This is the page I send people who ask which article to read first. It skips the chronology and groups a small set of recommendations by what you want to learn.

LLM architectures From scratch Reasoning models PyTorch Beginner-friendly Glossary
Best LLM architecture articles

I would begin with the large comparison and keep the gallery open as a visual reference. The other pieces zoom in on attention, DeepSeek, and the workflow I use to read model code and configuration files.

The Big LLM Architecture Comparison My broad comparison of how recent open-weight LLMs arrange attention, experts, normalization, and related components. A Visual Guide to Attention Variants in Modern LLMs Use this when MHA, GQA, MLA, sliding windows, and hybrid stacks start to blur together. LLM Architecture Gallery The architecture figures and specifications in one place, with links back to the longer explanations. My Workflow for Understanding LLM Architectures The config-and-code workflow I use when a model release lacks a complete architecture description. From GPT-2 to gpt-oss A side-by-side bridge from the small GPT-2 template to the much newer gpt-oss design. From DeepSeek V3 to V3.2 DeepSeek's path from V3 to sparse attention in V3.2, including what changed and what stayed.

See all LLM articles →

Best from-scratch implementation articles

If you learn best by coding, start with self-attention or the KV cache. The remaining articles apply the same approach to tokenization, Qwen3, low-rank adaptation, and LLM evaluation.

Understanding and Coding the KV Cache in LLMs from Scratch Build and inspect the cache that prevents each decoding step from recomputing the full attention history. Self-Attention from Scratch Derive the attention equations, then turn them into a small PyTorch implementation. Byte Pair Encoding Tokenizer from Scratch Train a byte-pair tokenizer on a small corpus instead of treating tokenization as a black box. Understanding and Implementing Qwen3 from Scratch Work through a Qwen3-style model in code, component by component. LoRA and DoRA from Scratch Implement both methods and see exactly where their low-rank weight updates differ. The 4 Main Approaches to LLM Evaluation from Scratch Code examples for four evaluation setups, along with the assumptions behind each one.

See all LLM articles →

Best reasoning model articles

The overview is the easiest entry point. From there, the two scaling articles cover what happens at inference time, while the reinforcement-learning piece focuses on how reasoning models are trained.

Understanding Reasoning LLMs My overview of the training and inference methods that turn a base LLM into a reasoning model. Categories of Inference-Time Scaling How sampling, search, verification, and extra test-time tokens fit into the same taxonomy. The State of Reinforcement Learning for LLM Reasoning A closer look at GRPO and the reinforcement-learning recipes behind recent reasoning releases. Inference-Time Compute Scaling Methods The longer treatment of test-time scaling methods and the tradeoffs between them. First Look at Reasoning From Scratch A short preview that connects the theory to the Reasoning from Scratch code.
Best practical PyTorch articles

These are hands-on pieces for learning PyTorch or debugging an inefficient training setup. The one-hour tutorial is the best starting point if you have not used the library before.

PyTorch in One Hour A one-hour refresher that starts with tensors and ends with a complete training loop. Optimizing Memory Usage in PyTorch Ways to cut activation and optimizer memory, with code examples for both LLMs and vision transformers. Make Your PyTorch Models Train Faster A checklist of changes I actually try when a training loop is slower than expected. Mixed-Precision Techniques for LLMs When float16 and bfloat16 help, where they differ, and how they affect model training. Finetuning Falcon LLMs with LoRA and Adapters A worked Falcon finetuning example comparing LoRA with adapter-style updates. Single-GPU LLM Finetuning with Gradient Accumulation Use gradient accumulation to emulate a larger batch when one GPU cannot hold it.

See all PyTorch articles →

Best beginner-friendly articles

These need less background than the implementation articles above. I would pick the reading list for LLM fundamentals, the development overview for the training pipeline, or the coding-agent article for a more applied entry point.

Understanding Large Language Models: A Reading List The paper trail I would follow to build a first mental model of language models. Developing an LLM: Building, Training, Finetuning The full lifecycle in one article, from pretraining through instruction finetuning. New LLM Pre-training and Post-training Paradigms A tour of newer pretraining and post-training recipes, with enough context for non-specialists. Keeping Up With AI Research and News How I filter papers and news without trying to read everything. Getting the Most Out of a Technical Book A few study habits that help turn a technical book into code and notes you can reuse. Components of a Coding Agent A small map of the moving parts in a coding agent, including tools, memory, and repository context.