OpenAI Codex Works Best in Non-Code-Editors

  • Codex works poorly for most serious synthesis/domain-specific tasks.
  • Hypothesis: non text editor environments are where codex will excel

Like:

  • Cell/Node-Based Environments where functions are compact and contexts are specified in a structured way
    • Blockly?
  • Strongly typed languages where type annotations specify a lot
  • Define sprites/assets/params, then ask for game behavior in prose, then iterate
  • Well-specified primitives/standard libraries
    • Numpy/OpenCV - "solve for" operations in data pipelines
    • Autodiff