A beginner's learning hub

Learn AI, from scratch

Six guides that take you from "I have no idea how this works" to the real engineering frontier of on-device AI. The first five need no prior AI or math; the last is the advanced deep end.

Suggested path: Edge LLMs → Attention & Transformers → Embeddings → Fine-tuning & LoRA → How Training Works → Advanced Edge Internals. Each is self-contained, has diagrams, and ends with a quick quiz. Roughly 15–20 min each.
The guides
1

How a Language Model Runs on a Phone

start here

The absolute basics (models, tokens, weights), what "edge" means, the one golden rule of speed, then the tricks that make it possible: quantization, the KV cache, hardware backends, speculative decoding, memory-sharing, and serving.

tokens & weightsquantizationKV cacheCPU / GPU / NPU
Open guide 1 →
2

Attention & the Transformer

the core engine

Opens up the mechanism behind every LLM: how words become vectors, how attention lets each word look at the others (Query, Key, Value), multi-head attention, word order, the full Transformer block, and how the next word is chosen.

embeddingsQ / K / Vself-attentionTransformer block
Open guide 2 →
3

Embeddings

most useful, least effort

Turning text into numbers you can compare by meaning. The meaning-map, cosine similarity, the "king − man + woman ≈ queen" trick, semantic search, and a full walkthrough of RAG — letting an AI answer from your own documents.

vectorssimilaritysemantic searchRAG
Open guide 3 →
4

Fine-tuning & LoRA

make it yours

How to specialize a general model without a data center. Prompting vs RAG vs fine-tuning, what full fine-tuning costs, the LoRA shortcut, QLoRA, a when-to-use-which guide, and why tiny swappable adapters are perfect for the edge.

fine-tuningLoRAQLoRAadapters
Open guide 4 →
5

How Training Actually Works

where models come from

The loop under everything else: random weights, the loss (how wrong it is), gradient descent (rolling downhill), backpropagation (assigning blame), epochs and batches, and overfitting vs generalizing — plus how it all connects back to the other four guides.

lossgradient descentbackpropagationoverfitting
Open guide 5 →
6

Advanced Edge Internals

the deep end

The engineering frontier: K-quants and mixed precision beyond Q4_0, squeezing the CPU with weight repacking and big-core threadpools, the reality of NPUs and QNN, and continuous batching built on paged-KV — mapped to EdgeLM's honest built-vs-in-progress status.

K-quantsSIMD / i8mmQNN / NPUcontinuous batching
Open guide 6 →
You've reached the end of the set. These six guides cover the whole arc — from what a token is to the edge's real engineering frontier. Want to go further? Good next steps outside this hub: reading real llama.cpp / LiteRT-LM internals, or trying quantization and a small fine-tune yourself.