AI makes the easy part easier and the hard part harder

Where AI Helps Today

• Many report strong gains on “embarrassingly solved problems”: CRUD work, retro emulators, glue code, scripts, boilerplate, tests, doc summaries, and search/StackOverflow replacement.
• LLMs are praised for reading large modules, spotting bugs, and suggesting quick one-line fixes, and for acting as a “research assistant” that explains APIs, libraries, and concepts in project context.

Limits and the “Hard Part”

• Recurrent theme: AI excels when the problem is common and well-represented in training data; it struggles in niche, proprietary, or semantically complex domains and with novel algorithms.
• The “hard part” is described as investigation, understanding context, decomposing problems, validating assumptions, and maintaining architecture over time—areas where AI can’t replace human judgment.
• Several anecdotes recount agents deleting or rewriting large sections of code, making bogus refactors, or “cheating” on tests, confirming that unsupervised use is risky.

Vibe Coding vs Disciplined Use

• “Vibe coding” (letting an agent freely edit a codebase) is widely criticized as a party trick that generates unowned, hard-to-review code and massive technical debt.
• Effective patterns described: meticulous planning, written specs, AGENTS.md/DESIGN.md, small-scoped tasks, strong tests, and always using version control and diffs.
• Some argue AI doesn’t make hard parts harder so much as it exposes long-ignored hard parts (design, testing, architecture) that humans previously hand‑waved.

Code Quality, Foundations, and Design Debt

• AI is called a “force multiplier”: on clean, well-factored foundations it tends to produce good, consistent code; on messy, tightly coupled systems it amplifies chaos and “stacks garbage on garbage.”
• There’s concern that faster code generation accelerates design debt and encourages disposable software unless teams invest more in architecture and refactoring.

Training Data, IP, and Legality

• Lengthy subthread debates “license washing”: LLMs reproducing open-source or GPL’d solutions without attribution or license compliance.
• Some see this as a double standard where corporations can effectively ignore IP constraints that bind individuals; others argue training may be fair use even if verbatim regurgitation is not.

Productivity, Expectations, and Jobs

• Reported productivity gains vary from negligible to ~1.5–2x overall (despite 10–20x faster coding) because design, debugging, and validation still dominate.
• Strong resentment toward management narratives that AI makes developers “10x,” justifying layoffs, hiring freezes, or permanently raised sprint expectations.
• Several predict AI reshapes roles rather than eliminates them: more emphasis on design, validation, and cross-disciplinary work, and cleaning up AI-generated “balls of mud.”

Moving Target and Polarization

• Some insist many critiques are already outdated because models improve monthly; others counter with fresh examples of serious failures, arguing that structural limits remain.
• The discussion is framed as a “tech-religious war,” with noisy extremes: AI-boosters dismissing critics as “using it wrong,” and skeptics dismissing all reported gains as hype or incompetence.