>>13411105That's true sometimes of something like Stockfish. Not of AlphaZero, which represents a unique training strategy that I'm referencing. The AlphaZero blogs are very interesting reads, feel free to check them out.
>>13411085You said that "artificial intelligence is only intelligent through sheer brute forcing and doesn't just copy to learn," which doesn't really make any sense. For one thing, GANs do learn to copy, but we can even ignore those and go for something cooler in that it is literally exactly what you think is impossible - an actual, somewhat-obscure example of learning rules by copying is DreamCoder. You can feed DreamCoder three lists, and those three lists in reverse. Given the ability to manipulate lists, it will learn in its own coding language the transformation that reverses any arbitrary list. So what you're looking for is actually a dynamic programming method that already exists, and I'm sure its authors are working on generalizing it anyhow. But your fundamental rationale that something must learn by copying is flawed. Unsupervised representation learning methods using transformers - "brute force" memorization, as you'd call it - are functionally indistinguishable from the few-sample dynamic programming task (DreamCoder) if you treat the system as a black box.
