Advanced coding in low-resource programming languages
Advances in large language models are enabling AI systems to write and migrate code in older or niche programming languages like COBOL, Fortran, and even custom or esoteric syntaxes, raising the prospect that language choice may become far less constraining for software projects. Commenters highlight both impressive feats—such as LLMs reimplementing SQLite or helping reverse‑engineer legacy game platforms—and serious concerns about relying on “vibe‑coded” AI output for safety‑critical or financial infrastructure, where maintenance, verification, and test coverage remain paramount. Many see potential in using AI to understand and modernize legacy systems rather than replace them outright, especially as expert human maintainers age out of the workforce.
Safety-Critical Systems and “Vibe Coding”
- Strong pushback against using LLM-generated “vibe code” for banking, nuclear, and avionics systems.
- Some joke that such systems might eventually be “too important” to leave to fallible humans, with counterexamples raised of humans averting disasters.
- Commenters note that these systems rely on formal methods, proofs, and long-tested code; LLMs might help understand legacy code but shouldn’t author core safety logic.
- Maintenance is seen as the real bottleneck; many would enthusiastically write fresh Rust but far fewer want to maintain it long term.
Language Choice Becoming More Fungible
- Several see it as “sci‑fi” that language choice could become as temporary as a variable name, thanks to LLMs and agents.
- Others argue frequent rewrites signal poor planning and increase risk; they prefer evolving, well-tested code.
- Counterview: with good specifications and tests, implementation language is an implementation detail; easier language migration can reduce tech debt.
- Example motivations: prototyping in Python then rewriting in Rust, or moving from Smalltalk to Go for deployment reasons.
LLMs and Low-Resource / Niche Languages
- Multiple reports that frontier models already handle custom or obscure languages well when given rules or skills in-context.
- Niche communities (e.g., around legacy multimedia tech) are using LLMs to accelerate reverse engineering and preservation efforts.
- Some surprise that COBOL performance lags while Lisps and other languages do relatively well; unclear if this is data scarcity or deeper limitations.
- Interest in how models generalize to esoteric or freshly invented languages.
SQLite-from-Scratch and Benchmarks
- The internal OpenAI task of reimplementing SQLite, even in languages like PowerShell or PHP, is seen as impressive but also somewhat theatrical.
- Some stress that doing an SQLite-compatible implementation with strong test suites is a tough but plausible capability benchmark, not a production replacement.
- Comparisons are made to failed attempts at LLM-generated C compilers; debate over whether SQLite or a C compiler is actually harder.
Assembly and New Low-Level Syntaxes
- Discussion of new, higher-level assembly syntaxes (e.g., with C-like types, structured blocks) as potentially more LLM-friendly.
- View that LLMs are already good at raw mnemonics, so benefits may be more about human readability than AI needs.
- Examples cited of prior similar efforts and skepticism about broad adoption, though some see a possible resurgence if LLM workflows favor such targets.
Broader Reflections on Legacy & Niche Languages
- Nostalgia for mainframe-era languages and resource constraints; sense that AI will shift programmers from “typists” to solution designers.
- Mild disappointment that some beloved niche languages (Haskell, Clojure, Forth) are omitted from the article’s focus, though many expect LLMs to handle them reasonably well.
- Overall consensus: language choice matters less for LLMs than for humans, but safety, testing, and long-term maintenance still matter a lot.