It's not so surprising to me. It's like how Markov chains get better at passing for human the more N-grams they memorize. larger models will continue getting marginally better at predicting the distribution (human language.) but that doesn't translate into improved intelligence.