Learn how modern language and multimodal AI systems work and how to use them responsibly.

Generative AI and language systems

Status: Advanced outline. Use the sample AI exercises as preparation.

Prerequisites

  • Python functions and JSON
  • Basic probability and evaluation
  • AI exercises
  • A clear understanding that generated text is not automatically evidence

Learning objectives

  • explain tokens, embeddings, retrieval, and model context;
  • write prompts with explicit constraints and structured outputs;
  • evaluate groundedness, usefulness, and failure cases; and
  • design permissions and human review for tools or agents.

Ordered subtopics

  1. Tokens, language models, and probability
  2. Embeddings and semantic similarity
  3. Transformers and attention
  4. Prompt design and structured output
  5. APIs, local models, and cost awareness
  6. Retrieval-augmented generation
  7. Evaluation, grounding, and citations
  8. Tool use and simple agents
  9. Hallucinations, bias, privacy, and prompt injection
  10. Multimodal systems and accessibility

Planned project: build a citation-based study assistant over a small, approved document set.

Completion evidence

Submit a small application with an approved document set, citations, at least ten evaluation questions, a hallucination/error log, a privacy boundary, and a human-review step.

Best practices

  • Use the smallest model and context that can solve the task.
  • Treat retrieved documents and user content as untrusted input.
  • Never place secrets or private student records in prompts.
  • Validate structured output before using it in a program.
  • Use read-only, narrow tools first; require approval before consequential actions.
  • Measure repeated test cases instead of showcasing one impressive response.

References