There's no shortage of LLM explainers, but most of them stop at attention. Intuitive AI Academy is the rare curriculum that goes all the way from tokenization to RLHF, with real implementations and real case studies. It's the kind of reference I wish I'd had when I started.
What it covers
- Architecture - tokenization, embeddings, positional encoding, attention, GPT-2 from scratch.
- Pre-training - scaling laws (Chinchilla), data engineering, training infrastructure, LLaMA 3 case study.
- Post-training - SFT, DPO/preference optimization, tool use, safety, Tulu 3 case study.
- Advanced - distillation, LoRA, MoE, optimizers, RL/RLHF.
150,000+ words across 24+ chapters, structured as a technical reference rather than a tutorial.
Why it matters for inference engineers
The topics map directly to what serving teams actually need:
- Attention mechanisms - the thing you're optimizing in every kernel.
- MoE routing - directly relevant to serving DeepSeek, GLM, and the frontier MoE models.
- Distillation - related to speculative decoding draft-model training.
- Scaling laws - the framework for thinking about capacity.
Mental model
Most explainers hand you a map of the city. This one walks you through the streets, block by block, and shows you how the buildings are actually constructed.
The takeaway
Understanding the model is the prerequisite to serving it well. A curriculum that goes from token to RLHF is the fastest way to build that understanding.
Sources
- Intuitive AI Academy: the curriculum, 24+ chapters.
- Scaling Laws (Chinchilla): the compute-optimal training framework the curriculum teaches.
- Tulu 3: the post-training case study.
- LLaMA 2: the pre-training case study.