Translation Problems

Type: LLM capability pattern
Referenced from: audit-test-automate-ai-delegation

Definition

LLMs are strongest at translation problems — converting between two well-documented formalisms. The transformer architecture was literally built for translation, and these tasks align perfectly with how LLMs learn.

Why Translation Works

  1. Both formalisms are well-documented — The training data contains abundant examples on both sides.

  2. Translation forces tacit assumptions to surface:
    - Vernacular hides assumptions: "the system handles errors gracefully"
    - Formal target languages can't accept that: you must spell out what happens, under which conditions, with which fallback
    - Same with math: "nothing" and "zero" are basically the same in vernacular, but radically different in programming

  3. The LLM usually picks the most conventional answer — which is often the right one for structured translation.

Canonical Examples

  • Code ↔ Documentation — Engineering ideas to API docs, code comments to specifications
  • Spec ↔ Tests — Requirements to test cases, behavior descriptions to assertions
  • Prose ↔ Structured Data — Natural language to JSON, free text to database records
  • Engineering Ideas → Patent Claims — Technical innovations to legal claim language
  • Requirements → Architecture — Product specs to system design documents

Why This Matters for dark-factory-kb

The KB pipeline IS a translation problem:
- Source articles (prose) → Structured concepts (formalism)
- Concepts → HTML pages (formalism)
- Log entries → Structured reports (formalism)

This is the LLM's sweet spot, which explains why the pipeline works well with agent delegation.