SK / ENGINEERING BLOG · 2026 EDITION
From NLP
to Large Language Models.
A patient, illustrated, start-from-zero guide. We begin with how a computer reads a sentence, build the Transformer piece by piece, and follow it all the way to modern LLMs, reasoning models, agents and the art of measuring them.
Start reading →Contents · 10 chapters
Language, Tokens and Word Vectors
A computer sees text as bytes, not meaning. This chapter starts at zero: what NLP is, the kinds of problems it solves, how raw text is cut into…
CHAPTER 02 · 7 SECTIONS · ~14 MINFrom RNNs to the Transformer
Recurrent networks and LSTMs read text one word at a time. We see why that became a bottleneck, how attention fixed it, and then build the complete…
CHAPTER 03 · 10 SECTIONS · ~39 MINInside Modern Attention: Positions, Normalization and Efficiency
The details that separate the 2017 Transformer from today’s models: multi-head variants (MQA, GQA, MLA), every major positional encoding including…
CHAPTER 04 · 14 SECTIONS · ~13 MINBERT and Encoder-Only Models
How keeping only the encoder half of the Transformer gives a model that understands text bidirectionally: BERT’s architecture, pretraining tasks,…
CHAPTER 05 · 19 SECTIONS · ~32 MINLarge Language Models: Experts, Decoding and Prompting
What makes a language model “large”, how Mixture-of-Experts scales parameters cheaply, how text is actually generated token by token, and how…
CHAPTER 06 · 24 SECTIONS · ~34 MINFast Inference and Training at Scale
The engineering that makes giant models usable: KV caching, PagedAttention, speculative decoding, scaling laws, parallelism, ZeRO, FlashAttention and…
CHAPTER 07 · 32 SECTIONS · ~59 MINFinetuning and Aligning LLMs
From a raw pretrained model to a helpful assistant: supervised finetuning, LoRA and QLoRA, then preference tuning with reward models, PPO, Best-of-N…
CHAPTER 08 · 20 SECTIONS · ~38 MINReasoning Models
How models learn to “think before answering”: chain of thought, test-time scaling, reinforcement learning with verifiable rewards, GRPO, the…
CHAPTER 09 · 29 SECTIONS · ~74 MINRetrieval, Tools and Agents
Connecting a frozen model to the world: retrieval-augmented generation, tool calling, the Model Context Protocol, the ReAct loop, multi-agent systems…
CHAPTER 10 · 35 SECTIONS · ~72 MINEvaluation and the Road Ahead
How to tell whether a model is any good — human ratings, BLEU and ROUGE, LLM judges, factuality, agent evaluation, benchmarks — plus vision…
How to use this blog
Read the chapters in order: each one introduces its own vocabulary, explains the intuition first, then shows the math or mechanism, and finishes with the trade-offs that matter in practice. Equations always come with a plain-English reading of every symbol, so you can skip the math on a first pass and return to it later.
- Chapters 1–2 build the foundations: text, tokens, embeddings, recurrent networks, attention and the Transformer.
- Chapters 3–4 open the Transformer up: modern attention variants, positional encodings, normalization, and encoder-only models such as BERT.
- Chapters 5–7 turn the architecture into LLMs: generation, scaling, efficient inference and training, finetuning and alignment.
- Chapters 8–10 cover reasoning models, retrieval and agents, evaluation, and where the field is heading.