Raw LLM Responses

I don’t believe for one second that AI (or English—or Bhojpuri or Tongan for that matter—😂 human language)—won’t replace developers. Their first attempt at doing so failed, but that is expected. My prediction is that Cursor, Windsurf, replit and the like are nothing but the first generation of AI code tools which produced functioning code for the most part but led to tech debt problems like structurally less diverse code and less thorough coverage of edge cases. “Vibe coding” is now firmly cemented as a paradigm but we haven’t quite figured out the best practices yet nor the best tech for it. Y Combinator released a helpful video on empirically proven best practices for vibe coding some time ago, but I think that will amount to something more like taking stock of the current state of the art in 2025 when we look back on that video ten years from now. Yes developers are safe for now, but you can bet that researchers are actively working on addressing the issues of AI assisted code generation 1.0. I can expect Gen 2.0 of this paradigm to be much more sophisticated. Think about the natural language translation industry as a reference for historical comparison. Human translators in the U.S. who got BA degrees in Spanish/French and MA’s in Translation Studies used to be able to put children through colleges with translator salaries. They had a specialized white collar skill which required hours upon hours of instruction to master the grammar, vocab, syntax etc. of their target translation language. Google Translate 1.0 was based on modeling rules symbolically and used to get so much wrong that it wasn’t reliable. Human translators must have let out sighs of relief at the time. But Google Translate 2.0 came back strong with statistical approaches. Vibe translation requiring a reduced number of translators overall is now firmly embedded in a lot of translation work, including game localization for the video game industry, leaving only select domains like highbrow literature and poetry intact.