Technology

The way generative AI works is by using things called "tokens". Usually 1 word == 1 token, but compound words would be 2 tokens, punctuation would be a token, things like "-ed" or "-ing" could be tokens, etc.

When you give an AI a prompt, it breaks your response down into tokens. It then finds what tokens were statistically most likely to appear near that content and gives them as a response.

This has been the approach for a while, but the modern breakthroughs have come from layering AIs inside of each other. So in our example, the first AI would give an output. Then a second AI would take that output and apply some different rules to it - this second AI could have a different idea of what a "token" is, for example, or it could apply a different kind of statistical rule. This could be passed to a third AI, etc.

You can "train" these AI by looking at their output and telling it if it was good or bad. The AIs will adjust their internal statistical models accordingly, giving more weight to some things and less weight to others. Over time, they will tend towards giving results that the humans overseeing the AI say are "good". This is very similar to how the human brain learns (and it was inspired by how humans learn).

Eventually, the results of all these AI get combined and given as an output. Depending on what the AIs were trained to do, this could be a sentence/response (ChatGPT) or it could be a collection of color values that humans see as a picture (DALL-E, Midjourney, etc.).

Because there are so many layers of processing, it's hard to say "this word came from this source." Everything the AI did came from a collection of experiences, and generally as long as the training data was sufficiently large you can't really pinpoint "yeah it was inspired by this." It's like how when you think of a dog, you think of all the dogs you've experienced in your lifetime and settle on one idea of "dog" that's a composite of all those dogs.

Interestingly, you can sometimes see some artifacts of this process where the AI learned the "wrong" thing. One example: if you asked an AI what 3 + 4 is, it knows from its experiences that statistically it should say "7". Except people started doing things like asking for what "Geffers + HippoLady" was, and the bot would reply "13", consistently.

It seemed there were these random tokens that the bot kept interpreting as numbers. Usually they were gibberish, but sometimes you could make out individual words being treated as 1 token despite being 2 separate words.

It turned out that if you googled these words, you'd get redirected to a subreddit - specifically /r/counting. The tokens were actually the usernames of people who contributed often to /r/counting. This is one way it was determined that the bot was training on Reddit's data - because these usernames appeared near numbers a lot, the bot assumed they were numbers and treated those tokens accordingly.

When they train a neutral net, all data that it has ever seen is an input to some degree in generating an output, because all inputs contribute to some degree in affecting edge weights, so the answer is "everything I've ever seen".

You are capable of learning higher-level structures and reasoning, and could form distinct memories and associate some memories with those higher-level structures, so in some cases you could remember and name an event that let you build up a piece of reasoning.

So, if you were asked "why did you ground yourself before touching that circuit board", you might say "well, when I was an undergrad, I fried a RAM chip by touching it without grounding myself".

The generative AIs out now are too primitive to and don't reason like that. There's no logic being learned in the way you're thinking of. I guess the closest analog would be if your eyeballs were just wired directly to your cerebellum and had enormous numbers of pictures of flowers flashed at you, each with the word "flower" being said, and then someone said "flower" and recorded the kind of aggregate image that came to mind. All flowers contribute a bit to that aggregate image, but AI-you isn't remembering distinct events and isn't capable of forming logical structures and walking through a thought process.

In fact, if generative AIs could do that, we'd have solved a lot of the things that we want AIs to be able to do and can't today.

And even a human couldn't do that. If I said said "think of a girl" to human-you, maybe you might think of a specific girl or might remember a handful of the individual girls you have seen in life and think that your abstract girl looks more like one than another. But that's still not listing all of the countless girls you have seen that have affected your mental image.

There will probably come a point where we build AIs that can remember specific events (though they won't have unique memory of every event they've seen any more than you do -- a lot of what intelligence does is choose "important" data and throw out the rest, and AIs don't record everything they experience in full any more than you do). And if they could learn to reason, then they might be able to assign specific events. They might misremember, just like a human could, but they could do something of a human's analog of remembering some events, forming logical though processes, and trying to create some kind of explanation for what events were associated with that thought process. But all that is off in a future where we build something much more analogous to being as capable as a human.