Hacker News new | ask | show | jobs
by mellosouls 252 days ago
(2022) This is particularly important here as the essay makes no mention of LLMs or coding agents (which were still in their infancy in development environments; this article is post original copilot/codex but pre ChatGPT).
2 comments
aDyslecticCrow 252 days ago
> This is particularly important here

No actually. Why is that important? I dont quite see why that is relevant. Could you elaborate?

link
mellosouls 252 days ago
In at least two ways, which you might see as alternatively positive and negative wrt whether expertise in programming languages is useful.

Firstly, when using them to create software it's pretty obvious that experienced devs and people who understand theory have a greater ability to guide, curate and control them.

Secondly, as they improve in ability we can see a paradigm change for people using them at least as significant as the jump from assembly to high level languages. Most programmers would have no need to study assembly these days.

Either way, their omission (while appropriate for the year, if somewhat lacking in foresight) is a significant one that renders it somewhat dated already.

link
cubefox 252 days ago
Yeah. Now the trend goes in an entirely different direction: telling an LLM in natural language what to achieve using those programming languages. How quickly times change.

Edit: I assume this comment gets downvoted because people don't like where we are heading, not because they really think LLM programming capabilities won't continue to improve at a staggering pace.

link
aDyslecticCrow 252 days ago
> LLM programming capabilities won't continue to improve at a staggering pace.

The error rate of models make language design, tooling, testing methodology and human review more important than ever before. This demands language evolution. You could get faar with lax testing and language tooling with enough caution and skill. but when LLMs enter the picture, that no longer flies.

we need tooling, static analysis, testing paradigms, language design that restrict how dangerous the LLM is allowed to act.

natural language is faar to fuzzy to replace programming (system specification is already famously impossible thing to do right). If you think it truly will replace code, i highly suspect you work om webbdesign, where testing and reliability was always a secondary concern.

And even then, I think were already on the convergence platoe of LLM code. The companies are raising prices as diminishing improvement and balooning compute costs.

link