Research Notes: LLM Principles, Explained

1. Source map

Primary sources

• Vaswani et al., Attention Is All You Need
• Brown et al., Language Models are Few-Shot Learners
• Devlin et al., BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• Ouyang et al., Training language models to follow instructions with human feedback
• Bai et al., Constitutional AI: Harmlessness from AI Feedback
• Carlini et al., Extracting Training Data from Large Language Models
• Nasr et al., Scalable Extraction of Training Data from (Production) Language Models
• Lewis et al., Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
• Guu et al., REALM: Retrieval-Augmented Language Model Pre-Training
• Liu et al., Lost in the Middle: How Language Models Use Long Contexts

Authoritative context

• Kudo and Richardson, SentencePiece and related subword-tokenization work
• Public material from Anthropic and OpenAI on instruction following and agentic workflows, used as context rather than core evidence

Secondary context

• Bender et al., On the Dangers of Stochastic Parrots

2. Evidence notes

• Tokenization is treated as splitting text into units the model can process.
• Embeddings are treated as high-dimensional vector representations of tokens.
• Transformer self-attention is explained as direct context relation modeling.
• Pretraining is treated as learning broad generalization through next-token prediction.
• Inference is treated as runtime candidate generation using fixed weights.
• Context length is treated as the maximum input the model can directly condition on.
• “The model stores knowledge” is written carefully to distinguish distributed retention from verbatim memorization.

3. Inclusion and exclusion decisions

• Include: non-specialist explanations of the core pipeline, minimal technical terms, and practical cautions.
• Include: a self-attention diagram and an explanation of Q/K/V as role names.
• Include: the accuracy and limits of “knowledge storage” through memory, generalization, and extractability.
• Include: context-length limits and the practical case for retrieval alongside the model.
• Exclude: derivations, loss-function details, optimizer comparisons, and benchmark minutiae.
• Exclude: current vendor feature comparisons, pricing, and fast-changing product specs.

4. Rejected or downgraded sources

• Secondary blog explainers were not used as core evidence, even when they were readable.
• News-style product commentary was downgraded because the article is about first principles.
• Interpretations that treat attention weights as a direct explanation were not centered because they can mislead readers.

5. Open questions

• The failure threshold for long-context use depends strongly on the task.
• The rate of verbatim memorization varies by model, duplication rate, and evaluation method.
• The word “understanding” is philosophically loaded, so the article keeps the practical decision axis in focus.