How do you define 'understand'?
There is plenty of AI which learns the rules of games like Alpha Zero.
LLMs might not have the architecture to 'learn', but it also might. If it optimizes all possible moves one chess peace can do (which is not that much to learn) it can easily only 'move' from one game set to another by this type of dictionary.
Understanding a rules-based system (chess) means to be able to learn non-probabilistic rules (an abstraction over the concrete world). Humans are a mix of symbolic and probabilistic learning, allowing them to get a huge boost in performance by admitting rules. It doesn't mean a human will never make an illegal move, but it means a much smaller probability of illegal move based on less training data. Asymptotically, performance from humans and purely probabilistic systems converge. But that also means that in appropriate situations, humans are hugely more data-efficient.
> in appropriate situations, humans are hugely more data-efficient
After spending some years raising my children I gave up the notion that humans are data efficient. It takes a mind numbing amount of training to get them to learn the most basic skills.
You could compare childhood with the training phase of a model. Still think humans are not data-efficient ?
Yes, that is exactly the point I am making. It takes many repetitions (epochs) to teach them anything.
Compared to the amount of data needed to train an even remotely impressive 'AI' model , that is not even AGI and hallucinates on a regular basis ? On the contrary, it seems to me that humans and their children are hugely efficient.
> On the contrary, it seems to me that humans and their children are hugely efficient.
Does a child remotely know as much as ChatGPT? Is it able to reason remotely as well?
I'd say the kid knows more about the world than ChatGPT, yes. For starters, the kid has representations of concepts such as 'blue color' because eyes... ChatGPT can answer difficult questions for sure, but overall I'd say it's much more specialized and limited than a kid. However, I also think that's mostly comparing apples and oranges, and that one's judgement about that is very personal. So, in the end I don't know.
A baby learns to walk and talk in 1 year. Compared to the number of PHDs and compute training these models, the baby is so far ahead in efficiency I marvel way more at their pace.
Neither AlphaZero nor MuZero can learn the rules of chess from an empty chess board and a pile of pieces. There is no objective function so there’s nothing to train upon.
That would be like alien archaeologists of the future finding a chess board and some pieces in a capsule orbiting Mars after the total destruction of Earth and all recorded human thought. The archaeologists could invent their own games to play on the chess board but they’d have no way of ever knowing they were playing chess.
AlphaZero was given the rules of the game, but it figured out how to beat everyone else all by itself!
All by itself, meaning playing against itself...
Interestingly, Bobby Fischer did it in the same way. Maybe AlphaZero also hates chess ? :-)