One Month Without AI

Original Article ↗ Hacker News Discussion ↗

AI coding assistants and agentic tools are enabling developers to ship more features faster, but many engineers report feeling their skills, code quality, and architectural control eroding as they lean on models to generate large, complex changes. Commenters describe a trade-off between short‑term velocity and long‑term maintainability, with issues like bloated code, harder reviews, fragile workflows and atrophying expertise appearing alongside genuine gains in debugging, boilerplate generation and learning unfamiliar stacks. The emerging consensus is that AI can be powerful when tightly constrained and well‑integrated into existing practices, but unmanaged use—especially for end‑to‑end feature implementation—risks turning codebases into opaque, brittle systems that even their nominal owners no longer fully understand.

Scope of AI Use in Coding

  • Some describe near-total “vibe-coding”: no manual typing or reading code for months; features (e.g., 2FA, games, complex apps) are specified in natural language and implemented by agents end‑to‑end.
  • Others use AI narrowly: boilerplate, refactors, unit tests, small scripts, or as an advanced search/debug assistant, while keeping core implementation manual.
  • A subset of developers intentionally avoid AI for new projects to regain control and stay sharp.

Quality, Security, and Verification

  • Critics argue blindly trusting AI for security‑sensitive features (like 2FA) is irresponsible; you’re still accountable even if “cyber models” are good at pentesting.
  • Several report reviewing AI-generated systems and finding subtle, serious bugs (asynchronous race conditions, missing backpressure, silent error suppression).
  • Proposed mitigation: stronger static analysis, richer type systems, formal verification, and treating tests as only one layer of proof.

Productivity, WIP, and Complexity

  • Some teams claim order‑of‑magnitude speedups and rapid delivery of long‑delayed features.
  • Others find that agents generate huge PRs and documentation walls, creating review bottlenecks and encouraging hand‑waving.
  • One team’s “go hard on agents” experiment led to high WIP, stalled integrations, and unhappy stakeholders; they reverted to strict WIP limits and more human collaboration.
  • Commenters stress that AI tends to increase complexity unless tightly constrained.

Skill Atrophy and Abstraction

  • Many report feeling their implementation skills erode quickly; brief “no‑AI” sessions can be anxiety‑inducing but also reassuring.
  • Some see this as natural abstraction drift (like moving from assembly to high‑level languages); others insist you can’t design well without actively coding.

When AI Works Well

  • Debugging complex production issues by correlating logs, code, and weird edge cases.
  • Learning unfamiliar stacks (e.g., BLE, asyncio, MicroPython) faster than with today’s SEO‑cluttered web search.
  • Generating unit tests, small tools, and scaffolding where code quality and long‑term maintainability matter less.

Agentic Development and Control Loss

  • Agent swarms tempt people into multitasking and deferring review; tasks that could be hand‑coded in 20 minutes become multi‑day review exercises.
  • Some admit they no longer fully understand what a large AI‑written codebase does and rely on end‑to‑end testing and “duck tests” (does it appear to work?).

Writing and Non-code Uses

  • Book‑writing experiments with LLMs yield fast drafts but generic style, logical inconsistencies, cliché‑ridden prose, and months of cleanup—often not worth it.
  • Several say models have gotten better at code but worse or blander at long‑form prose.

Model Behavior and Tooling Variance

  • Experiences vary sharply: some see frontier models as “better than most devs”; others get hallucinations, wrong file usage, miscounts, and invented variables.
  • Quality depends heavily on model version, provider, and integration (chat vs IDE/agent with tools); some suspect silent downgrades or quantization, others dispute this.

Cultural Attitudes and Analogies

  • Analogies range from “AI as cocaine/slot machine” to “AI as autopilot or CNC machine”; some see moral panic, others see genuine addiction‑like overuse.
  • There’s visible polarization between “use it or be left behind” advocates and those worried about slop, dependency, and long‑term maintainability.