Practical LLMs: useful tools and failure modes
Simon Willison discusses prompting, data extraction and model limitations in his PyCon US 2024 keynote.
What it covers
Simon Willison's annotated transcript follows his PyCon keynote through prompting, retrieval, tool use, Code Interpreter and structured extraction. It suits technical readers evaluating concrete uses alongside hallucinations and prompt injection. Coding examples assume some familiarity with developer tools.
Extend the safety lesson
His separate 2025 security article examines combining private data, untrusted content and external communication. Read it before connecting assistants to sensitive workflows; it explains how instructions embedded in incoming material can trigger unintended data sharing.
The keynote is a 2024 snapshot: its model comparisons and product examples need that historical context.
