Chapter 1

Language Models: From Next-Token Prediction to Useful Systems

September 2026 · Study outline

Chapter: Introduction, in Daniel Jurafsky and James H. Martin, Speech and Language Processing, 3rd edition draft, August 19, 2026.

Topics to learn

Page numbers refer to the linked chapter PDF.

• What makes a language model large pp. 3-5
  • Parameters as weights and biases p. 5
  • How model size and training tokens scale with compute p. 5
  • Size, inference cost, and resource tradeoffs p. 5
  • Monolingual and multilingual models p. 4
  • Tokens versus words, including shorter tokens outside English p. 5
• What next-token prediction can teach pp. 6-8
  • Conditional distributions and autoregressive generation p. 4
  • Prompting as conditional generation p. 7
  • The distributional hypothesis p. 6
  • The Shannon game, information, and compression p. 8
• Neural networks, embeddings, and model access p. 9
  • Weights as the learned computation p. 9
  • Open-weight and closed or proprietary models p. 9
• History of the field pp. 10-11
  • Statistical and neural foundations p. 10
  • The symbolic turn p. 11
  • The revival of empiricism p. 11
  • Prompting and the current language-model period p. 11
• Language structure and interaction pp. 12-13
  • Syntax, semantic roles, coreference, and discourse p. 12
  • Conversational common ground and grounding p. 13
  • Sycophancy and overconfidence in dialogue p. 13
  • Probing a representation versus showing the model uses that structure p. 12
  • Using linguistic structure in cognitive and social science p. 13
• Training pp. 13-18
  • Pretraining data, including Common Crawl, curation, deduplication, and quality or toxicity filters p. 15
  • Filter failures, copyright, privacy, skew, and provenance p. 15
  • Instruction tuning p. 15
  • Preference pairs, and PPO and DPO by name p. 17
  • Reinforcement learning with verifiable rewards, including formal output constraints p. 17
• Inference pp. 19-21
  • System, user, and assistant roles p. 20
  • Zero-shot and few-shot prompting p. 19
  • What demonstrations contribute, and when more of them stop helping p. 19
  • Chain-of-thought and test-time compute p. 20
  • The full path from an assembled prompt to a stopped generation p. 20
• Agents p. 22
  • Acting by calling programs p. 22
  • The reason, act, and observe loop p. 22
  • Appending observations to the context p. 22
• Evaluation pp. 23-26
  • Accuracy on benchmarks such as MMLU, BIG-Bench Hard, GSM8K, and AIME24 p. 23
  • Perplexity p. 25
  • Human rubrics and LLM-as-a-judge, singly or pairwise p. 25
  • Proxy overlap metrics such as chrF, BLEU, and ROUGE p. 26
  • Goodhart's law and statistical significance p. 26
  • Latency, energy, and fairness, including worst-group comparisons p. 26
• Safety, alignment, and wording pp. 27-29
  • Dependence, de-skilling, and sycophancy p. 27
  • Representational harm and unequal performance p. 28
  • Privacy, misuse, and environmental cost p. 28
  • Prompt injection and other agent risks p. 28
  • Value-sensitive design, review boards, constitutional AI, and red teaming pp. 28-29
  • Anthropomorphism and the intentional stance p. 29

FAQ