{
  "id": 7969848,
  "title": "AI made translation cheap. The hard part was knowing what architecture to keep.",
  "url": "https://urgent.news/2026/09/17/ai-made-translation-cheap-the-hard-part-was-knowing-what-architecture",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-17T06:28:58.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/michaeltruong/ai-made-translation-cheap-the-hard-part-was-knowing-what-architecture-to-keep-4hp3"
  },
  "original_language": "en",
  "account": "Codenames AI, a web game that allows an LLM to play the game with players, faced the challenge of supporting multi-language selection, starting with Simplified Chinese. The author, not a native Chinese speaker, utilized AI to generate candidate translations rapidly, bypassing their own judgment. However, the author needed a system to ensure the translations were natural, ambiguous in the right way, and not simple mistakes. This involved deterministic checks for verification and explicit review points for human judgment. Initially, the author assumed a strict substring-collision validator would protect card-word quality, but this proved half right. Treating collision purity as a hard invariant made the Chinese translations worse, pushing them towards single-character fragments and awkward abbreviations. Relaxing the rule allowed for review, and natural, compact cards were restored. The author realized that elevating the signal of collision-free translations to the same severity as duplicates and empty values was a mistake. Despite passing validation, the translations may not always be natural. Codenames is built on ambiguity in English, while Chinese often forces a choice of meaning explicitly. Certain English concepts do not survive as single clean Chinese words without choosing a specific meaning. Thus, the author differentiated between authoring process and runtime requirement. The architecture initially modeled Simplified Chinese as a localization of English Classic pack, but later realized that a native Chinese Codenames list is not a translation of the English Classic pack. The runtime requirement was smaller: wordSet × language → playable word pool, with Classic English, Classic Chinese, Extended English, and Extended Chinese using the same path. The author removed the mapping layer, generator scripts, and old comparison artifacts, retaining the shipped words. The process learned that AI made producing translations cheap, but it did not decide what deserved to become permanent structure. The confusion between machinery that helped build a feature and the product's actual model led to these mistakes.",
  "summary": "I ship Codenames AI , a web game where an LLM plays Codenames with you. One of my longtime readers, @xulingfeng , asked whether the game could support multi-language selection, and that was enough to pull forward work that had been sitting on the backlog: new word sets and languages, including Simplified Chinese. I am not a native Chinese speaker. AI could generate candidate translations far…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}