Trying to do things off the beaten path with LLMs is rarely successful, especially if there's a related much more popular option.
I'm convinced that programmers' bias towards LLMs is strongly correlated with the weirdness of their work. Very often my strange ideas pushed to LLMs look like solutions but are rather broken and hallucinated attempts which only vaguely represent what needs to be done.
The JS ports of FFmpeg (or WASM port if you want the in-browser approach) are very old and would be more than present in modern LLM training datasets, albeit likely not enough of a proportion for LLMs to understand it well.
