Translation

Machine translation (MT) is the automated conversion of text or speech from a source language into a target language while preserving meaning, tone, and (in good systems) cultural nuance. Neural machine translation dominates today, having replaced older statistical/rule-based systems; dedicated translation models (Google Translate, DeepL) use encoder-decoder transformer architectures trained on massive parallel corpora (aligned sentence pairs across languages), while general-purpose LLMs (GPT-4, Claude, Gemini) have also become highly competent translators — often outperforming dedicated MT systems on nuance, idiom, and context-aware translation (understanding that "it's raining cats and dogs" isn't literal, or that a product name shouldn't be translated) because they bring broader world knowledge and can be steered with instructions like "translate formally for a legal document" or "keep the casual tone." Why it matters for SaaS builders: translation underlies i18n/localization tooling (auto-translating a product's UI strings and marketing content into new markets), real-time chat/support translation (letting a support agent and customer converse across languages), subtitle/caption localization for video platforms, and cross-border e-commerce (auto-translating product listings). A crucial distinction for production SaaS: machine translation is fast and cheap but must never be used verbatim for brand-critical, legal, or medical content without human review — subtle mistranslations in these domains carry real liability. A concrete worked example — a SaaS product localizing its UI: (1) the codebase's English `en.json` locale file contains `{"dashboard.welcome": "Welcome back, {name}!"}`; (2) a translation job sends each string plus surrounding context (which screen it appears on, whether it's a button label or a full sentence) to the LLM: "Translate this UI string to Turkish, preserving the {name} placeholder exactly and keeping it appropriately informal for a SaaS product dashboard: 'Welcome back, {name}!'"; (3) the model returns `"Tekrar hoş geldin, {name}!"`; (4) the string is written to `tr.json` but flagged as `needs_review: true` in the translation management system, and held for a native-speaker review pass before it ships to end users — never a Day-1 bulk auto-dump for a brand-new launch locale, since raw, unreviewed machine translation frequently produces text that's technically correct but reads as stiff, overly formal, or subtly "translated-feeling" in ways a native speaker immediately notices and that measurably damages trust when entering a new market. Builders localizing at scale typically batch-translate everything quickly with an LLM as a first pass, then gate actual publication behind human review, staggering the review-and-release process so translated content ships incrementally rather than all at once.