← Blogs
Complete Since April 27, 2026 7 days · 18 min
Overview of Modern Nets
I'd used transformers but couldn't explain the architecture end to end without stumbling. A revision pass, piece by piece: tokenizers, positional encoding, attention, BERT vs GPT, and the GPT family.
Coverage
7 of 7 strands covered
Day
1 Tokenization
2 Embeddings and position
3 Attention
4 Architecture
5 Training objectives
6 Model families
7 Tooling
01 The encoder block, BPE, and how tokens are made
02 Multi-head attention and why we use dot products
03 Masking, padding, and BERT vs GPT
04 The transformer architecture, drawn from scratch
05 Pretraining, feedforward, residual connections, and layer norm
06 LLaMA, Mistral, and the road to reasoning models
07 RAG, LangChain, LangGraph, and the LLM tooling landscape
Syllabus
24 of 28 topics written up
- 1 Tokenization 1/1
- 2 Embeddings and position 2/3
- 3 Attention 3/3
- 4 Architecture 4/5
- 5 Training objectives 2/2
- 6 Model families 5/7
- BERT vs GPT
- What LLaMA did
- What Mistral did differently
- Reasoning models
- Multimodals and image generation
- The chain-of-thought paper
- Diffusion models
- 7 Tooling 7/7
Index
Everything written, tangents included, A to Z
L
M
P
R
V