(You can skip this. It’s autobiographical)
Generative AI crept up on me. It shouldn’t have. When I was rather young my world centered around science fiction, and sentient robots were a close second only to space ships that went faster than light.
When I got older I was interested enough in this stuff to devote more than a precious decade of the prime of my life to understanding brains. I got distracted and went in too deep: I had intended to build brains, but I ended up studying them.
I spent too much time on that and at some point I ragequit and went off to do something else and the next thing I know, the damn computers are writing poetry and making sketches.
I should have been paying more attention.
On the computer side of things I had left the field at the point when speech recognition was hard (“How do you deal with accents?”, “What happens where there is background noise?”, “Cepstrum”) and computer vision was begin derided (“In 1960 some genius at MIT thought Conputer Vision was a summer project for an intern. Those idiots.”).
I vaguely remember using a voice based menu on a landline one day and thinking “Hmm, they seem to have made some progress on the recognition front, don’t they?”
Then I saw people using dictation software and then “Siri” but I kept doing what I was doing.
I remember reading, with some amusement, people fiddling around with “hallucinations” generated by forcing noise “backwards” through networks that, when run forwards, did pedestrian things like classify images.
Then I blinked and the world turned into this.
In later posts I will talk a little about where I personally find value in generative A.I. and what I think our world will look like with this genie out of the bottle.
Today, I want to talk about the fact that robots now write poetry and make art. You know, the things we assumed only humans could do.
When the first large language models came on the scene, I stared slack jawed at their outputs. Many people focus on the negatives, such as that they were hallucinations, that they “made up facts.”
These people are missing the forest for the trees.
In 2000 if you told me that a computer would generate a novel coherent sentence, let alone several paragraphs of grammatically correct English I would have called you a liar.
The fact that you can feed a large body of well written English into an algorithm that then creates a matrix which then can be repeatedly cranked to spit out, one word (or even semi-word “token”) by word, whole paragraphs that, at the very least, would get you an A in grammar, says something profound about language, if not intelligence.
It says that language and thought has some kind of regularity that can be learned statistically.
Much effort has been spent in studying language acquisition in humans. One of many observations is that we acquire language by repeated example. Even though we get formal schooling in grammar, strictly speaking, that is a recent phenomenon. For most of history and for most people, we have learned language (vocabulary and grammar) by listening to other people talk and then mimicking them.
So, in a way, it is not surprising that once we got a large enough neural network, we figured out a way to train it with enough examples of sentences and paragraphs to get it to reproduce them with enough variability.
The trickier thing for me is to comprehend how this leads to coherent paragraphs. The simplest explanation is perhaps that “all the thoughts have been thought” and “all the ideas are out there”. Basically, once you have the combined writings of a few million people you have a statistical model of every likely conversation - you have the questions and the answers that come with it.
And this raises something interesting about thinking and what thinking is. It would seem that generative A.I. is proof by demonstration that thinking is the regurgitation (there is probably a more dignified word that is currently eluding me) of thoughts we’ve experienced other express to us with some amount of novelty (randomness) sprinkled in.
On top of that, of course, is that each of us has a slightly different data set we’ve been trained on (our personal experiences differ a bit).
I guess, after typing this all out, I have come to the conclusion that thinking is actually an illusion. Our brain is one giant statistical model, a mirror, of our environment. It is “merely” a device to reflect things that comes into it with a little bit of distortion.
That’s a somewhat disappointing thought, but it has the virtue of being a non-mystical description of what we are.
Of course, there is always the possibility that these generative A.I. things are just shallow imitations of the richness that is a human mind, and once the novelty has worn off, we, I, will start to see the flaws in all that it produces and realize that there is some spark missing that will not be replicated even if we have a neural network a thousand times bigger and a training data set a thousand times larger.
But I have a feeling that is not the case. Generative A.I. has exposed the fact that our brains are statistical modeling machines, and ones that are not even that large.