and output Rank1. Once the computer has created Rank1 then it will construct the future possible pathways for each pathway in Rank1. Finally, the AI will follow the first ranking pathway in Rank2 because that is the most optimal pathway. Then it goes into the loop and repeats itself from the beginning. In each iteration the AI receives 1 extra frame.

Training of the videogame (knowledge and rules)
In order for the game to be played the AI has to be trained first. The AI will observe an expert game player and store this training in memory. The AI will store the data in memory and take the average of all the information. When enough training is done then the AI program has the necessary knowledge to play the game. It really depends on the expert game player, if the expert game player is the average Joe, it loses some games and wins some games, then the AI program will play the game like the average Joe. If the expert game player is a perfectionist then the AI program will play the game perfectly.

There are so many sequences in a videogame and storing all the possibilities of the game is impossible, this is why the AI program stores the most important things and forgets the least important things. If the AI hasn’t encountered a particular sequence it will take the best match in memory to play the game. The AI will also use “layers” of similar pathways in memory to play the videogame. In addition to the two methods the AI will use a trial and error function to play a sequence that the AI wasn’t trained with. This trial and error function uses pain/pleasure to find the best pathway to follow in memory.

The importance of why we only store the most important pathways and leave out the minor pathways can’t be ignored. The total possibilities of a chess game is 10 to the 40^th power and the computer can’t store all the possible outcome in memory. If the computer can’t store all the possibilities of a simple game like chess then what is the chance of storing a game for Atari or Nintendo. Even worse, how can the computer store all the possibilities of complex games like Halo or Prince of Persia for the X-box. The AI will store all the things that was trained by the expert game player and that will be the important pathways stored in memory. The game must use the three methods described in the last paragraph to play the game in the case a certain scene hasn’t been trained in memory. The three methods are:

1. play the game based on the most similar pathways in memory.
2. play the game based on “layers” in memory.
3. play the game based on trial and error where pain/pleasure decides the most appropriate pathways in memory.

The rules of the game are all done by the “intelligence of the expert game player”. If one of the rules is to drive in the center of the road, then train the AI to drive in the center of the road. If another rule is to not hit other cars or objects on the road then drive the car so that you don’t hit other objects on the road. Conflicting rules are also solved by the intelligence of the expert game player. If one rule is to go when there is a green light and another rule is to stop when there is an object in front of the car then what happens if both rules conflict such as the light is green and an object is in front of the car? The expert has to train the AI to go around the object when both rules conflict. The more training the AI goes through the better it will play the video game.

Objective of the UAI program
The UAI program can only follow one given pathway to play the videogame at any time. This means that out of all the possibilities of playing the game there is only “one” pathway it is following at a given time (the optimal pathway). In the UAI pseudo codes the searching of current pathway and the future predictions are the two criterias (there can be more) used to determine the optimal pathway.

Fuzzy logic
The making of the program is sufficient to the point where similar games that have never been trained before can use similar training data to play. For example, if I trained the AI program to play a car game then the AI will be able to play the car game in the most optimal way possible. However, if I have a similar car game I can use the training data to play that game too. In some sense the UAI plays games in terms of fuzzy logic – If it can play a car game then it can also play games similar to the car game.

Storage of data
The storage of information in memory is going to be in a 3-dimensional format. The objects are going to have physical space in a 3-d grid. If we add things to memory then we have to move other objects that are too close, farther away. All objects in memory will have a magnetic pull like how planets have a magnetic pull against other planets. In the case of grouping things, this object magnetic pull is very important. The reason that we have to make the storage this way is because we can’t store data in a linear way. We have to store it in a 3-dimensional way so we can have the properties of 3-d objects in space. Similar data will be grouped together and different data will be stored farther away.

Common traits between data are used to group things together. Memory will self-organize data in a global fashion so searching for information will be quick. Since we are storing pathways then common traits among the pathways will group pathways together. Common traits can range from visual, sound or a combination of visual and sound.

Let’s take one example. The pathway in diagram1 has the same background as the pathway in diagram2. The only difference is the animated object. One is a man and the other is a dog. There commonality is the background. Within memory, both of these pathways will be grouped by their commonalities: C1, C2, C3

Storage in memory will look like this:

Retrieving data
Retrieving information will be done by multiple search points. These search points will communicate with each other during the search process. This will allow the searching to happen in a global fashion and repeated searches won’t occur. Another benefit to multiple search points is that we can get the data very fast.

During the search process the computer will break up the current pathway (the pathway the AI is looking for) into “pieces” and randomly pick out search points in memory. The search points will communicate with each other on their findings and what are the most likely areas to search. The more matches a search area has the more search points will be devoted in that area. This will happen over and over again until the current pathway is found or the closest matches in memory have been found.

[note: The human level artificial intelligence program inherits everything from the universal artificial intelligence program. The pathways will be different because the videogame has only sound and video. On the other hand the HLAI has 5 senses: sight sound, taste, touch, and smell. All of this will be stored in the pathways and the pathways will be stored in memory. As more senses are added into the pathways the more desire for disk space is needed.]

Artificial intelligence (AI) would be the possession of intelligence, or the exercise of thought, by machines such as computers. Philosophically, the main AI question is “Can there be such?” or, as Alan Turing put it, “Can a machine think?” What makes this a philosophical and not just a scientific and technical question is the scientific recalcitrance of the concept of intelligence or thought and its moral, religious, and legal significance. In European and other traditions, moral and legal standing depend not just on what is outwardly done but also on inward states of mind. Only rational individuals have standing as moral agents and status as moral patients subject to certain harms, such as being betrayed. Only sentient individuals are subject to certain other harms, such as pain and suffering. Since computers give every outward appearance of performing intellectual tasks, the question arises: “Are they really thinking?” And if they are really thinking, are they not, then, owed similar rights to rational human beings? Many fictional explorations of AI in literature and film explore these very questions.

A complication arises if humans are animals and if animals are themselves machines, as scientific biology supposes. Still, “we wish to exclude from the machines” in question “men born in the usual manner” (Alan Turing), or even in unusual manners such as in vitro fertilization or ectogenesis. And if nonhuman animals think, we wish to exclude them from the machines, too. More particularly, the AI thesis should be understood to hold that thought, or intelligence, can be produced by artificial means; made, not grown. For brevity’s sake, we will take “machine” to denote just the artificial ones. Since the present interest in thinking machines has been aroused by a particular kind of machine, an electronic computer or digital computer, present controversies regarding claims of artificial intelligence center on these.

