We should be more tired than the model
AI coding assistants and “agentic” workflows are reshaping how software is written, boosting output but often eroding developers’ understanding of their own code. Commenters describe a growing tension between short‑term productivity gains and long‑term skill atrophy, code quality, and job security, especially as companies push for faster delivery without sharing the benefits. Many advocate using LLMs as scaffolding, refactoring helpers, or Socratic tutors rather than full automators, arguing that humans must stay mentally engaged — and even “more tired than the model” — to retain control, taste, and accountability.
LLMs as learning aids and documentation helpers
- Several use LLMs to surface relevant docs, deep references, and practice problems, but insist on checking originals.
- People ask if LLMs are “OK” for theory; responses suggest coupling them with testable exercises so claims can be verified.
Comprehension, abstraction, and “being more tired than the model”
- Central concern: bottleneck is understanding, not typing. Agentic workflows can flood you with code you don’t grok.
- Some report more cognitive exhaustion: less mechanical typing, more decision-making and code vetting.
- Debate over whether LLMs provide “real” abstraction: they raise the level of expression (English→code) but remain non‑deterministic, which clashes with traditional notions of abstraction.
Productivity, fatigue, and quality
- Reported gains range from modest (~1.25x) to large (up to ~4x output) but often with increased mental load.
- Many see no corresponding jump in quality; some think overall product/code quality may be declining despite huge infra spend.
- Others say delegating low-level implementation frees attention for design, complexity analysis, and product thinking.
Tooling, refactoring, and agents vs IDEs
- Big subthread on using agents for refactoring vs classic IDE/LSP refactors.
- Critics argue many AI refactor prompts duplicate mature, deterministic IDE features that are faster and free.
- Proponents counter that agents can orchestrate multi-step changes (tests, docs, cross-file edits) from a single prompt.
Labor, incentives, and job security
- Strong skepticism that productivity gains will benefit workers absent unions or political action; expectation is more monitoring, pressure, and layoffs.
- Others foresee smaller, more LLM-assisted teams rather than mass job loss, but this is contested.
Workflows to retain understanding and skills
- Strategies: deliberate refactoring sessions, Socratic quizzing by agents, strict review gates, limiting when to call the agent, or not using LLMs at all.
- Some treat skill decay as acceptable if “taste” and high-level problem-solving remain; others fear long-term atrophy and dependency.
UX, flow, and future of coding tools
- Current “chat + slot machine” UX is seen as anti-flow and addictive.
- Desired direction: tighter IDE integration, context-aware suggestions, and tools that keep humans in the driver’s seat rather than fully agentic coding.