Advanced coding in low-resource programming languages

Commenters discussed whether large language models can meaningfully program in low‑resource or niche languages and what that implies for software practice. Some, like nicfio, expressed nostalgia for tight, resource‑constrained programming and welcomed LLMs as tools that change who writes code, while others such as kragen highlighted surprising internal tests like reimplementing SQLite as evidence of rapidly improving capability. Practitioners reported that modern models can learn custom syntaxes from context (giancarlostoro, vegnus, ch4s3), and hobbyists described using LLMs to accelerate reverse engineering and preservation of legacy runtimes. Raphlinus and Dan Luu noted weaker performance on assembly and argued clearer intent and compiler verification could remedy model shortcomings. Dooglius raised the important distinction between true generalization and simply producing lots of training data for a given language.

Opinion divides sharply on risk, maintenance, and signaling. Several commenters warned against “vibe coding” safety‑critical systems like banking and nuclear control (anonymous908213, AnimalMuppet, bayindirh), and others emphasized maintenance and tech‑debt problems when code is generated but not maintainable by teams (npalli, gonzalohm). Some framed models’ language‑agnosticism as evidence of good design and adaptability (retrac, malisper), while critics pointed to proprietary test suites and past failures (favorited, sarchertech) as reasons for caution. Overall, the thread balances excitement about lowered barriers to language creation against concern for long‑term verification, governance, and who ultimately bears responsibility.