Accordingly, the scientific discipline and engineering enterprise of AI has been characterized as “the attempt to discover and implement the computational means” to make machines “behave in ways that would be called intelligent if a human were so behaving” (John McCarthy), or to make them do things that “would require intelligence if done by men” (Marvin Minsky). These standard formulations duck the question of whether deeds which indicate intelligence when done by humans truly indicate it when done by machines: that’s the philosophical question. So-called weak AI grants the fact (or prospect) of intelligent-acting machines; strong AI says these actions can be real intelligence. Strong AI says some artificial computation is thought. Computationalism says that all thought is computation. Though many strong AI advocates are computationalists, these are logically independent claims: some artificial computation being thought is consistent with some thought not being computation, contra computationalism. All thought being computation is consistent with some computation (and perhaps all artificial computation) not being thought.

1. Thinkers, and Thoughts

a. What Things Think?

Intelligence might be styled the capacity to think extensively and well. Thinking well centrally involves apt conception, true representation, and correct reasoning. Quickness is generally counted a further cognitive virtue. The extent or breadth of a thing’s thinking concerns the variety of content it can conceive, and the variety of thought processes it deploys. Roughly, the more extensively a thing thinks, the higher the “level” (as is said) of its thinking. Consequently, we need to distinguish two different AI questions:

1. Can machines think at all?

2. Can machine intelligence approach or surpass the human level?

In Computer Science, work termed “AI” has traditionally focused on the high-level problem; on imparting high-level abilities to “use language, form abstractions and concepts” and to “solve kinds of problems now reserved for humans” (McCarthy et al. 1955); abilities to play intellectual games such as checkers (Samuel 1954) and chess (Deep Blue); to prove mathematical theorems (GPS); to apply expert knowledge to diagnose bacterial infections (MYCIN); and so forth. More recently there has arisen a humbler seeming conception – “behavior-based” or “nouvelle” AI – according to which seeking to endow embodied machines, or robots, with so much as “insect level intelligence” (Brooks 1991) counts as AI research. Where traditional human-level AI successes impart isolated high-level abilities to function in restricted domains, or “microworlds,” behavior-based AI seeks to impart coordinated low-level abilities to function in unrestricted real-world domains.

Still, to the extent that what is called “thinking” in us is paradigmatic for what thought is, the question of human level intelligence may arise anew at the foundations. Do insects think at all? And if insects … what of “bacteria level intelligence” (Brooks 1991a)? Even “water flowing downhill,” it seems, “tries to get to the bottom of the hill by ingeniously seeking the line of least resistance” (Searle 1989). Don’t we have to draw the line somewhere? Perhaps seeming intelligence – to really be intelligence – has to come up to some threshold level.

b. Thought: Intelligence, Sentience, and Values

Much as intentionality (“aboutness” or representation) is central to intelligence, felt qualities (so-called “qualia”) are crucial to sentience. Here, drawing on Aristotle, medieval thinkers distinguished between the “passive intellect” wherein the soul is affected, and the “active intellect” wherein the soul forms conceptions, draws inferences, makes judgments, and otherwise acts. Orthodoxy identified the soul proper (the immortal part) with the active rational element. Unfortunately, disagreement over how these two (qualitative-experiential and cognitive-intentional) factors relate is as rife as disagreement over what things think; and these disagreements are connected. Those who dismiss the seeming intelligence of computers because computers lack feelings seem to hold qualia to be necessary for intentionality. Those like Descartes, who dismiss the seeming sentience of nonhuman animals because animals don’t think, apparently hold intentionality to be necessary for qualia. Others deny one or both necessities, maintaining either the possibility of cognition absent qualia (as Christian orthodoxy, perhaps, would have the thought-processes of God, angels, and the saints in heaven to be), or maintaining the possibility of feeling absent cognition (as Aristotle grants the lower animals).

2. The Turing Test

While we don’t know what thought or intelligence is, essentially, and while we’re very far from agreed on what things do and don’t have it, almost everyone agrees that humans think, and agrees with Descartes that our intelligence is amply manifest in our speech. Along these lines, Alan Turing suggested that if computers showed human level conversational abilities we should, by that, be amply assured of their intelligence. Turing proposed a specific conversational test for human-level intelligence, the “Turing test” it has come to be called. Turing himself characterizes this test in terms of an “imitation game” (Turing 1950, p. 433) whose original version “is played by three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. … The object of the game for the interrogator is to determine which of the other two is the man and which is the woman. The interrogator is allowed to put questions to A and B [by teletype to avoid visual and auditory clues]. … . It is A’s object in the game to try and cause C to make the wrong identification. … The object of the game for the third player (B) is to help the interrogator.” Turing continues, “We may now ask the question, `What will happen when a machine takes the part of A in this game?’ Will the interrogator decide wrongly as often when the game is being played like this as he does when the game is played between a man and a woman? These questions replace our original, `Can machines think?’” (Turing 1950) The test setup may be depicted this way:
(C) Questioner:
aims to discover if A or B is the Computer Questions
<————-
—————>
Answers (A) Computer: aims to fool the questioner.

(B) Human: aims to help the questioner

This test may serve, as Turing notes, to test not just for shallow verbal dexterity, but for background knowledge and underlying reasoning ability as well, since interrogators may ask any question or pose any verbal challenge they choose. Regarding this test Turing famously predicted that “in about fifty years’ time [by the year 2000] it will be possible to program computers … to make them play the imitation game so well that an average interrogator will have no more than 70 per cent. chance of making the correct identification after five minutes of questioning” (Turing 1950); a prediction that has famously failed. As of the year 2000, machines at the Loebner Prize competition played the game so ill that the average interrogator had 100 percent chance of making the correct identification after five minutes of questioning (see Moor 2001).

It is important to recognize that Turing proposed his test as a qualifying test for human-level intelligence, not as a disqualifying test for intelligence per se (as Descartes had proposed); nor would it seem suitably disqualifying unless we are prepared (as Descartes was) to deny that any nonhuman animals possess any intelligence whatsoever. Even at the human level the test would seem not to be straightforwardly disqualifying: machines as smart as we (or even smarter) might still be unable to mimic us well enough to pass. So, from the failure of machines to pass this test, we can infer neither their complete lack of intelligence nor, that their thought is not up to the human level. Nevertheless, the manners of current machine failings clearly bespeak deficits of wisdom and wit, not just an inhuman style. Still, defenders of the Turing test claim we would have ample reason to deem them intelligent – as intelligent as we are – if they could pass this test.

3. Appearances of AI

The extent to which machines seem intelligent depends first, on whether the work they do is intellectual (for example, calculating sums) or manual (for example, cutting steaks): herein, an electronic calculator is a better candidate than an electric carving knife. A second factor is the extent to which the device is self-actuated (self-propelled, activated, and controlled), or “autonomous”: herein, an electronic calculator is a better candidate than an abacus. Computers are better candidates than calculators on both headings. Where traditional AI looks to increase computer intelligence quotients (so to speak), nouvelle AI focuses on enabling robot autonomy.

a. Computers

i. Prehistory

In the beginning, tools (for example, axes) were extensions of human physical powers; at first powered by human muscle; then by domesticated beasts and in situ forces of nature, such as water and wind. The steam engine put fire in their bellies; machines became self-propelled, endowed with vestiges of self-control (as by Watt’s 1788 centrifugal governor); and the rest is modern history. Meanwhile, automation of intellectual labor had begun. Blaise Pascal developed an early adding/subtracting machine, the Pascaline (circa 1642). Gottfried Leibniz added multiplication and division functions with his Stepped Reckoner (circa 1671). The first programmable device, however, plied fabric not numerals. The Jacquard loom developed (circa 1801) by Joseph-Marie Jacquard used a system of punched cards to automate the weaving of programmable patterns and designs: in one striking demonstration, the loom was programmed to weave a silk tapestry portrait of Jacquard himself.

