Your analogy simply does not hold here. If you’re having an AI train itself to play chess, then you have adversarial reinforcement learning. The AI plays itself (or another model), and reward metrics tell it how well it’s doing. Chess has the following:
- A very limited set of clearly defined, rigid rules.
- One single end objective: put the other king in checkmate before yours is or, if you can’t, go for a draw.
- Reasonable metrics for how you’re doing and an ability to reasonably predict how you’ll be doing later.
Here’s where generative AI is different: when you’re doing adversarial training with a generative deep learning model, you want one model to be a generator and the other to be a classifier. The classifier should be given some amount of human-made material and some amount of generator-made material and try to distinguish it. The classifier’s goal is to be correct, and the generator’s goal is for the classifier to pick completely randomly (i.e. it just picks on a coin flip). As you train, you gradually get both to be very, very good at their jobs. But you have to have human-made material to train the classifier, and if the classifier doesn’t improve, then the generator never does either.
Imagine teaching a 2nd grader the difference between a horse and a zebra having never shown them either before, and you hold up pictures asking if they contain a horse or a zebra. Except the entire time you just keep holding up pictures of zebras and expecting the child to learn what a horse looks like. That’s what you’re describing for the classifier.
Dude, I’m sorry, I just don’t know how else to tell you “you don’t know what you’re talking about”. I’d refer you to Chapter 20 of Goodfellow et al.'s 2016 book on Deep Learning, but 1) it tragically came out a year before transformer models, and 2) most of it will go over your head without a foundation from many previous chapters. What you’re describing – generative AI training on generative AI ad infinitum – is a death spiral. Literally the entire premise of adversarial training of generative AI is that for the classifier to get better, you need to keep funneling in real material alongside the fake material.
You keep anthropomorphizing with “AI can already understand X”, but that betrays a fundamental misunderstanding of what a deep learning model is: it doesn’t “understand” shit about fuck; it’s an unfathomably complex nonlinear algebraic function that transforms inputs to outputs. To summarize in a word why you’re so wrong: overfitting. This is one of the first things you’ll learn about in a ML class, and it’s what happens when you let a model train on the same data over and over again forever. It’s especially bad for a classifier to be overfitted when it’s pitted against a generator, because a sufficiently complex generator will learn how to outsmart the overfitted classifier and it will find a cozy little local minimum that in reality works like dogshit but outsmarts the classifier which is its only job.
You really, really, really just fundamentally do not understand how a machine learning model works, and that’s okay – it’s a complex tool being presented to people who have no business knowing what a Hessian matrix or a DCT is – but please understand when you’re talking about it that these are extremely advanced and complex statistical models that work on mathematics, not vibes.