- Mitchell Hashimoto distinguishes two ways of using language models: interactive prompting (ChatGPT-style, human-driven discussions) and transactional prompting (treating the model like a function, software-driven, high volume, specific tasks).
- Prompt engineering is primarily about transactional prompting, aiming for accuracy and low cost, and its main benefit lies in transactional use cases rather than interactive ones.
- Another key dimension is objectivity vs. subjectivity: prompt engineering works well for objectively correct outputs (e.g., information extraction, classification) but is less effective for subjective tasks (e.g., creative writing, art generation).
- Code generation can be both objective (small, well-specified tasks) and subjective (larger, less specific tasks), with prompt engineering more applicable to the former.
- Prompt engineering for interactive prompting is like being an effective Googler, while for transactional prompting it resembles being a data scientist.
- Agents, though automated interactive processes, are primarily transactional in nature, so prompt engineering applies to them as well.