Musings of a Restless Mind

The two most frequent and equally dangerous responses to so-called AI are fear and envy.

1. AI will take over the world and kill us all!
2. I need more of that AI right now, or we’re all going to be losers, dead losers.

First let’s dispense with the bear in the convenience store. AI is really, as we in the biz like to say, NBI (Nothing But Initials). There’s nothing either artificial or intelligent about what most of us are talking about. Sure Elon Musk uses the term correctly when he fear mongers about AI, but what he’s talking about is a lot farther off than his other project, you know, the one about putting a colony on Mars. There is no soul in the machine, and you can’t put yours there any more than you can grow one all on its own in Victor Frankenstein’s lab. Ain’t gonna’ happen, Bubba! But we don’t need to have that discussion now, because Artificial Intelligence is not what we’re all talking about.

First, you need to understand some of the basics. For that I refer you to my brief article explaining the concepts. That notes that the 3 pillars of Big Data (now usually, but erroneously called AI) are machine learning, ontologies, and statistical investigation. Now that we know the concepts, here’s the important part. ChatGPT, and any Big Data tool that creates images from text, are all about machine learning but with a new technique. Their creators attempt to substitute mathematical models for ontology creation via Large Language Models (LLM). They do this in ways mostly incomprehensible to even them. In their arrogance, they also cavalierly dispense with a human-labeled training data set. “Hey, this is math. We don’t need no stinking human validation. We have the almighty data set.”

Now some specifics about what’s going on that’s relatively new.

Most of us are familiar with latitude and longitude that can pinpoint your whereabouts on the surface of the earth in 2 dimensions using what mathematician’s call 2 vectors, the good old Cartesian coordinates of x and y. If you had high school geometry or physics, they introduced you to a third dimension, known as height, elevation, or altitude, usually simply called z ’cause mathematicians be like that, and don’t like to spell things out, especially for things that have 3 different words for it. Just give it a letter, and act like it means all of those things. Ambiguity gets in their way. The thing is though, those aren’t 3 terms for exactly the same thing. The mathematicians just waved their magic wand and told you those 3 are the same for their purposes, but that’s pretend. Those differences didn’t really go away just because it made the math work to ignore them.

Next, mathematicians don’t limit themselves to the 3 dimensions we can comprehend. Theoretically you can have as many dimensions (and the vectors measuring them) as you want. One of my wife’s math profs. once quipped, “You could consider God an infinite vector space.” Some LLMs boast of a thousand vectors that define where a word is placed in this incomprehensible thousand-dimensional multitude of madness. The math says cat and toy are much closer together than cat and president, so let’s go with cat toy rather than cat president, right? Anyway that’s the concept, and most of the experts don’t actually understand it any better than that because, who can keep track of a thousand value coordinate tag much less billions of items, each with these thousand vector tags.

“So, what does this all mean to me,” you ask? You may notice that your brain has wandered while I’ve hit you with all that technical mumbo jumbo. That’s the purposeful magician’s distraction, so you don’t see the trick. Remember how, in the article I linked to above, I told you the importance of a human-labeled training data set and a similar human-labeled test data set? That’s a lot of work. That’s because I’m interested in finding the truth. The purveyors of the new, “this time we mean it”, machine learning that they call AI (because AI is not trademarked and can mean anything they want it to mean and that they can make you believe it to mean) don’t actually care about the truth or the facts. With another mathemagician’s trick, they just want to pass the Turing Test aptly named The Imitation Game by Turing. If they can make you believe it’s real, then they win! Of course that varies with the sophistication of “you” the observer.

In the art realm, non-profit organizations boast of their data set of almost 6 billion image-text pairs. Here’s the important question, “Who made sure that text accurately matched the image?” The inevitable answer is, “We just took it at face value, by whoever posted it on the web.” Yeah, that always works. “But you, the pranksters, fraudsters, and idiots are overwhelmed by the innocent posters,” they explain. Maybe, but do you know how the web works? There’s hundreds of businesses that specialize in search engine optimization, by which they mean, tagging your posts with whatever gets the most attention. Pure survival by joining the biggest herd. As Yossarian said to the classic, ‘suppose everyone thought the same way you do’ question, “Then I’d be a damn fool to think any different.” They’ve created a self-reinforcing feedback loop. Truth, facts, wisdom are all irrelevant, or, as the purveyors of such nonsense tell themselves, are taken care of by “the wisdom of crowds”.

Where does all this lead? Here’s an example:

Simon Fraser University is a well-funded public research university in Canada (i.e., not some fringe place funded by the Koch brothers, George Soros, or Glenn Beck). Their Discourse Processing Lab put up a fake news detection web app at http://fakenews.research.sfu.ca. If you would like to check it out, try following the link and typing:

Nancy Pelosi bathes in the blood of virgins.

It will return:

“According to our model, text is mostly based on FALSE INFORMATION.”

Change the first two words so the text reads:

Donald Trump bathes in the blood of virgins

and you get this response:

“According to our model, text is mostly based on FACTS.”

Whatever you might say about Donald Trump, I see no reason to believe that facts support that he bathes in the blood of virgins even if you think he would happily do that. Do they have receipts? Where does he get them all? That pair of responses is just sad. Why would one trust the judgement of such an AI app?

For those interested in art, what happens when you just have billions of images trained by the tags anonymous folks put on them? You ask for a pretty superheroine and get a bunch of images with women with 40DD breast sizes. You tell the app that you want a picture of a pretty woman, and you get an image from a database with 100 images of Venus de Milo and 100,000 images of Kim Kardashian. Maybe you get a pretty woman with three arms, or 7 fingers on each hand. Why? Because this is anatomy without ontology. Did the creators of the data set ask sketch artists and anthropologists what constitutes a pretty face, or what human anatomy really consists of? No, they didn’t. Machine learning without ontologies or human validation, never gets beyond the learning level of a brillaint, autistic 3 year-old. Most things may be right, but many things are wildly wrong. The 3 year-old doesn’t know what’s right or what’s true, he just knows one thing seems to go with the other. This is “training” on the level of dog training, learning by association and only by association.

And that’s only the non-malicious tagging. What happens if some group of trolls post tens of thousands of images of Hillary Clinton with a dog’s face pasted over her own? Chances are your so-called AI tool will begin to think Hillary Clinton looks like a dog. A human has to get involved to remove those fraudulent images from the data set.

So to sum up, there is no AI. AI is still just a wish a science fiction writer’s heart makes. There’s no I, Robot. There’s no HAL 9000. What we currently have are giant, uncurated data sets sorted by machine learning algorithms that even their own “architects” frequently don’t understand. Yes, their results can be impressive in the same way a stage magician’s tricks can be, and for the same reason. Can they do impressive things? Yes, if you use them for what they’re good at, like asking for an image of a spooky cemetery, or a subject that imitates a well-known artists style. Just don’t fool yourself into thinking they are intelligent or can make decisions, or find new cures for cancer. They can assist in all those tasks if you know what you’re doing with them. But they’re just doing the research faster than you can possibly do. They can narrow possibilities, but then it’s you who has to do the real research.

You may not be the smartest kid on the block, but you’re still smarter than a 3 year-old with a brain filled with potentially connected facts whose connection he doesn’t actually understand. Swear at “AI” like a rioting teen-ager, or submit to “AI” like an abject servant, and you’ll look like a fool. Act like an adult human with a fun, useful tool, and you’ll potentially get amazing things from our modern machine learning tools.