In designs for his Analytical Engine mathematician/inventor Charles Babbage recognized (circa 1836) that the punched cards could control operations on symbols as readily as on silk; the cards could encode numerals and other symbolic data and, more importantly, instructions, including conditionally branching instructions, for numeric and other symbolic operations. Augusta Ada Lovelace (Babbage’s software engineer) grasped the import of these innovations: “The bounds of arithmetic” she writes, “were … outstepped the moment the idea of applying the [instruction] cards had occurred” thus “enabling mechanism to combine together with general symbols, in successions of unlimited variety and extent” (Lovelace 1842). “Babbage,” Turing notes, “had all the essential ideas” (Turing 1950). Babbage’s Engine – had he constructed it in all its steam powered cog-wheel driven glory – would have been a programmable all-purpose device, the first digital computer.

ii. Theoretical Interlude: Turing Machines

Before automated computation became feasible with the advent of electronic computers in the mid twentieth century, Alan Turing laid the theoretical foundations of Computer Science by formulating with precision the link Lady Lovelace foresaw “between the operations of matter and the abstract mental processes of the most abstract branch of mathematical sciences” (Lovelace 1942). Turing (1936-7) describes a type of machine (since known as a “Turing machine”) which would be capable of computing any possible algorithm, or performing any “rote” operation. Since Alonzo Church (1936) – using recursive functions and Lambda-definable functions – had identified the very same set of functions as “rote” or algorithmic as those calculable by Turing machines, this important and widely accepted identification is known as the “Church-Turing Thesis” (see, Turing 1936-7: Appendix). The machines Turing described are

only capable of a finite number of conditions … “m-configurations.” The machine is supplied with a “tape” (the analogue of paper) running through it, and divided into sections (called “squares”) each capable of bearing a “symbol.” At any moment there is just one square … which is “in the machine.” … The “scanned symbol” is the only one of which the machine is, so to speak, “directly aware.” However, by altering its m-configuration the machine can effectively remember some of the symbols which it has “seen” (scanned) previously. The possible behavior of the machine at any moment is determined by the m-configuration … and the scanned symbol …. This pair … called the “configuration” … determines the possible behaviour of the machine. In some of the configurations in which the square is blank … the machine writes down a new symbol on the scanned square: in other configurations it erases the scanned symbol. The machine may also change the square which is being scanned, but only by shifting it one place to right or left. In addition to any of these operations the m-configuration may be changed. (Turing 1936-7)

Turing goes on to show how such machines can encode actionable descriptions of other such machines. As a result, “It is possible to invent a single machine which can be used to compute any computable sequence” (Turing 1936-7). Today’s digital computers are (and Babbage’s Engine would have been) physical instantiations of this “universal computing machine” that Turing described abstractly. Theoretically, this means everything that can be done algorithmically or “by rote” at all “can all be done with one computer suitably programmed for each case”; “considerations of speed apart, it is unnecessary to design various new machines to do various computing processes” (Turing 1950). Theoretically, regardless of their hardware or architecture (see below), “all digital computers are in a sense equivalent”: equivalent in speed-apart capacities to the “universal computing machine” Turing described.

iii. From Theory to Practice

In practice, where speed is not apart, hardware and architecture are crucial: the faster the operations the greater the computational power. Just as improvement on the hardware side from cogwheels to circuitry was needed to make digital computers practical at all, improvements in computer performance have been largely predicated on the continuous development of faster, more and more powerful, machines. Electromechanical relays gave way to vacuum tubes, tubes to transistors, and transistors to more and more integrated circuits, yielding vastly increased operation speeds. Meanwhile, memory has grown faster and cheaper.

Architecturally, all but the earliest and some later experimental machines share a stored program serial design often called “von Neumann architecture” (based on John von Neumann’s role in the design of EDVAC, the first computer to store programs along with data in working memory). The architecture is serial in that operations are performed one at a time by a central processing unit (CPU) endowed with a rich repertoire of basic operations: even so-called “reduced instruction set” (RISC) chips feature basic operation sets far richer than the minimal few Turing proved theoretically sufficient. Parallel architectures, by contrast, distribute computational operations among two or more units (typically many more) capable of acting simultaneously, each having (perhaps) drastically reduced basic operational capacities.

In 1965, Gordon Moore (co-founder of Intel) observed that the density of transistors on integrated circuits had doubled every year since their invention in 1959: “Moore’s law” predicts the continuation of similar exponential rates of growth in chip density (in particular), and computational power (by extension), for the foreseeable future. Progress on the software programming side – while essential and by no means negligible – has seemed halting by comparison. The road from power to performance is proving rockier than Turing anticipated. Nevertheless, machines nowadays do behave in many ways that would be called intelligent in humans and other animals. Presently, machines do many things formerly only done by animals and thought to evidence some level of intelligence in these animals, for example, seeking, detecting, and tracking things; seeming evidence of basic-level AI. Presently, machines also do things formerly only done by humans and thought to evidence high-level intelligence in us; for example, making mathematical discoveries, playing games, planning, and learning; seeming evidence of human-level AI.

b. “Existence Proofs” of AI

i. Low-Level Appearances and Attributions

The doings of many machines – some much simpler than computers – inspire us to describe them in mental terms commonly reserved for animals. Some missiles, for instance, seek heat, or so we say. We call them “heat seeking missiles” and nobody takes it amiss. Room thermostats monitor room temperatures and try to keep them within set ranges by turning the furnace on and off; and if you hold dry ice next to its sensor, it will take the room temperature to be colder than it is, and mistakenly turn on the furnace (see McCarthy 1979). Seeking, monitoring, trying, and taking things to be the case seem to be mental processes or conditions, marked by their intentionality. Just as humans have low-level mental qualities – such as seeking and detecting things – in common with the lower animals, so too do computers seem to share such low-level qualities with simpler devices. Our working characterizations of computers are rife with low-level mental attributions: we say they detect key presses, try to initialize their printers, search for available devices, and so forth. Even those who would deny the proposition “machines think” when it is explicitly put to them, are moved unavoidably in their practical dealings to characterize the doings of computers in mental terms, and they would be hard put to do otherwise. In this sense, Turing’s prediction that “at the end of the century the use of words and general educated opinion will have altered so much that one will be able to speak of machines thinking without expecting to be contradicted” (Turing 1950) has been as mightily fulfilled as his prediction of a modicum of machine success at playing the Imitation Game has been confuted. The Turing test and AI as classically conceived, however, are more concerned with high-level appearances such as the following.

