FROM FIRST PRINCIPLES TO SYSTEMS
How language models actually work
From first principles to efficient inference in 106 complete lessons per language. Every claim cited, every lab grounded in real numbers, nothing behind a login.
Systems reference for CPU inference with weight streaming: Kimi K3 in C.
- 10
- published tracks
- 106
- lessons
- 2
- EN + PT-BR
- INTUITION
- MECHANISM
- EVIDENCE
- SYSTEMS
SECTION 01 · TOKEN TO SYSTEM
LEARNING PROTOCOL
Every lesson is a complete investigation
You do not just consume content. You follow the mechanism, test the idea, and prove you can explain it.
- 01Concept
The actual mechanism, with no hand-waving.
- 02Analogy
The same idea in plain language.
- 03Lab
Real numbers, computed live.
- 04Teach-back
You rebuild the idea in your own words.
- 05Quiz
The right answer comes with the reason.
A lesson finishes only when your explanation has substance and the quiz is right. Your writing stays in this browser.
CURRICULUM SURVEY
The complete path, layer by layer
10 published tracks form one continuous descent, from foundations to the silicon that runs the model.
- 01Orientation & FoundationsBuild an accurate mental model of language models and the mathematics needed for everything that follows.Enter track↘
- 02From Text to TensorsSee how text becomes tokens, vectors, and numerical structures a neural network can process.Enter track↘
- 03Neural Network FundamentalsLearn how neural networks represent functions, measure error, propagate gradients, and improve through optimization.Enter track↘
- 04Sequence ModelsFollow the path from recurrent state to encoder-decoder models and attention, the immediate ancestors of the Transformer.Enter track↘
- 05The TransformerAssemble the architecture behind modern LLMs, beginning with self-attention and its query, key, and value operations.Enter track↘
- 06Pretraining at ScaleDesign the objectives, data, optimization, parallelism, and cost model behind a serious language-model pretraining run.Enter track↘
- 07Post-training & AlignmentTurn pretrained predictors into useful assistants through instruction tuning, preference learning, reinforcement learning, and distillation.Enter track↘
- 08Inference & EfficiencyUnderstand how trained models generate text and how decoding choices trade off diversity, coherence, latency, and cost.Enter track↘
- 09Building & Serving StacksMaster the software that actually runs Qwen3.8-27B — PyTorch, Transformers, vLLM, SGLang, TensorRT-LLM, llama.cpp, Ollama, MLX, and Modular MAX — and learn which one each situation calls for.Enter track↘
- 10Hardware & InfrastructureUnderstand the silicon beneath Qwen3.8-27B — NVIDIA, AMD, Apple, Qualcomm, and Cerebras — and the bandwidth, roofline, and cost arithmetic that decides what runs where, and for how much.Enter track↘
106 complete lessons per language across 10 tracks, with EN + PT-BR parity.
DEPTH WITHOUT SURVEILLANCE
No account. No paywall. No tracking.
Your progress belongs to your device. The code and curriculum remain open.
Start the descent