> Why not use a transformer-based model architecture actually designed for translation?
Because translation requires a thorough understanding of the source material, essentially up to the level of AGI or close to it. Long-range context matters, short-range context matters, idioms, short-hand, speaker identity, etc... all matters.
Current LLMs do great at this, the older translation algorithms based on "mere" deep learning and/or fancy heuristics fail spectacularly in the most trivial scenarios, except when translating between closely related languages, such as most (but not all) European ones. Dutch to English: Great! Chinese to English: Unusable!
I've been testing modern LLMs on various translation tasks, and they're amazing at it.[1] I've never had any issues with hallucinations or whatever. If anything, I've seen LLMs outperform human translators in several common scenarios!
Don't assume humans don't make mistakes, or that "organic mistakes" are somehow superior or preferred.
[1] If you can't read both the source and destination language, you can gain some confidence by doing multiple runs with multiple frontier models and then having them cross-check each other. Similarly, you can round-trip from a language you do understand, or round-trip back to the source language and have an LLM (not necessarily the same one!) do the checking for you.
Because translation requires a thorough understanding of the source material, essentially up to the level of AGI or close to it. Long-range context matters, short-range context matters, idioms, short-hand, speaker identity, etc... all matters.
Current LLMs do great at this, the older translation algorithms based on "mere" deep learning and/or fancy heuristics fail spectacularly in the most trivial scenarios, except when translating between closely related languages, such as most (but not all) European ones. Dutch to English: Great! Chinese to English: Unusable!
I've been testing modern LLMs on various translation tasks, and they're amazing at it.[1] I've never had any issues with hallucinations or whatever. If anything, I've seen LLMs outperform human translators in several common scenarios!
Don't assume humans don't make mistakes, or that "organic mistakes" are somehow superior or preferred.
[1] If you can't read both the source and destination language, you can gain some confidence by doing multiple runs with multiple frontier models and then having them cross-check each other. Similarly, you can round-trip from a language you do understand, or round-trip back to the source language and have an LLM (not necessarily the same one!) do the checking for you.