llms-for-R.md

LLM for code generation

Currently three main styles:

1. Autocomplete

   • e.g. github copilot, windsurf
   • Inline "ghost text" as you type
   • Sometimes amazingly good; often pretty useless
   • Need to train yourself to ignore spurious suggestions

2. Chat

   • e.g. ChatGPT, Claude, positron assistant
   • Claude best for R code

3. "Agentic"

   • e.g. claude code, cursor.
   • Has ability to run code and change files
   • Comes up with a plan, iterates, and works away on it
   • Can take a long time; cost a lot of tokens.

LLM pros

• Amazing for quickly generating 95% demos

  • e.g Shiny app from a hand drawing
  • "Create an data frame containing data about X"

• Great at translations:

  • commad line curl -> httr(2)
  • Latex -> Quarto
  • R code -> STAN
  • JSON -> unit tests

• Very helpful at finding words/algorithms that you don't know about.

• Don't forget that it doesn't just generate code; it can explain and critique code too.

LLM cons

• Less good at making incremental changes to large existing codebases

• Doesn't know about latest package versions. At some level is a weighted average of all code on the internet, so tends to older/more popular idioms rather than modern approaches.

• Output always looks plausible, but functions and arguments can be hallucinations. You can not (MUST NOT!) blindly trust its output.

• Gets stuck in local optima in long conversations; start fresh sessions frequently