A Radically Different Approach to AI: Rethinking Machine Intelligence for Green Mobility
Summary
A researcher reflects on four decades of work toward genuine artificial cognition, rejecting Turing-test trickery in favor of logic humans can follow. The long-dormant, now-public-domain design raises timely questions about how AI should be built as it reshapes transportation and energy systems.
About forty years ago, I decided I would develop artificial intelligence in my spare time. I knew about the Turing test, and I believed that a computer tricking people into thinking it was human did not indicate intelligence. There was a program called ELIZA that did that already, and it was clearly not intelligent at all.
I wanted to produce real AI, a type of thought that might take some effort for a human being to figure out. Maybe I would have to ask the computer how it reached the conclusion it did. But it would do so using logic a human could follow, examining each step. It would be a process a person could examine.
My goal, I believed, would best be reached by first inventing something I called Artificial Cognition. This would be a native way for computers to deal with data in the form of bit-fields, which were native data for computers, as translations of human words, in the same way that computers use IEEE bitwise representations of real numbers. Once this step was done, it would be easy to string words together in sentences, which could be directly translated into and out of human language. And the sentences would tell computers things they could “understand.”
With such translations, a computer could translate a work, such as Shakespeare’s Henry V, into its own language. Then it could translate the text into whatever language we might want. Some subtleties would be lost, but human translators do the same thing.
My idea was that computers should be able to work on ideas natively. At first, I thought this would require a very wide CPU, with registers that might be 256 bits, or even wider. In time, I came to believe that it would be best to stick to CPUs that were in use. As 64-bit CPUs were clearly coming soon, when I was working on this in 1986, I decided to use that data size. I also thought I might have to design a CPU with an additional instruction set, but I never got to the point of needing one.
So my design goal ended up having a word to translate almost any word (or even phrases) of English, occupying 64 bits of memory. And these could be understood and manipulated in standard registers of a standard computer.
I believe that I did achieve that goal, though the project never went into commercial use. A patent on the final design was issued by the US PTO, but that was long ago, and as I understand it, the whole thing is now in the public domain.
My reason for writing about this is that AI is a very contentious issue today. Anthropic has even warned potential stockholders in its IPO documents that its product could have catastrophic effects if things go wrong, possibly to the point of terminating all human life. I think it even points out that a major part of the value of its stock would be lost if there were no human beings left to buy it.
So I started to work on what I called Artificial Cognition in about 1986. Doing this in my spare time was a long slog. I went through a bunch of design iterations, starting, as I said, with 256-bit words. I won’t go through the work that had to be done determining how many designs there were and why each would not work. Each attempt took about a year or more, but eventually I got to the point of having a design in my fifth try.
I will give a simple explanation of the design in this article. Before I do, I will describe what the design could do at the time the patent was applied for, over twenty years ago. To reiterate, it could do these things:
- The words occupy a 64-bit register. I did design a CPU that had the mask bits in registers also, but this hardware is not required for basic operations.
- The words could be operated on using a computer’s logical operations, including OR, AND, XOR, etc. Some operations, such as XOR, are very useful.
Eventually, I defined about 10,000 words for this design. Doing this was a full-time project that took two years, during which I also tested the design. At a point where I had about 2,500 words, I built a simple program that took two inputs of English words, translated them into the computer words, compared them to see that they were valid for compares, and held them for the record. Then it went through the entire vocabulary, taking each word (or set of words) in turn and finding any word that compared to it in precisely the same way the first two words compared.
This might sound like it did far more work than it needed to do. Because most words are not comparable, most compares produced faults and were rejected immediately. So the process did not have to do the millions of compares one might imagine. This was even easier than one might suppose, because whole huge sets of words could be rejected together. So if the input words were a pair like “train” and “airplane,” all words for colors or pieces of clothing could be rejected as sets.
One thing about this that is interesting is how it is done in the computer CPU. With a pair of 64-bit registers loaded, an XOR instruction will result in the difference between the two, or it will produce a fault. The XOR requires a single clock cycle. Then, with the word from the dictionary loaded into a register, a single XOR with the difference produces the word that gives the analogy, or it will produce a fault.
If there is a fault, we go on to the next word in the vocabulary. Otherwise, we have an analogy. In practice, for example, I ran this using original inputs of “car” and “small passenger plane.” The results came back, including the examples following, for “as a car is to a small passenger plane” (please note that each set of words in quotes is a single 64-bit word for the computer):
- As “commuter train” is to “commuter plane”
- As “main battle tank” is to “heavy bomber”
- As “ambulance” is to “air ambulance”
- As “cargo train” is to “cargo plane”
- As “scout car” is to “scout plane”
- As “race car” is to “racing plane”
- As “taxi” is to “small commercial passenger plane”
- As “lion” is to “eagle”
Whoops! What was that?
There were many other results, but the combination of lion and eagle stood out so clearly that I looked specifically to see what the comparison entailed. Why “lion” and “eagle,” but not “tiger” and “owl,” or “house cat” and “hawk”? There were about half a dozen predatory land animals and at least four predatory birds in that small dictionary, but “lion” and “eagle” were the only pair returned by the program. What gives?
It turned out that things had gone precisely the way they should have. The reason that “lion” and “eagle” were chosen while others were not is that the definitions of these two had connotative values as symbols of authority. That information was in their 64-bit definitions. You can do a lot with 64 bits, if you know how.
