
Day-to-Day LLM Stack and Learning to Use It

AI tools for software engineers, but without the hype – with Simon Willison (Co-Creator of Django)
Topic 3 of 7
Published Sep 25, 2024 · Clip 10:27–33:35 (23:08)
Simon describes his day-to-day LLM stack: Claude 3.5 Sonnet as default, GPT-4o for code interpreter and voice mode (used during dog walks), GitHub Copilot autocomplete (heavily used, learned through comment-based prompting), his open-source llm command-line tool, and local models like Llama and Mistral for research and learning. He explains his role as an independent researcher, learning Rust with ChatGPT during Advent of Code, and stresses these tools are difficult to use — you must build an intricate mental model of what they're good and bad at. He debunks fine-tuning for adding knowledge and praises Claude's prompting documentation while criticizing the prevalence of junk prompting advice and superstition.