Four papers across thirty days. One per week. Each one sits on a Sunday, after you have already built the thing the paper describes.
That order is deliberate. Reading Attention Is All You Need before you have written a tokenizer is an exercise in nodding at equations. Reading it after Day 3 and Day 5, when you have already felt why the model cannot count letters and why brute-force search does not scale, is a different experience entirely.
Thirty minutes. Not three hours.
You do not read the whole paper. You read the sections listed below, then you write one page in plain English explaining the core idea using an analogy you made up yourself. That page goes in your repo.
If you cannot explain it without the paper open, you have not finished.
| Week | Day | Date | Paper | Read these sections | Write this |
|---|---|---|---|---|---|
| 1 | Day 7 | Sun, Oct 4 | Attention Is All You Need (Vaswani et al., 2017) | Sections 1, 2 and 3 | One page on how Multi-Head Attention works, using a real-world analogy |
| 2 | Day 14 | Sun, Oct 11 | Retrieval-Augmented Generation for Knowledge-Intensive NLP (Lewis et al., 2020) | Abstract and Introduction | One page breaking down the three failure modes of enterprise RAG |
| 3 | Day 21 | Sun, Oct 18 | ReAct: Synergizing Reasoning and Acting in Language Models (Yao et al., 2022) | Abstract and Introduction | One page contrasting pure Chain-of-Thought with ReAct on real tasks |
| 4 | Day 28 | Sun, Oct 25 | Efficient Memory Management for LLM Serving with PagedAttention (Kwon et al., 2023) | Sections 1 to 3 | Covered inside the Capstone live session rather than a standalone write-up |
Weeks 1, 2 and 3 each have a dedicated Read Day. Week 4 does not. The PagedAttention paper is folded into the Day 28 Capstone session instead, because the paper is directly load-bearing for the system design you build that day.
If you want the same shape as the other three weeks, read it on Saturday, October 24 alongside your Docker ship day and write the page anyway. It takes thirty minutes and Question 78 in the interview bank comes straight out of it.
They are the four papers that explain why your builds work.
The Transformer paper explains what the model is. The RAG paper explains why you are retrieving instead of fine-tuning. The ReAct paper explains why agents ground themselves in tool output instead of thinking harder. The PagedAttention paper explains why serving is expensive and what the industry did about it.
Every one of them shows up in interviews. Not as a trivia question, as a reference point that a strong candidate can reach for and a weak one cannot.