To be clear, it's the original purpose of LLMs.

The whole LLM scene today came about because context was really important to translations. The "attention is all you need" paper was by the Google Translation team as they came up with ideas to improve how to map context of words and carry them across in translations.

At some point people started asking the translation to "translate from English to English as if you're an AI assistant".

Anyway it shouldn't surprise anyone that LLMs are good at translation. The real surprise to everyone is how powerful translation engines that understood context could be!

One distinction is the original transformer was an encoder/decoder while (most?) LLMs today are encoder only.

The translation transformer also was able to peek ahead in the context window while (most?) LLM's now only consider previous tokens.

They're usually thought as "decoder only"

Oops yes thank you, was late when I replied.

I like to think of it as if the LLM is simply translating questions into answers.

>They are also quite good at translating poorly written and only semi coherent writing, which can be incredibly useful if the person you are communicating with is quite sloppy.

You see this with recent automated translation on YouTube. If the creator of (say) an English-language video doesn't upload subtitles, YouTube automatically creates them based on the audio, but they lack punctuation and have nonsense phrases. The AI-driven translation of those subtitles to other languages cleans up the text along the way, so the end result is that non-English speakers get better subtitles than English speakers.

Bringing it back to the other comments, they should do EN->EN translation on the transcription.

It also makes sense that they would be good at translating from English to programming languages, for the same reasons.