AI Needs To Feed On Fresh Meat, Or It Dies

In Asimov’s original Foundation, a novel set amidst a fading and decaying Empire, one saturated in hubris and self-esteem, we meet Lord Dorwin, a celebrated scholar from the Galactic Empire, and his encounter with a man from the Foundation, a group that has fled to the far side of the universe to preserve the Good and True.

Pirenne heard Lord Dorwin’s idea of scientific research. Lord Dorwin thought the way to be a good archaeologist was to read all the books on the subject—written by men who were dead for centuries. He thought that the way to solve archaeological puzzles was to weigh the competing authorities…Don’t you see there’s something wrong with that?

I trust you do see.

But maybe you don’t see what this has to do with AI.

The largest success of AI—semi-statistical models—is in making fake pictures, and even fake voices, which seem real. Or real enough to pass casual inspection. So vivid are these creations that there has been much speculation whether real people will be needed in the future to, say, act in movies or television. Couldn’t AI just do it all?

No.

Not in pictures or in text. The real things—real people, genuine texts, real-life observations—will always be needed.

AI fits it models using observations taken from Reality. It can only ever approximate it, which is to say, only ever model the causes of things. Suppose you took away the new observations used to fit the models. It can keep all the old ones, but no new ones can be taken.

But what if we used AI output as new observations? Could these make up for the food AI needs to feed on?

Here is a portion of the text of the story Family Fun, which I believe is used in the literature portion of the final exam for bachelor’s degrees at Harvard for Women’s Studies students:

See It Go

“Look,” said Dick. “See it go. See it go up.”

Jane said, “Oh, look! See it go. See it go up.”

“Up, up,” said Sally. “Go up, up, up.”

Down it comes.

“Run, Dick. We can find it.”

“See me run,” said Sally. “See Spot run. Oh, oh! This is fun.”

A dirt-simple version of an AI language model works like this. We first gather a “corpus”, or a body of literature with which to “train” our model. And, as I tire of repeating, AI is only a model. This corpus is our observations.

Our “training” will consist of walking through every unique word and building a table of the chance each other word follows it. For instance, it looks like the most common word after up is (again) up. And, for example, it never follows up.

In a larger model we’d do the same for sentence length and punctuation, building in some hard-coded rules of English grammar (“All quotation marks need to be closed”, etc.). And we’d consider more than just the word that came immediately after each word, and also look at that chance of other words “downstream” from the current one. We don’t need any of that complexity here since it’s not needed to make our point. The complexity of the AI model is not relevant.

After training, we can generate “new” AI texts, just like Google’s Gemini. Why not generated some two-word sentences?

We first pick a word at “random” from this “corpus”. We next pick a word by the probability it has coming after it. Suppose our first word was up. The word with the highest chance of coming after is also up, so we have a good chance of generating that as the second word. If so, our two-word AI-generated sentence would be “Up up”.

Others might be “Up Jane”, “Up down” or “Up said.” If the first word was Dick, the two word sentences would be “Dick see”, “Dick we”. For see we get “See it”, “See me”, “See Spot.” Most of these are passable sentences in English, as long as we accept slang.

Suppose folks generated lots of these sentences. The temptation to use these as if they were genuine new essays would be overwhelming, for some. As we have already seen with other AI-generated content (students using them for essays, etc.). This some of these generations would end up on Harvard student essays, others in blog posts, in Substacks and places like that. Some would quote the text while acknowledging it was AI, and others would simply us it.

Since we got our original corpus from “scrapping” sources just like this, the result is that some of the generated text would end up back in a new training corpus!

The new corpus would consist of the old one plus the new text, which was model output. We would then fit our New & Improved AI using the augmented corpus, which included the output from the first model.

I hope it is easy to see that if we continued this process enough times, the resulting generated sentences would all look something like “Up up! Up up up up. Up. Up up up Spot.”

This would in essence be models of models of models of models of, etc. etc., of models of the original corpus.

The same exact thing would happen if we were to feed the results of AI-generated images back into the training sets of future AI models. It’s more of a danger here because AI (model)-generated images are much more common than model-generated texts. So these would more easily be sucked up into new training datasets than text. If this dangerous feedback wasn’t closely monitored, pretty soon all generated images would be four-fingered young Asian women in computer-game armor. And very, very bland looking.

Models, which is to say AI, can only manipulate what it is given using the rules set by the coders. Real life will always be needed, because models have no hope of capturing the full complexity of Reality.

Incidentally, sharp readers will recognize we have seen elsewhere the nonsense that results of models feeding other models which feed other models, and so on.

Subscribe or donate to support this site and its wholly independent host using credit card click here. Or use the paid subscription at Substack. Cash App: $WilliamMBriggs. For Zelle, use my email: matt@wmbriggs.com, and please include yours so I know who to thank.