There are scaling laws which show LLMs can benefit from an order of magnitude more training data than the current state of the art, suggesting that far beyond G

There are scaling laws which show LLMs can benefit from an order of magnitude more training data than the current state of the art, suggesting that far beyond GPT-4 level performance should be possible in 4GB of RAM with enough training data and compute time. So, kinda?