Close Enough 0% ◫Browse AaGlossary
8.1
Part 8 · How we read papers

Paper club

One paper per chapter, read with everything in this book at hand. We say what it claims in our own words, re-measure what we can, and say plainly when something looks off.

~3 min

The first seven parts built a toolkit: embeddings and their geometry, indexes, quantizers, hashing, entropy. This part turns the toolkit around and points it at research papers. Each chapter takes one paper and asks what it tells us about vectors, and whether we believe it.

1Why a paper club

Papers in this field are written to be accepted, not to be understood by the next reader. They use their own notation, report the experiments that worked, and present design choices as inevitable. A paper club reads against that grain. The questions are the ones you would ask a colleague who had just presented the paper:

  • What is the idea, in our words? Strip the paper down to one or two mechanisms and restate them in the language of this book: cosine or MIPS, recall or estimation error, bits per vector, −log⁡2p-\log_2 p.
  • What does it tell us about vectors? The interesting part is rarely the headline number. It is the measurement or the observation in Section 3 that says something about how real vectors are shaped.
  • What did we check ourselves? Wherever a claim can be tested on a laptop or one GPU, we test it. These checks appear in boxes marked Measured, and the scripts are in experiments/.
  • What looks off? Missing baselines, a guarantee that is weaker than it sounds, a number that does not match its own table. We say so, explain why, and keep the doubts separate from the facts.

2How a chapter is told

Each chapter is a story, not a summary. It follows the paper’s argument as a chain of questions, and each step should make you want the next.

  1. A hook. A puzzle or a number that makes the problem feel real, before any notation.
  2. The idea, played with. A toy you can drag, small enough to see the whole mechanism, before the formula that names it.
  3. The catch. Every good idea runs into a wall; the wall is what motivates the next idea.
  4. The real thing. The same picture on real vectors, re-measured by us on an open model.
  5. Raised eyebrows. What we would ask the authors, each with a one-line verdict.
  6. What it suggests. What changes in how we think about embeddings, search or compression.

3The papers

  • MagicPIG: don’t pick the tallest Chen et al., 2024 (ICLR 2025). Fast attention reads only the top-scoring keys, which is like estimating average height from the tallest people. Sample instead, and let SimHash tell you the odds. Touches Parts 1, 5 and 7.