Comment by monk_grilla
3 days ago
I prefer to get bleeding edge LLMs to use old, time-hardened technologies. They are more stable, I trust they're not going to change out from under me, and it is more transferable. The ergonomics of the actual markup don't matter to me, since I'm not writing it, so I'd rather the more powerful, more training data option.
Just one example: Every university and academic publication in the world give LaTeX formatting templates, examples courses etc.
Ergonomics matter even for LLMs. Also, compile speed, quality of error messages, how easy it is to get something right (especially that LLMs are "blind" unless you want to burn a lots of tokens, so they sort of just reason without the visual output) absolutely matter and can turn a one-shot LLM prompt into a multi-turn agent churning wasting tokens for no good reason.
I'm not saying that it's necessarily the case here, but typst is past the very early phase and LLMs are decent at generating it. Also, many formatting templates have been converted to typst as well.