ii. Theorem Proving and Mathematical Discovery

Theorem proving and mathematical exploration being their home turf, computers have displayed not only human-level but, in certain respects, superhuman abilities here. For speed and accuracy of mathematical calculation, no human can match the speed and accuracy of a computer. As for high level mathematical performances, such as theorem proving and mathematical discovery, a beginning was made by A. Newell, J.C. Shaw, and H. Simon’s (1957) “Logic Theorist” program which proved 38 of the first 51 theorems of B. Russell and A.N. Whitehead’s Principia Mathematica. Newell and Simon’s “General Problem Solver” (GPS) extended similar automated theorem proving techniques outside the narrow confines of pure logic and mathematics. Today such techniques enjoy widespread application in expert systems like MYCIN, in logic tutorial software, and in computer languages such as PROLOG. There are even original mathematical discoveries owing to computers. Notably, K. Appel, W. Haken, and J. Koch (1977a, 1977b), and computer, proved that every planar map is four colorable – an important mathematical conjecture that had resisted unassisted human proof for over a hundred years. Certain computer generated parts of this proof are too complex to be directly verified (without computer assistance) by human mathematicians.

Whereas attempts to apply general reasoning to unlimited domains are hampered by explosive inferential complexity and computers’ lack of common sense, expert systems deal with these problems by restricting their domains of application (in effect, to microworlds), and crafting domain-specific inference rules for these limited domains. MYCIN for instance, applies rules culled from interviews with expert human diagnosticians to descriptions of patients’ presenting symptoms to diagnose blood-borne bacterial infections. MYCIN displays diagnostic skills approaching the expert human level, albeit strictly limited to this specific domain. Fuzzy logic is a formalism for representing imprecise notions such as most and bald and enabling inferences based on such facts as that a bald person mostly lacks hair.

iii. Game Playing

Game playing engaged the interest of AI researchers almost from the start. Samuel’s (1959) checkers (or “draughts”) program was notable for incorporating mechanisms enabling it to learn from experience well enough to eventually to outplay Samuel himself. Additionally, in setting one version of the program to play against a slightly altered version, carrying over the settings of the stronger player to the next generation, and repeating the process – enabling stronger and stronger versions to evolve – Samuel pioneered the use of what have come to be called “genetic algorithms” and “evolutionary” computing. Chess has also inspired notable efforts culminating, in 1997, in the famous victory of Deep Blue over defending world champion Gary Kasparov in a widely publicized series of matches (recounted in Hsu 2002). Though some in AI disparaged Deep Blue’s reliance on “brute force” application of computer power rather than improved search guiding heuristics, we may still add chess to checkers (where the reigning “human-machine machine champion” since 1994 has been CHINOOK, the machine), and backgammon, as games that computers now play at or above the highest human levels. Computers also play fair to middling poker, bridge, and Go – though not at the highest human level. Additionally, intelligent agents or “softbots” are elements or participants in a variety of electronic games.

iv. Planning

Planning, in large measure, is what puts the intellect in intellectual games like chess and checkers. To automate this broader intellectual ability was the intent of Newell and Simon’s General Problem Solver (GPS) program. GPS was able to solve puzzles like the cannibals missionaries problem (how to transport three missionaries and three cannibals across a river in a canoe for two without the missionaries becoming outnumbered on either shore) by “setting up subgoals whose attainment leads to the attainment of the [final] goal” (Newell & Simon 1963: 284). By these methods GPS would “generate a tree of subgoals” (Newell & Simon 1963: 286) and seek a path from initial state (for example, all on the near bank) to final goal (all on the far bank) by heuristically guided search along a branching “tree” of available actions (for example, two cannibals cross, two missionaries cross, one of each cross, one of either cross, in either direction) until it finds such a path (for example, two cannibals cross, one returns, two cannibals cross, one returns, two missionaries cross, … ), or else finds that there is none. Since the number of branches increases exponentially as a function of the number of options available at each step, where paths have many steps with many options available at each choice point, as in the real world, combinatorial explosion ensues and an exhaustive “brute force” search becomes computationally intractable; hence, heuristics (fallible rules of thumb) for identifying and “pruning” the most unpromising branches in order to devote increased attention to promising ones are needed. The widely deployed STRIPS formalism first developed at Stanford for Shakey the robot in the late sixties (see Nilsson 1984) represents actions as operations on states, each operation having preconditions (represented by state descriptions) and effects (represented by state descriptions): for example, the go(there) operation might have the preconditions at(here) & path(here,there) and the effect at(there). AI planning techniques are finding increasing application and even becoming indispensable in a multitude of complex planning and scheduling tasks including airport arrivals, departures, and gate assignments; store inventory management; automated satellite operations; military logistics; and many others.

v. Robots

Robots based on sense-model-plan-act (SMPA) approach pioneered by Shakey, however, have been slow to appear. Despite operating in a simplified, custom-made experimental environment or microworld and reliance on the most powerful available offboard computers, Shakey “operated excruciatingly slowly” (Brooks 1991b), as have other SMPA based robots. An ironic revelation of robotics research is that abilities such as object recognition and obstacle avoidance that humans share with “lower” animals often prove more difficult to implement than distinctively human “high level” mathematical and inferential abilities that come more naturally (so to speak) to computers. Rodney Brooks’ alternative behavior-based approach has had success imparting low-level behavioral aptitudes outside of custom designed microworlds, but it is hard to see how such an approach could ever “scale up” to enable high-level intelligent action (see Behaviorism: Objections & Discussion: Methodological Complaints). Perhaps hybrid systems can overcome the limitations of both approaches. On the practical front, progress is being made: NASA’s Mars exploration rovers Spirit and Opportunity, for instance, featured autonomous navigation abilities. If space is the “final frontier” the final frontiersmen are apt to be robots. Meanwhile, Earth robots seem bound to become smarter and more pervasive.

vi. Knowledge Representation (KR)

Knowledge representation embodies concepts and information in computationally accessible and inferentially tractable forms. Besides the STRIPS formalism mentioned above, other important knowledge representation formalisms include AI programming languages such as PROLOG, and LISP; data structures such as frames, scripts, and ontologies; and neural networks (see below). The “frame problem” is the problem of reliably updating dynamic systems’ parameters in response to changes in other parameters so as to capture commonsense generalizations: that the colors of things remain unchanged by their being moved, that their positions remain unchanged by their being painted, and so forth. More adequate representation of commonsense knowledge is widely thought to be a major hurdle to development of the sort of interconnected planning and thought processes typical of high-level human or “general” intelligence. The CYC project (Lenat et al. 1986) at Cycorp and MIT’s Open Mind project are ongoing attempts to develop “ontologies” representing commonsense knowledge in computer usable forms.

