Rule-based translation is a traditional technique used in compilers and controlled natural languages (CNLs). It can be made more reliable than translation with large language models (LLMs), which on the other hand needs less work and is therefore easier to scale up. But there is a way to combine the advantages of both techniques: let LLMs generate translation rules. To guarantee reliability, these rules have to be checked for correctness, but this can be expected to require less work than writing the rules from scratch.
This talk will report on recent experiments in using LLMs for extending the scope of Informath, a translation system between formal and informal mathematics. Vibe coding has been used for learning both grammar rules in GF (Grammatical Framework) and wrapper code in Haskell, based on mathematics text books and formalizations,, existing grammars, and prompts.