Let's talk about LLMs
Large language models are prompting sharp disagreement over how much they truly change software development versus acting as yet another productivity tool. Commenters weigh claims of “paradigm shift” and 10x gains against evidence of buggy code, fragile workflows, regulatory and reliability constraints, and the risk of overhyping benchmark wins while real‑world impact remains modest. Many see LLMs as powerful for debugging, refactoring, and auxiliary tasks that augment skilled engineers, but doubt they will soon replace professional expertise or fundamentally solve the hard parts of building and operating software systems.
LLM Discourse Fatigue vs Ongoing Obsession
- Several posters are tired of yet another “what will LLMs do to society” take and compare the hype cycle to crypto.
- Others push back: if you’re bored, ignore it; clearly many still want to argue about it.
Tool vs Paradigm Shift
- One camp: LLMs are just powerful new tools (like calculators, CAD, or drill drivers). They help with “accidental difficulty” but don’t alter the fundamentals of software engineering.
- Opposing camp: LLMs are a genuine paradigm shift; coding is shifting toward orchestration, tooling, and governance of AI-generated code, with agentic workflows becoming standard.
- Some note that in practice their job changed dramatically within a year, which feels paradigm-level even if theory says “just a better tool.”
Productivity, “10x,” and Silver Bullet Arguments
- The thread revisits No Silver Bullet: essential vs accidental complexity and skepticism about 10x productivity.
- Critics argue: LLMs mostly reduce typing and boilerplate, and empirical studies so far suggest modest gains with stability risks.
- Supporters counter that coding/agent tools are already huge productivity boosts, especially for debugging, ops, reporting, and internal tooling.
- Debate over “10x programmers” and whether LLMs can move organizations anywhere near that; wide disagreement, from “no such thing” to “we’re already close.”
Quality, Reliability, and “Vibe Coding”
- Many praise LLMs for debugging, code review, refactoring, test writing, and documentation; coding from scratch is described as mixed and fragile.
- Reports of impressive small/greenfield projects contrast with failures on more complex or high-stakes systems.
- Some see “vibe coding” as democratizing; others warn it produces fragile “big balls of mud,” especially dangerous in regulated or mission-critical domains.
Future Trajectory and Scaling Laws
- Pro-AI posters lean on scaling laws, benchmarks, and rapid capability gains, arguing there’s no clear ceiling yet.
- Skeptics mention regressions in newer models, benchmark overfitting, possible asymptotes, and lack of visible macroeconomic impact beyond growing debt.
- Both sides agree current models are imperfect; the dispute is whether improvements will plateau below or surpass broadly competent human programmers.