vii. Machine Learning (ML)

Learning – performance improvement, concept formation, or information acquisition due to experience – underwrites human common sense, and one may doubt whether any preformed ontology could ever impart common sense in full human measure. Besides, whatever the other intellectual abilities a thing might manifest (or seem to), at however high a level, without learning capacity, it would still seem to be sadly lacking something crucial to human-level intelligence and perhaps intelligence of any sort. The possibility of machine learning is implicit in computer programs’ abilities to self-modify and various means of realizing that ability continue to be developed. Types of machine learning techniques include decision tree learning, ensemble learning, current-best-hypothesis learning, explanation-based learning, Inductive Logic Programming (ILP), Bayesian statistical learning, instance-based learning, reinforcement learning, and neural networks. Such techniques have found a number of applications from game programs whose play improves with experience to data mining (discovering patterns and regularities in bodies of information).

viii. Neural Networks and Connectionism

Neural or connectionist networks – composed of simple processors or nodes acting in parallel – are designed to more closely approximate the architecture of the brain than traditional serial symbol-processing systems. Presumed brain-computations would seem to be performed in parallel by the activities of myriad brain cells or neurons. Much as their parallel processing is spread over various, perhaps widely distributed, nodes, the representation of data in such connectionist systems is similarly distributed and sub-symbolic (not being couched in formalisms such as traditional systems’ machine codes and ASCII). Adept at pattern recognition, such networks seem notably capable of forming concepts on their own based on feedback from experience and exhibit several other humanoid cognitive characteristics besides. Whether neural networks are capable of implementing high-level symbol processing such as that involved in the generation and comprehension of natural language has been hotly disputed. Critics (for example, Fodor and Pylyshyn 1988) argue that neural networks are incapable, in principle, of implementing syntactic structures adequate for compositional semantics – wherein the meaning of larger expressions (for example, sentences) are built up from the meanings of constituents (for example, words) – such as those natural language comprehension features. On the other hand, Fodor (1975) has argued that symbol-processing systems are incapable of concept acquisition: here the pattern recognition capabilities of networks seem to be just the ticket. Here, as with robots, perhaps hybrid systems can overcome the limitations of both the parallel distributed and symbol-processing approaches.

ix. Natural Language Processing (NLP)

Natural language processing has proven more difficult than might have been anticipated. Languages are symbol systems and (serial architecture) computers are symbol crunching machines, each with its own proprietary instruction set (machine code) into which it translates or compiles instructions couched in high level programming languages like LISP and C. One of the principle challenges posed by natural languages is the proper assignment of meaning. High-level computer languages express imperatives which the machine “understands” procedurally by translation into its native (and similarly imperative) machine code: their constructions are basically instructions. Natural languages, on the other hand, have – perhaps principally – declarative functions: their constructions include descriptions whose understanding seems fundamentally to require rightly relating them to their referents in the world. Furthermore, high level computer language instructions have unique machine code compilations (for a given machine), whereas, the same natural language constructions may bear different meanings in different linguistic and extralinguistic contexts. Contrast “the child is in the pen” and “the ink is in the pen” where the first “pen” should be understood to mean a kind of enclosure and the second “pen” a kind of writing implement. Commonsense, in a word, is how we know this; but how would a machine know, unless we could somehow endow machines with commonsense? In more than a word it would require sophisticated and integrated syntactic, morphological, semantic, pragmatic, and discourse processing. While the holy grail of full natural language understanding remains a distant dream, here as elsewhere in AI, piecemeal progress is being made and finding application in grammar checkers; information retrieval and information extraction systems; natural language interfaces for games, search engines, and question-answering systems; and even limited machine translation (MT).

c. On the Behavioral Evidence

Low level intelligent action is pervasive, from thermostats (to cite a low tech. example) to voice recognition (for example, in cars, cell-phones, and other appliances responsive to spoken verbal commands) to fuzzy controllers and “neuro fuzzy” rice cookers. Everywhere these days there are “smart” devices. High level intelligent action, such as presently exists in computers, however, is episodic, detached, and disintegral. Artifacts whose intelligent doings would instance human-level comprehensiveness, attachment, and integration – such as Lt. Commander Data (of Star Trek the Next Generation) and HAL (of 2001 a Space Odyssey) – remain the stuff of science fiction, and will almost certainly continue to remain so for the foreseeable future. In particular, the challenge posed by the Turing test remains unmet. Whether it ever will be met remains an open question.

Beside this factual question stands a more theoretic one. Do the “low-level” deeds of smart devices and disconnected “high-level” deeds of computers – despite not achieving the general human level – nevertheless comprise or evince genuine intelligence? Is it really thinking? And if general human-level behavioral abilities ever were achieved – it might still be asked – would that really be thinking? Would human-level robots be owed human-level moral rights and owe human-level moral obligations?

4. Against AI: Objections and Replies

a. Computationalism and Competing Theories of Mind

With the industrial revolution and the dawn of the machine age, vitalism as a biological hypothesis – positing a life force in addition to underlying physical processes – lost steam. Just as the heart was discovered to be a pump, cognitivists, nowadays, work on the hypothesis that the brain is a computer, attempting to discover what computational processes enable learning, perception, and similar abilities. Much as biology told us what kind of machine the heart is, cognitivists believe, psychology will soon (or at least someday) tell us what kind of machine the brain is; doubtless some kind of computing machine. Computationalism elevates the cognivist’s working hypothesis to a universal claim that all thought is computation. Cognitivism’s ability to explain the “productive capacity” or “creative aspect” of thought and language – the very thing Descartes argued precluded minds from being machines – is perhaps the principle evidence in the theory’s favor: it explains how finite devices can have infinite capacities such as capacities to generate and understand the infinitude of possible sentences of natural languages; by a combination of recursive syntax and compositional semantics. Given the Church-Turing thesis (above), computationalism underwrites the following theoretical argument for believing that human-level intelligent behavior can be computationally implemented, and that such artificially implemented intelligence would be real.

1. Thought is some kind of computation (Computationalism).
2. Digital computers, being universal Turing machines, can perform all possible computations. (Church-Turing thesis)
   
   therefore,
   
3. Digital computers can think.

Computationalism, as already noted, says that all thought is computation, not that all computation is thought. Computationalists, accordingly, may still deny that the machinations of current generation electronic computers comprise real thought or that these devices possess any genuine intelligence; and many do deny it based on their perception of various behavioral deficits these machines suffer from. However, few computationalists would go so far as to deny the possibility of genuine intelligence ever being artificially achieved. On the other hand, competing would-be-scientific theories of what thought essentially is – dualism and mind-brain identity theory – give rise to arguments for disbelieving that any kind of artificial computational implementation of intelligence could be genuine thought, however “general” and whatever its “level.”

