AI Grimoire

Shelves

Categories

Seven shelves, ordered roughly the way a forward pass is built and then taken apart again. Each divides into branches — 34 of them in all — and an entry lives on exactly one branch. Cross-references do the rest.

Contents

01
Architectures
attention variants, position schemes, normalisation, sparse blocks
54 entries · 2 stubs
02
Training
optimisers, schedules, sharding, RL objectives, stability failures
40 entries · 1 stub
03
Fine-tuning
PEFT families, adapter merging, data curation, catastrophic drift
5 entries · 1 stub
04
Inference
kernels, KV-cache layouts, quantisation, speculative decoding
8 entries · 1 stub
05
Interpretability
circuits, SAEs, causal interventions, probing, feature geometry
9 entries · 2 stubs
06
Adversarial ML
evasion, extraction, poisoning, privacy leakage, agent hijacks
6 entries
07
Evaluation
metrics, preference aggregation, contamination, statistical power
5 entries
08
Miscellany
numerics, tensor conventions, the things every other shelf assumes
3 entries
09
Tooling
what the software is for, when to reach for it, and what it hides
13 entries