Ok, on the risk of sounding exactly like one of the "sentience" guys, but "statistical probabilities of relationships between strings of characters" is a non-answer that makes the whole thing sound simpler than it is without actually providing any explanation how it works. What kind of statistical probabilities exactly?
Right now, research is still ongoing what kind of knowledge is actually captured inside the models, but there are some hints that higher-level, "semantic" knowledge might emerge during training: https://thegradient.pub/othello/
Right now, research is still ongoing what kind of knowledge is actually captured inside the models, but there are some hints that higher-level, "semantic" knowledge might emerge during training: https://thegradient.pub/othello/