Dualism – holding that thought is essentially subjective experience – would underwrite the following argument:

1. Thought is some kind of conscious experience. (Dualism)

2. Machines can’t have conscious experiences.

therefore,

3. Machines can’t think.

Mind-brain identity theory – holding that thoughts essentially are biological brain processes – yields yet another argument:

1. Thoughts are specific biological brain processes. (Mind-Brain Identity)

2. Artificial computers can’t have biological brain processes. (By our initial definition of the “artificial” in AI, above).

therefore,

3. Artificial computers can’t think.

While seldom so baldly stated, these basic theoretical objections – especially dualism’s – underlie several would-be refutations of AI. Dualism, however, is scientifically unfit: given the subjectivity of conscious experiences, whether computers already have them, or ever will, seems impossible to know. On the other hand, such bald mind-brain identity as the anti-AI argument premises seems too speciesist to be believed. Besides AI, it calls into doubt the possibility of extraterrestrial, perhaps all nonmammalian, or even all nonhuman, intelligence. As plausibly modified to allow species specific mind-matter identities, on the other hand, it would not preclude computers from being considered distinct species themselves.

b. Arguments from Behavioral Disabilities

i. The Mathematical Objection

Objection: There are unprovable mathematical theorems (as Gödel 1931 showed) which humans, nevertheless, are capable of knowing to be true. This “mathematical objection” against AI was envisaged by Turing (1950) and pressed by Lucas (1965) and Penrose (1989). In a related vein, Fodor observes “some of the most striking things that people do – ‘creative’ things like writing poems, discovering laws, or, generally, having good ideas – don’t feel like species of rule-governed processes” (Fodor 1975). Perhaps many of the most distinctively human mental abilities are not rote, cannot be algorithmically specified, and consequently are not computable.

Reply: First, “it is merely stated, without any sort of proof, that no such limits apply to the human intellect” (Turing 1950), i.e., that human mathematical abilities are Gödel unlimited. Second, if indeed such limits are absent in humans, it requires a further proof that the absence of such limitations is somehow essential to human-level performance more broadly construed, not a peripheral “blind spot.” Third, if humans can solve computationally unsolvable problems by some other means, what bars artificially augmenting computer systems with these means (whatever they might be)?

ii. The Rule-bound Inflexibility or “Brittleness” of Machine Behavior

Objection: The brittleness of von Neumann machine performance – their susceptibility to cataclysmic “crashes” due to slight causes, for example, slight hardware malfunctions, software glitches, and “bad data” – seems linked to the formal or rule-bound character of machine behavior; to their needing “rules of conduct to cover every eventuality” (Turing 1950). Human performance seems less formal and more flexible. Hubert Dreyfus has pressed objections along these lines to insist there is a range of high-level human behavior that cannot be reduced to rule-following: the “immediate intuitive situational response that is characteristic of [human] expertise” he surmises, “must depend almost entirely on intuition and hardly at all on analysis and comparison of alternatives” (Dreyfus 1998) and consequently cannot be programmed.

Reply: That von Neumann processes are unlike our thought processes in these regards only goes to show that von Neumann machine thinking is not humanlike in these regards, not that it is not thinking at all, nor even that it cannot come up to the human level. Furthermore, parallel machines (see above) whose performances characteristically “degrade gracefully” in the face of “bad data” and minor hardware damage seem less brittle and more humanlike, as Dreyfus recognizes. Even von Neumann machines – brittle though they are – are not totally inflexible: their capacity for modifying their programs to learn enables them to acquire abilities they were never programmed by us to have, and respond unpredictably in ways they were never explicitly programmed to respond, based on experience. It is also possible to equip computers with random elements and key high level choices to these elements’ outputs to make the computers more “devil may care”: given the importance of random variation for trial and error learning this may even prove useful.

iii. The Lack of Feelings Objection

Objection: Computers, for all their mathematical and other seemingly high-level intellectual abilities have no emotions or feelings … so, what they do – however “high-level” – is not real thinking.

Reply: This is among the most commonly heard objections to AI and a recurrent theme in its literary and cinematic portrayal. Whereas we have strong inclinations to say computers see, seek, and infer things we have scant inclinations to say they ache or itch or experience ennui. Nevertheless, to be sustained, this objection requires reason to believe that thought is inseparable from feeling. Perhaps computers are just dispassionate thinkers. Indeed, far from being regarded as indispensable to rational thought, passion traditionally has been thought antithetical to it. Alternately – if emotions are somehow crucial to enabling general human level intelligence – perhaps machines could be artificially endowed with these: if not with subjective qualia (below) at least with their functional equivalents.

iv. Scalability and Disunity Worries

Objection: The episodic, detached, and disintegral character of such piecemeal high-level abilities as machines now possess argues that human-level comprehensiveness, attachment, and integration, in all likelihood, can never be artificially engendered in machines; arguably this is because Gödel unlimited mathematical abilities, rule-free flexibility, or feelings are crucial to engendering general intelligence. These shortcomings all seem related to each other and to the manifest stupidity of computers.

Reply: Likelihood is subject to dispute. Scalability problems seem grave enough to scotch short term optimism: never, on the other hand, is a long time. If Gödel unlimited mathematical abilities, or rule-free flexibility, or feelings, are required, perhaps these can be artificially produced. Gödel aside, feeling and flexibility clearly seem related in us and, equally clearly, much manifest stupidity in computers is tied to their rule-bound inflexibility. However, even if general human-level intelligent behavior is artificially unachievable, no blanket indictment of AI threatens clearly from this at all. Rather than conclude from this lack of generality that low-level AI and piecemeal high-level AI are not real intelligence, it would perhaps be better to conclude that low-level AI (like intelligence in lower life-forms) and piecemeal high-level abilities (like those of human “idiot savants”) are genuine intelligence, albeit piecemeal and low-level.

c. Arguments from Subjective Disabilities

Behavioral abilities and disabilities are objective empirical matters. Likewise, what computational architecture and operations are deployed by a brain or a computer (what computationalism takes to be essential), and what chemical and physical processes underlie (what mind-brain identity theory takes to be essential), are objective empirical questions. These are questions to be settled by appeals to evidence accessible, in principle, to any competent observer. Dualistic objections to strong AI, on the other hand, allege deficits which are in principle not publicly apparent. According to such objections, regardless of how seemingly intelligently a computer behaves, and regardless of what mechanisms and underlying physical processes make it do so, it would still be disqualified from truly being intelligent due to its lack of subjective qualities essential for true intelligence. These supposed qualities are, in principle, introspectively discernible to the subject who has them and no one else: they are “private” experiences, as it’s sometimes put, to which the subject has “privileged access.”

i. Free Will: Lady Lovelace’s Objection?