This work was done in a 32-bit machine, simulating the 64-bit machines just about to come to market. With registers loaded (in a 64-bit machine), each compare would take one clock cycle. So the whole thing, with computers pre-loaded, would take two clocks. In a 2-GHz machine, the set of compares took one billionth of a second. One billionth of a second in a 64-bit machine anyone could buy. One billionth of a second in a Raspberry Pi costing $45.
AI companies are bringing products to market that they warn could end the lives of all human beings. Anthropic says its IPO, which comes out on Thursday, will establish the value of the company at $1.5 trillion (NOTE: with a “T”). But Anthropic is warning that a risk for stockholders is that it could kill all humanity, including you, and your entire family, even if you live in a modest home and get no benefit from the company in any way. And everyone else could go with you. That, I read, is what the IPO papers say, but some people believe it could be very worthwhile to do.
I believe this is not artificial intelligence. It might be artificial neglect. Maybe even artificial willful ignorance. Maybe, with artificial rational thought, we could produce artificial due diligence, but AI is not known for rationality.
I wonder how high the value of Anthropic stock will go if only computers could buy it because there were no people left. Of course, without a Fed made up of humans, they might figure out a way to have infinite money to play with.
How it works
It is all very fuzzy.
The iteration of my invention that works depends on 64 bits being divided into categories, each represented by 8 bits. These are organized into eight supersets (3 bits), each with four groups of subsets (2 bits), with each group having eight basic subsets (3 bits). The total is 8 bits per set value, so there are 256 categories.
Of the eight supersets in each category, one is “agent.” This is anything that does anything. It is divided into four groups of subsets, two of which are “living thing” and “mechanism.” These are divided by type into subsets, so “living thing” includes “animal” and “plant,” among others.
The category “location and time” is similar. This is divided into groups of subsets, one of which is “geography.” Within this are eight subsets, among which are “air,” “land,” “water,” “space,” and four others.
We can see the basic way the vocabulary works by taking these for the most significant 8 bits in a word, the field I call “BAS,” and the next most significant field, which I call “ALT,” because I seem to be unable to come up with a better 3-letter name for it. (Sorry, I do get a bit dull sometimes. Often.)
Using just these two fields, we can see that a car, a train, a taxi, a bus, and a bicycle are all in a group, “land vehicle” (for ALT and BAS). An airplane of any type is an “air vehicle.” So are a blimp and a kite carrying prayers to the clouds asking them to make rain.
Similarly, a lion, tiger, bear, or dog would be a “land animal,” and most birds are “air animal.” Please note that I say “most birds.” They are not divided into biological groups, but according to where they function. Some birds only function on land. Similarly, a bat is an “air animal.” And Melville was right, Moby Dick was a “water animal,” just like any fish.
This broadly divides our dictionary into potential word values using 16 bits. This would limit us to just 65,536 potential words, if they were 16 bits long. Fortunately, we are only a quarter of the way through the 64 bits of our potential word definitions.
The next parts of the definition use 8 bits for each of several fields. A SRC field could represent where something comes from, why it has come into existence, or even what fuels it. For a car, this could be a value that decodes as “chemical,” or a value meaning “electricity.” For a sail-powered car, this field might mean “air.” For just “car,” with no motive force specified, this would be left to the value 0, meaning “any” or “whatever.”
An MTD field represents how it is operated. For a car, this could be decoded by a word meaning “engine” or “motor.” That is, unless the car is powered by sail, in which case neither of these values would work, but something other than 0 would be used, because 0 could mean a superset that engines and motors would match to. In the case of a buggy, this field would say “animal.”
A DST field may mean a destination or purpose. For a car, this could be a value meaning “human being,” for what it carries.
In addition to these, there is a CXT field, which can carry context. In the case of a lion or eagle, this is a value meaning “authority,” which is why they matched to each other but nothing else in the test.
So far, I have described six fields. They provide us with far more potential words than any human language has by a huge factor. That should be enough, but there are other uses for bits. For example, I decided some time ago that there should be eight bits assigned to negation bits, each corresponding to a single field in the word.
Clearly, the values that make up the fields in the word are very fuzzy in their meanings. However, in the design, applying how these are to be interpreted has to be done strictly according to the sets of words involved. Implementing such a design means rules have to be developed according what part of the vocabulary is being worked on.
Altogether, I had a bit over 10,000 words defined about twenty years ago. I believed that these could be built up just a bit for software that could read a dictionary. As it happened, no one was willing to look at the invention to evaluate it, so funding of my invention ceased.
Before I finish up, I should mention that this sort of AI does not do images or other types of jobs that require pattern matching. That isn’t what I set out to do. I always assumed that such things would come from people interested in them. But I wasn’t interested.
I have a problem with the AI being produced commercially today. All of it looks to me like it is based on the same sort of pattern matching that is done by a human mind to some extent. And this is a problem because human beings are not naturally rational creatures. So AI companies are basically building irrational products.
As for Anthropic and the other organizations that claim to have products that threaten human life, that looks like artificial insanity requiring artificial restraint. Or artificial constraint. At least until someone can develop artificial ethics and artificial caution. I don’t think Anthropic has that yet. But I think there is a way to do it.
Featured photo: Fuzzy definition, by Luke Jones, via Unsplash
Comments
No approved comments yet. Be the first to comment.
Leave A Reply