Can someone explain the intuition behind the en-gram idea? I know DeepSeek published a paper about it a few months ago and the Gemma models have a lightweight version of it; but it hasn’t clicked for me yet