In 1958 a couple of graduate students concocted a device that analyzed images from a connected camera.
The images were 20 by 20 pixels in resolution.
There was a camera that took the pictures and it was connected to some sort of logic unit that analyzed the pictures – 20 by 20 resulted in 400 units to be analyzed per picture.
Each picture was of a rectangle; it was either on the left or the right side of the picture.
The logic device was a simulated neuron: it had a sensing input - dendrite - an action unit - a neuron - and an output - an axon; it was a computer version of a single neuron, the basic logic unit of the human brain.
So 400 pixels were feeding a neuron to transfer information to an axon which transferred it to a brain.
Each pixel was different in shade.
The brain was asked which side of the full picture a rectangle was on.
It was supposed to figure that out by doing some logical math resulting from the aggregate mathematical information passed from the 400 pixels.
Initially gibberish, rather than accurate answers resulted from this essentially random information.
But the inventors of the device constantly adjusted and refined the pixel information: they trained the brain toward correct answers.
So this thing – by the way, it was called a Perceptron – had a sensor, the camera taking the images, a processor - the logic unit, and some kind of thing that communicated the logic unit’s judgment as to where the rectangle was.
It was designed to perform as does the dendrite/neuron/axon structure of the brain.
As noted above, at first it didn’t perform.
But its creators hung in there.
Most important, as they trained it, they also fed it as much data – pictures – as was possible and it “learned” and it got really good at saying where the rectangle was.
A lot of scientists thought that the Perceptron was a rudimentary first step toward emulating the brain.
Time went on and other sorts of training got fed to Perceptron-like devices: such neural networks were fed masses of labeled data, so they could fairly accurately identify a dog from a rabbit or a cat.
But that was a simple act of using the logic gates of a neural network to use memorized stuff to identify newly encountered stuff.
That was kinda like intelligence, but it was more like the way some human cultures develop thier children into rote memory devices.
Along the way the PC came on the scene and it went into an explosive "Moore's Law" state of exponential expansion of power.
As is always the case with new technology, most of that power got assigned to human entertainment, in this case to the development and playing of ever more sophisticated games.
Those games quickly became so sophisticated that they needed a whole new type of processor working ahead of, but in tandem with the PC's ever-increasingly more powerful processor: games need to do a lot more than one thing at the same time; the unbelievably powerful base PC processor could only do one thing at a time.
When you put those two in tandem you could stack an array of neural networks that could go through labelled data sets so fast that they could tell you almost before you showed them a picture of a camel that they were looking at a camel.
So Super Rote got born.
That has utility, but you probably could get the same thing done with a room full of little rote conditioned human kids.
Somewhere along the line somebody decided to use that massively increased-power neural, human-like, network to do predictive text.
I guess the iPhone Word Cop must have been one of the early products of that research.
In it's simplest form that use of the massive neural network can predict the next word, or the next phrase or ...
All it needs is a lot of text, and some rules for parsing the text.
A recent issue of The Economist goes into useful detail as to what those parsing rules are and how they work.
And they have produced what we now know as Large Language Model Artificial Intelligence.
The unnerving thing is that almost immediately LLMAI began doing more than predicting the next word in some text editing application.
We have even heard about one of them propositioning a New York Times correspondent.
I used one to help me design the cover for a new book I am publishing.
As I begin to digest all that has happened in the last four months concerning LLMAI I am surprised how obvious it should have been that something of the really human intelligence sort has so quickly emerged.
***************************************
I recall having read over the years about the occasional discovery of a fairly young human who in some manner or other has spent its life with almost no external stimulation: family, nature, sound, smell - all the intense stimulation it is that massively besets most of us from birth onward.
Those children, for they have never been very old - it is thought that those that experience that lack of stimulation don't live very long - are essentially ciphers, human vegetables.
On the other hand, history is replete with examples of very young "prodigies" - children already performing at adult high genius level.
They have always come from environments of exposure to high intellectual stimulation - vast waves of information - sometimes even in the womb, in the case of music, and from, in all cases, usually genius parents and genius parents' friends.
So why should we be surprised that GPT 4, which has consumed all the text on the internet, and updates daily the latest increment, is moving beyond the iPhone Cop?
I corrected "freinds" in a pre-publishing version of the cover.
As to how that could have happened: "nobody's perfect".
No comments:
Post a Comment