NVIDIA · NCA-GENL
The weights thread
Where knowledge actually lives and how it gets changed. Opens in M01 with 'knowledge lives in parameters, not in a lookup table,' runs through embeddings and transformer internals as the shape that knowledge takes, and closes in M11 with RLHF — the only mechanism in the stack that moves human judgment into the weights.
T2 · 40 lessons across 5 modules
- M0101-01Next-token prediction: what a language model is trained to do
- M0101-02LLM parameters: what they are and where knowledge is stored
- M0101-03Tensor shapes in transformers: batch, sequence, and hidden size
- M0101-04Vectors, dot products, and cosine similarity
- M0101-05Loss functions and cross-entropy explained
- M0101-06Gradient descent and backpropagation
- M0101-07Train, validation, and test splits
- M0101-08How to build an evaluation set for an LLM project
- M0303-01What text embeddings are: learned dense vectors
- M0303-02Token embeddings vs sentence and document embeddings
- M0303-03How to choose an embedding model
- M0303-04How to test retrieval quality by hand
- M0303-05Vector arithmetic and word analogies (word2vec)
- M0404-01Self-attention and why context length costs quadratically
- M0404-02Positional encoding: how transformers represent word order
- M0404-03Encoder-only vs decoder-only vs encoder-decoder models (BERT, GPT, T5)
- M0404-04Autoregressive generation: how an LLM produces text
- M0404-05Temperature, top-k, top-p, and greedy decoding
- M0404-06The context window: what it is and how to budget it
- M0707-01Sparse retrieval and BM25 keyword search
- M0707-02Dense retrieval with embeddings
- M0707-03Limits of embedding search: negation, recency, and authority
- M0707-04Vector databases and ANN indexes (HNSW, IVF)
- M0707-05Access control and permissions in RAG retrieval
- M0707-06Hybrid search: combining keyword and vector retrieval
- M0707-07Reranking with a cross-encoder
- M0707-08Assembling context: chunk order and the lost-in-the-middle problem
- M0707-09The complete RAG pipeline, stage by stage
- M0707-10Debugging RAG: retrieval failure versus generation failure
- M0707-11Grounding, citations, and letting a model say I don't know
- M0707-12When RAG is the wrong tool
- M1111-01Pretraining vs continued pretraining vs instruction tuning
- M1111-02Supervised fine-tuning (SFT): what it can and cannot change
- M1111-03Catastrophic forgetting when fine-tuning
- M1111-04GPU memory requirements for training an LLM
- M1111-05LoRA and parameter-efficient fine-tuning (PEFT)
- M1111-06RLHF: reinforcement learning from human feedback explained
- M1111-07Reward models, reward hacking, and preference data
- M1111-08Prompt vs RAG vs fine-tuning: the full decision rule
- M1111-09Choosing a model adaptation strategy under real constraints
Part of the throughlines running across the NCA-GENL prep course.