My notes on AI — rewritten from first principles, in code, with interactive demos.
From counting words to training a neural language model.
NLPThe transformer, one concept at a time.
ArchitecturesOne codebase, model by model. Each stop shows only what changed.
Pre-TrainingPre-training and fine-tuning at scale.
InferenceDecoding, inference, and serving.
Deep LearningWhat DeepSeek invented: MLA, MoE, multi-token prediction, FP8.
RL + ReasoningRL, alignment, and reasoning — starting with no extra training.
NLPUsing and trusting LLMs: evaluation, RAG, agents, and safety.
SearchFrom the inverted index to agents that search for themselves.
ArchitecturesThe nine architectures from The Evolving Transformer, each as a change from the stop before it.
ArchitecturesThe five diagrams we have for The Evolving Transformer.
ResearchEvery paper cited here — searchable, tagged, one click away.
InterviewInterview question sets and Python practice. Reveal, track, quiz yourself.
PythonWrite and run Python in your browser. No install.
ArchivesEarly notes, 2018–2024. The path that led here.
CourseSix months. Ten phases. Applied AI research, end to end.