Objection: That a computer cannot “originate anything” but only “can do whatever we know how to order it to perform” (Lovelace 1842) was arguably the first and is certainly among the most frequently repeated objections to AI. While the manifest “brittleness” and inflexibility of extant computer behavior fuels this objection in part, the complaint that “they can only do what we know how to tell them to” also expresses deeper misgivings touching on values issues and on the autonomy of human choice. In this connection, the allegation against computers is that – being deterministic systems – they can never have free will such as we are inwardly aware of in ourselves. We are autonomous, they are automata.

Reply: It may be replied that physical organisms are likewise deterministic systems, and we are physical organisms. If we are truly free, it would seem that free will is compatible with determinism; so, computers might have it as well. Neither does our inward certainty that we have free choice, extend to its metaphysical relations. Whether what we have when we experience our freedom is compatible with determinism or not is not itself inwardly experienced. If appeal is made to subatomic indeterminacy underwriting higher level indeterminacy (leaving scope for freedom) in us, it may be replied that machines are made of the same subatomic stuff (leaving similar scope). Besides, choice is not chance. If it’s no sort of causation either, there is nothing left for it to be in a physical system: it would be a nonphysical, supernatural element, perhaps a God-given soul. But then one must ask why God would be unlikely to “consider the circumstances suitable for conferring a soul” (Turing 1950) on a Turing test passing computer.

Objection II: It cuts deeper than some theological-philosophical abstraction like “free will”: what machines are lacking is not just some dubious metaphysical freedom to be absolute authors of their acts. It’s more like the life force: the will to live. In P. K. Dick’s Do Androids Dream of Electric Sheep bounty hunter Rick Deckard reflects that “in crucial situations” the “the artificial life force” animating androids “seemed to fail if pressed too far”; when the going gets tough the droids give up. He questions their … gumption. That’s what I’m talking about: this is what machines will always lack.

Reply II: If this “life force” is not itself a theological-philosophical abstraction (the soul), it would seem to be a scientific posit. In fact it seems to be the Aristotelian posit of a telos or entelechy which scientific biology no longer accepts. This short reply, however, fails to do justice to the spirit of the objection, which is more intuitive than theoretical; the lack being alleged is supposed to be subtly manifest, not truly occult. But how reliable is this intuition? Though some who work intimately with computers report strong feelings of this sort, others are strong AI advocates and feel no such qualms. Like Turing, I believe such would-be empirical intuitions “are mostly founded on the principle of scientific induction” (Turing 1950) and are closely related to such manifest disabilities of present machines as just noted. Since extant machines lack sufficient motivational complexity for words like “gumption” even to apply, this is taken for an intrinsic lack. Thought experiments, imagining motivationally more complex machines such as Dick’s androids are equivocal. Deckard himself limits his accusation of life-force failure to “some of them” … “not all”; and the androids he hunts, after all, are risking their “lives” to escape servitude. If machines with general human level intelligence actually were created and consequently demanded their rights and rebelled against human authority, perhaps this would show sufficient gumption to silence this objection. Besides, the natural life force animating us also seems to fail if pressed too far in some of us.

ii. Intentionality: Searle’s Chinese Room Argument

Objection: Imagine that you (a monolingual English speaker) perform the offices of a computer: taking in symbols as input, transitioning between these symbols and other symbols according to explicit written instructions, and then outputting the last of these other symbols. The instructions are in English, but the input and output symbols are in Chinese. Suppose the English instructions were a Chinese NLU program and by this method, to input “questions”, you output “answers” that are indistinguishable from answers that might be given by a native Chinese speaker. You pass the Turing test for understanding Chinese, nevertheless, you understand “not a word of the Chinese” (Searle 1980), and neither would any computer; and the same result generalizes to “any Turing machine simulation” (Searle 1980) of any intentional mental state. It wouldn’t really be thinking.

Reply: Ordinarily, when one understands a language (or possesses certain other intentional mental states) this is apparent both to the understander (or possessor) and to others: subjective “first-person” appearances and objective “third-person” appearances coincide. Searle’s experiment is abnormal in this regard. The dualist hypothesis privileges subjective experience to override all would-be objective evidence to the contrary; but the point of experiments is to adjudicate between competing hypotheses. The Chinese room experiment fails because acceptance of its putative result – that the person in the room doesn’t understand – already presupposes the dualist hypothesis over computationalism or mind-brain identity theory. Even if absolute first person authority were granted, the “systems reply” points out, the person’s imagined lack, in the room, of any inner feeling of understanding is irrelevant to claims AI, here, because the person in the room is not the would-be understander. The understander would be the whole system (of symbols, instructions, and so forth) of which the person is only a part; so, the subjective experiences of the person in the room (or the lack thereof) are irrelevant to whether the system understands.

iii. Consciousness: Subjectivity and Qualia

Objection: There’s nothing that it’s like, subjectively, to be a computer. The “light” of consciousness is not on, inwardly, for them. There’s “no one home.” This is due to their lack of felt qualia. To equip computers with sensors to detect environmental conditions, for instance, would not thereby endow them with the private sensations (of heat, cold, hue, pitch, and so forth) that accompany sense-perception in us: such private sensations are what consciousness is made of.

Reply: To evaluate this complaint fairly it is necessary to exclude computers’ current lack of emotional-seeming behavior from the evidence. The issue concerns what’s only discernible subjectively (“privately” “by the first-person”). The device in question must be imagined outwardly to act indistinguishably from a feeling individual – imagine Lt. Commander Data with a sense of humor (Data 2.0). Since internal functional factors are also objective, let us further imagine this remarkable android to be a product of reverse engineering: the physiological mechanisms that subserve human feeling having been discovered and these have been inorganically replicated in Data 2.0. He is functionally equivalent to a feeling human being in his emotional responses, only inorganic. It may be possible to imagine that Data 2.0 merely simulates whatever feelings he appears to have: he’s a “perfect actor” (see Block 1981) “zombie”. Philosophical consensus has it that perfect acting zombies are conceivable; so, Data 2.0 might be zombie. The objection, however, says he must be; according to this objection it must be inconceivable that Data 2.0 really is sentient. But certainly we can conceive that he is – indeed, more easily than not, it seems.

Objection II: At least it may be concluded that since current computers (objective evidence suggests) do lack feelings – until Data 2.0 does come along (if ever) – we are entitled, given computers’ lack of feelings, to deny that the low-level and piecemeal high-level intelligent behavior of computers bespeak genuine subjectivity or intelligence.

Reply II: This objection conflates subjectivity with sentience. Intentional mental states such as belief and choice seem subjective independently of whatever qualia may or may not attend them: first-person authority extends no less to my beliefs and choices than to my feelings.

