There are maybe ~20-25 papers that matter. Implement those and you’ve captured ~90% of the alpha behind modern LLMs. Everything else is garnish.