University of Turin publishes practical textbook on large language models

The University of Turin published an educational document on arXiv (identifier 2610.37684v1), created as background material for a lecture within a master's program in artificial intelligence in medicine. The text was written as a practical and accessible introduction to large language models (LLMs) intended for scientists and professionals without programming knowledge, explaining the principles of tokenization, embeddings, training, inference, context window, prompting, and instruction hierarchy without mathematical detail.

A substantial part is devoted to the limitations of LLM outputs — hallucinations, uncertainty, reproducibility, biases, and the need for systematic verification and human oversight. The document also covers topics of privacy, governance, security, and the responsible handling of personal and sensitive data, and, in a more advanced section, Retrieval-Augmented Generation (RAG), search quality, chunking, vector search, agents, tools, and workflows.

It includes an extensive practical exercise in which the reader, step by step, runs a local LLM using the Ollama and Open WebUI tools and uses it to interact with a collection of scientific articles. The exercise focuses on evidence search, source verification, working with uncertainty, discrepancies between studies, and distinguishing between plausible-sounding text and scientifically grounded conclusions.