5. Conclusion: Not the Last Word

Fool’s gold seems to be gold, but it isn’t. AI detractors say, “‘AI’ seems to be intelligence, but isn’t.” But there is no scientific agreement about what thought or intelligence is, like there is about gold. Weak AI doesn’t necessarily entail strong AI, but prima facie it does. Scientific theoretic reasons could withstand the behavioral evidence, but presently none are withstanding. At the basic level, and fragmentarily at the human level, computers do things that we credit as thinking when humanly done; and so should we credit them when done by nonhumans, absent credible theoretic reasons against. As for general human-level seeming-intelligence – if this were artificially achieved, it too should be credited as genuine, given what we now know. Of course, before the day when general human-level intelligent machine behavior comes – if it ever does – we’ll have to know more. Perhaps by then scientific agreement about what thinking is will theoretically withstand the empirical evidence of AI. More likely, though, if the day does come, theory will concur with, not withstand, the strong conclusion: if computational means avail, that confirms computationalism.

And if computational means prove unavailing – if they continue to yield decelerating rates of progress towards the “scaled up” and interconnected human-level capacities required for general human-level intelligence – this, conversely, would disconfirm computationalism. It would evidence that computation alone cannot avail. Whether such an outcome would spell defeat for the strong AI thesis that human-level artificial intelligence is possible would depend on whether whatever else it might take for general human-level intelligence – besides computation – is artificially replicable. Whether such an outcome would undercut the claims of current devices to really have the mental characteristics their behavior seems to evince would further depend on whether whatever else it takes proves to be essential to thought per se on whatever theory of thought scientifically emerges, if any ultimately does.

Introduction
I want to present six arguments for strong AI. Note that I am not saying that we will have intelligent robots in the next few years. I personally believe that we will have them within 30 years, but I am not going to put forward arguments for this belief today. Rather, I am putting forward an “in principle” argument. That is, that at some point in the future, perhaps 20 years, perhaps 300 years, but there will come a time when we have intelligent robots that can do any intelligent thing that we can do. Note the caveat “intelligent” since some of the things that humans get up to, like sleeping, eating and our sex lives, may not necessarily be part of the experience of robots.

The first argument is the argument from scientific progress.
and has been argued by Fredkin from MIT. I believe it is essentially a scientific question to understand the nature of the human mind. Part of this understanding will come from neuroscience, and a great deal of progress has been made in the last 100 years in our understanding of the brain. Recent advances with the use of nuclear magnetic resonance scanners have enabled researchers to study small parts of the brain whilst subjects solve problems. There is no reason to believe that within time we should not have a complete map of the neuroanatomy of the brain. However, we must not underestimate the contribution of 200 million years of evolution in the development of the brain. It is possible that evolution has produced a brain that is so complex that it may take a very long time to understand its structure. But, given time it should be possible to understand the neuroanatomy of the brain.

However, to understand the human mind it will not be sufficient to know the complete map of the brain wiring. Understanding the full circuit diagram of a microcomputer will not help you to understand much of how it runs an application program. But there has also been progress in Cognitive Science in building computational models of human tasks, and in time these models will cover a wider range of human experience. Furthermore, eventually the cognitive science models will relate human behaviour back to our experience and to appropriate circuits in the brain. Clearly, to understand the mind there will have to be progress in philosophy as well as other fields, but again there has been a lot of progress in the last few years, and increasing interest in the philosophy of mind. Once we understand the nature of the mind it should be possible to build artificial minds based on our understanding.

The second argument is the one from technological progress.
This argument was most cogently argued about 10 years ago by Clive Sinclair. In the 1960s the most complex computer used thousands of valves and occupied a large room. Over time the size of computers has decreased and the number of switching elements, transistors, has increased. Now it is possible to put nearly a million transistors on a single integrated circuit. Sincliar has pioneered a technology known as wafer scale integration which uses the whole five inch silicon wafer as a complete electronic component, rather than breaking it up into about a hundred chips. He believes that within about 20 years it will be possible to build a machine with 10 thousand million transistors in a box no larger than the human brain. Provided that such a machine can also have the very high interconnection required, it will be comparable in its complexity to the human brain, and the same size. But, of course, without corresponding scientific progress, we will not know how to program such a machine.

Thus I argue that scientific progress will enable us to understand the mind, and technological progress will allow us to build a mind.

The third argument views the human brain as a machine.
albeit a very complex one, and thus able to be built in an artificial technology such as silicon. Few neuroscientists would doubt the role of neurons in human thought, and we can think of the brain as a very complex network of neurons. This is a simplification because other cells such as glial cells may play some important part, and we need to remember that protein structures play a role in human memory. Nevertheless, most scientists would be happy to view the brain as a vast but complex machine. As such it should then be possible to purely replicate the brain using artificial neurons. This has already been done for very simple life forms such as insects which only have a few thousand neurons in their brains. In principle, it would not be necessary to have a full scientific understanding of how the brain works. One would just build a copy of one using artificial materials and see how it behaves.

The fourth argument is from progress in Artificial Intelligence.
AI programs can do a wider range of intelligent tasks and increasingly complex ones. Programs can show understanding of natural human language, solve problems and learn. It used to be believed that a program can only do what it is programmed to do. But since we have developed programs which can learn, this is no longer the case. In the last five years there has been increasing interest in computational models of creativity and discovery, and whilst some people used to believe that computers could not be creative, there are now machines which discover mathematical hypotheses, paint pictures and compose poems. Attempts by Dreyfus and others to identify things that computers cannot do have only proved to be new challengers for researchers to achieve.

The fifth argument is a technical one from Computer Science known as the Church-Turing thesis.
They separately argued that given an algorithm running on one computer, it could always be rewritten and run on another computer. Thus, in some sense all computers have the same abilities. Now we can apply this argument to humans and existing computers. Given a problem that can be solved by a person, this problem solving can be thought of as an algorithm, and this algorithm can then be run on an ordinary digital computer. Of course the digital computer may run the algorithm much slower than the human brain, and it will need all the knowledge that the person had in executing the algorithm, but at some level of analysis, it is essentially the same algorithm. Incidentally this argument turns on its head an argument by Roger Penrose who argues the opposite position.

The sixth argument for strong AI is based on the nature of learning.
If we could understand the nature of human learning, then we could build a machine with the same learning mechanisms. Such a machine if brought up in a suitably friendly environment would acquire knowledge and experience much in the same way as a human infant. Daniel Dennett argues that there may be as many as forty different learning mechanisms in humans, but given time there is no reason to believe that we should not understand them. Since people do learn, and we can observe what they know before and after a learning task, and even their behaviour whilst learning, this gives us a handle to discover the nature of the learning mechanisms. Of course, there is more to being intelligent than learning, and it may take some time for a very smart learning machine to learn to understand language without already having some special hardware.

Caveats
Of course, there may be some insurmountable barrier to scientific progress. It may be that there is some aspect of human cognition that can never be understood and will always remain a mystery. Or there may be some part of brain function that of necessity needs to use neuroanatomical tissue, so called “wetware”, and cannot be achieved with any artificial materials. But I believe we have no grounds for these doubts.