Best code-first neural-network course: Neural Networks: Zero to Hero. Andrej Karpathy's video course building neural networks from scratch. Start here if you can code and want to understand the machinery behind LLMs.
Best visual transformer explanation: The Illustrated Transformer. Jay Alammar's visual guide to transformer architecture. Use it when attention, embeddings, and transformer blocks still feel abstract.
Best full LLM build path: Build a Large Language Model From Scratch. Sebastian Raschka's book and repository. Use it when you want to implement the pieces rather than only read diagrams.
Learn internals when you need mental models
You do not need to train a model to build useful AI apps, but model internals help when you want to understand tokens, embeddings, attention, training data, loss, fine-tuning, and why models fail in strange ways.
Karpathy is the best code-first route. Jay Alammar is the best visual route. Sebastian Raschka is the best long-form route when you want to build the pieces yourself.
Do not confuse internals with product skill
Model-internals resources are valuable, but they are not a substitute for app design, evals, retrieval, or workflow design. Learn them when they answer a specific question you keep running into.
A good path is to learn enough internals to reason clearly, then return to building and measuring systems. Understanding attention is useful; shipping reliable user-facing AI still needs engineering judgement.
Recommended courses and resources
-
Neural Networks: Zero to Hero
Video course · Andrej Karpathy · Intermediate
You want to understand neural networks and language models from code.
-
Build a Large Language Model From Scratch
GitHub repo / book · Sebastian Raschka · Intermediate
You want to build an LLM step by step.
-
The Illustrated Transformer
Visual guide · Jay Alammar · Beginner to intermediate
Transformer architecture still feels fuzzy.
-
Multi-Vector Embedding Models with Sentence Transformers
Practical guide · Hugging Face · Intermediate to advanced
You want a runnable guide to ColBERT-style late-interaction retrieval, including MaxSim scoring, retrieve-and-rerank, indexing, visual document retrieval, evaluation, and the storage-quality tradeoff versus dense embeddings.
-
Training and Finetuning Multi-Vector Embedding Models with Sentence Transformers
Training guide · Hugging Face · Advanced
You need an end-to-end August 26, 2026 recipe for adapting a late-interaction retriever to your own domain, from datasets and losses through training, evaluation, and index optimization.
