AI Has No Wisdom and Neither Will You
Engineers debate whether AI-generated code is inherently doomed to become an unmaintainable mess or simply reflects long‑standing human problems with bad architecture and tech debt. Many argue that large language models can already match or surpass average developers if guided by strong specs, tests, and “agent scaffolding,” but worry that over-reliance will erode institutional knowledge, junior developers’ skills, and the ability to make wise architectural tradeoffs over years. Others counter that coding itself is becoming a cheap, disposable commodity, and that the real differentiator will be how teams design systems, define constraints, and balance short-term velocity against long-term code health in an AI-saturated workflow.
AI‑generated code quality and “vibe coding”
- Many see LLM/agent‑written code as over‑complex, inconsistent, and prone to hidden bugs, especially when used with vague “do X” prompts.
- Others argue this is not fundamentally different from mediocre human code; AI just accelerates both feature delivery and code rot.
- Several note that AI quality is heavily shaped by constraints, examples, and review; unguided agents predictably produce slop.
Maintainability, metrics, and architecture
- Thread agrees that long‑term maintainability is hard to measure; common metrics (cyclomatic/cognitive complexity, SonarQube) are noisy and easily gamed.
- Some suggest RL environments and benchmarks for incremental edits could train models toward maintainability, but this is still emerging.
- Higher‑level architecture and boundary design are where agents currently fail most; tools can enforce low‑level rules, not good system design.
Learning, wisdom, and institutional knowledge
- A recurring worry: developers who “haven’t written code since 2025” may never build deep intuition, becoming dependent on AI.
- Comparisons are made to outsourcing manufacturing and losing industrial know‑how; AI could similarly erode software craftsmanship and domain understanding.
- Others counter that humans already rarely reach mastery and that “wisdom” about code is often tacit and hard to formalize even for teaching humans.
How people are actually using AI
- Some report large productivity gains: using agents for most implementation, then focusing on specs, architecture, and tests; doing major refactors in hours; generating more exhaustive testing and validation.
- Others find they spend too much time “nagging” models and fixing odd designs, preferring to code critical paths by hand.
- Techniques mentioned: AGENTS.md conventions, rich code comments as “infrastructure‑free memory,” multi‑agent cross‑review, and explicit refactoring passes to reduce accumulated slop.
Incentives, labor, and culture
- Many blame managerial focus on velocity and time‑to‑market for tolerating AI slop and penalizing engineers who slow down to improve quality.
- Hiring and evaluation become harder when applicants can use AI to mask weak fundamentals.
- Some argue caring about code quality is mostly a culture and incentive problem, not a tooling one.
Predictions about the future
- One camp expects ubiquitous AI, possibly fully agentic pipelines; “NO‑AI” policies seen as niche or unsustainable.
- Another predicts “NO‑AI” (or AI‑minimal) shops as a premium or safety‑driven niche, especially in high‑stakes domains.
- Unclear consensus on trajectory: some see AI coding quality improving rapidly; others see stagnation in core problems like maintainability and context limits.