LLM Use Cases

Some real-world “use cases” (half of which I disagree with).

1. Retrieval/Search
   • both inter- & intra-document (find doc in corpus & find in doc)
   • semantic matching (rather than keyword matching & link analysis)
   • e.g. DropBox Dash, CommandBar HelpHub
   • (personal opinion) seems a valid use case provided the output includes verifiable citations to original source
2. Summarisation/Classification
   • both inter- & intra-doc (whole-doc & per-paragraph)
   • e.g. Notion “Summarise”, Superpowered
   • (personal opinion) danger is users are very likely to assume 100% accuracy
3. Structured translation (for machine APIs)
   • e.g. Booking.com “Trip Planner”, Linear “AI Filters”
   • (personal opinion) identical to 1. Retrieval/Search, except with formatted (e.g. JSON) outputs – so not a separate use case
4. Unstructured translation (for human users)
   • proofreading, [paraphrasing], explaining, refactoring
   • e.g. Mutiny, Replit “Ghostwriter”
   • (personal opinion) “explaining” quality is currently unusably poor
   • (personal opinion) “proofreading/paraphrasing” quality is currently poor so only useful in the context of “highlighting confusing text”
5. Text Generation
   • e.g. Jasper AI, Tome
   • (personal opinion) this is terrible. AI artificial intelligence
     is not AGI artificial general intelligence
     , so is incapable of generation. It can paraphrase (at best) or plagiarise (at worst) – falling foul of fraud, misrepresentation, & IP theft laws.