Learning in public

Ideas in data and AI, explained the way I wish someone had explained them to me.

Intuition first, in an everyday analogy. Then the mechanism, the maths and the code. Then the part most write-ups skip: where it breaks.

Written by Ashish Pandey

Numbers measured, not quoted

Latest writing

kv-cache

inference

serving

A 16× speed-up with identical output, paid for in memory — and the reason your GPU serves fewer users than you expected.

attention

transformers

Every word gets repainted using the words around it. That single idea is what made modern AI work — and what makes long context expensive.

tokenization

A model never sees your words. It sees pieces — and which pieces it gets decides what it costs, how much fits, and how much meaning survives.