Course · 5 modules · 8 lessons · 15 min
Advanced LLM Concepts
A second-volume tour of the techniques pushing large language models forward — advanced training, modern inference and serving, retrieval and embeddings, alignment, and adversarial robustness.
The course at a glance
1 module · click any tile to jump to its lessons.
All lessons
№ 01Advanced Training
01Byte-Latent TransformersBLT replaces tokenization entirely — it operates directly on raw UTF-8 bytes, dynamically grouping them into variable-length patches based on local entropy.→02Catastrophic ForgettingCatastrophic forgetting is the abrupt loss of previously learned capability that happens when a neural network is trained on a new task or domain.→03Curriculum LearningCurriculum learning presents training examples in a meaningful order — usually easy to hard — instead of randomly, and modern LLM data mixing is its most consequential form.→04Gradient CheckpointingGradient checkpointing trades compute for memory by storing activations only at selected layers and recomputing the rest during backpropagation.→05GrokkingGrokking is the eerie phenomenon where a model first memorizes its training data with poor validation accuracy, then — long after training loss has plateaued — suddenly generalizes.→06In-Context LearningIn-context learning is the surprising ability of large language models to acquire a new task from a handful of examples shown in the prompt, without any parameter updates.→07Model CollapseModel collapse is the irreversible loss of distribution tails that occurs when AI models are trained, generation after generation, on data produced by other AI models.→08Multi-Token PredictionMulti-token prediction trains the model to predict several future tokens at once through parallel prediction heads, producing richer representations and faster inference.→