The problem is not AI code, but not knowing about system architecture or intent
Generative AI tools are increasingly writing large amounts of production code, but many engineers argue the real bottleneck has simply moved from typing to understanding: system architecture, intent, and long‑term maintainability. Commenters describe organizations where specs, tickets, code and even executive memos are AI‑generated, eroding human ownership, mental models and accountability while incentivizing “vibe coding” and quick feature output over reliability. Others counter that AI can be a powerful accelerator and even aid in understanding legacy systems, provided teams enforce human review, clear architectural abstractions and a “responsible human in the loop” who still knows what the system does and why.
AI “solved coding” vs reality
- Some argue LLMs have effectively “solved coding”: you can deliver features by working in natural language (requirements, specs, tests) and letting agents write code, often faster than human teams.
- Others strongly reject this: LLM code is viewed as insecure, inefficient, bug‑prone “slop” that might compile but fails on robustness, performance, and long‑term quality.
- A middle view: LLMs drastically reduce the cost of mechanical coding, but software engineering (requirements, tradeoffs, architecture) is unsolved and now more important.
Architecture, intent, and maintainability
- Many see the real bottleneck as system architecture, intent, and maintainability, not line‑level coding.
- Concern: if AI generates large chunks of code and no one deeply understands them, ownership and the ability to reason about systems erode.
- Others counter that AI can greatly help understand legacy systems (summaries, cross‑references) and refactor, making maintenance easier if used with discipline.
Organizational incentives and nihilism
- Several comments describe environments where executives mandate “AI everywhere,” use AI-written docs, tickets, and memos, and measure output via superficial KPIs.
- This encourages “slopmaxxing”: shipping unchecked AI output, deferring responsibility to tools, and treating jobs as temporary grifts before inevitable layoffs.
- AI is seen as a convenient way to launder commitment, responsibility, and credit: artifacts look polished but lack clear human intent or accountable authors.
Loss of understanding, learning, and mental models
- Widespread worry that constant reliance on agents weakens mental models: people debug by pasting into AI, not by reasoning, so they never internalize how systems work.
- Some engineers report onboarding and feature work feeling easy at first, then realizing months later they still don’t understand the codebase.
- Others argue motivated people can still learn deeply with AI as a tutor; lack of understanding is framed as a psychological and cultural problem, not a technical inevitability.
Changing roles and career paths
- Several predict a shift from “people who open pull requests” to hybrid roles closer to product management or “technical staff” guiding agents and systems.
- Fear that the traditional junior‑to‑senior pipeline (thousands of hours of hands‑on coding) will collapse, making future tech leads less grounded.
- Disagreement on whether this leads to more demand for high‑level system thinkers or to commoditization and job cuts.
Positive experiences and optimism
- Some report 10x productivity gains: rapid feature work, devops tasks, tests, and even complex changes to ugly legacy systems made tractable with AI.
- In domains with huge old codebases in obscure languages, AI is described as “god‑sent” for analysis and modernization.
- A few foresee AI making maintenance and refactoring cheaper, enabling continuous rewrites instead of living with deep technical debt.
Historical continuity and abstraction drift
- Multiple comments observe this is another step in a long trend: from assembly to C to high‑level languages to AI, each wave hides more complexity and shrinks the pool of people who understand lower layers.
- Some argue large organizations already had opaque decision‑making and fragile systems pre‑AI; AI mainly accelerates and exposes existing dysfunction rather than creating it from scratch.