The human-in-the-loop is tired
Engineers react to an essay on “human-in-the-loop” AI coding by describing how large language models can make them both more productive and more exhausted, shifting work from hands-on problem solving to supervising, reviewing, and herding opaque agentic workflows. Many lament a loss of craftsmanship, flow, and intrinsic rewards as coding becomes more like managing a slot machine or factory line, while others say careful, incremental use of tools like Claude reduces fatigue and lets them focus on higher-level design. Threaded through are worries about AI-generated “slop,” the sameness of modern apps, the irony of AI-written thinkpieces about AI burnout, and longer-term fears that development work may be devalued or reduced to a blame-bearing oversight role.
AI-written article debate
- Many commenters feel the article “reeks” of LLM style: corporate tone, repeated phrases like “the honest truth” / “that loss is real and worth naming,” em‑dash heavy, long-winded and generic.
- Others argue similar styles predated LLMs, note pre‑AI corporate/PM prose and CV conventions, and caution against over‑trusting AI detectors.
- A linked detector (Pangram) flags the text as AI; some defend its low false‑positive rate, others criticize its “supporting evidence” as misleading.
- Several readers say they’re too tired of “Claude-isms” to engage with the substance, which they otherwise find relatable.
Patterns of LLM-assisted coding
- Two main workflows emerge:
- “Assistant” mode: one chat, heavy upfront planning, incremental code generation, continuous review, often broken into stacked PRs.
- “Agent/vibe coding” mode: multiple parallel chats or agents churning large amounts of code, with later review.
- Many report the second style as chaotic, burnout‑inducing, and hard to review; the first is seen as slower but sustainable and more controllable.
- Some keep LLMs out of their editor, copying in/out snippets to maintain ownership and architectural understanding.
Burnout, reward, and loss of joy
- Strong resonance with the idea that LLMs automate the “fun” micro‑rewards (solving, designing, typing) and amplify the tiring parts (specifying, reviewing, debugging).
- Several compare LLM prompting to a slot machine or to being a novice again, guessing and retrying without clear causal understanding.
- Others say LLMs reduce fatigue: no more syntax grind, more focus on high‑level design, and a renewed “tool as imagination extender” feeling.
Code quality, “slop,” and craftsmanship
- Widespread anxiety about a flood of low‑quality, samey “slop” code and apps, often indistinguishable and uninteresting.
- Some accept ugly but functional LLM code, especially for boilerplate, arguing that user value and shipping speed matter more than elegance.
- Others mourn loss of craftsmanship and intrinsic satisfaction: “I told the model to do it” feels less meaningful than “I built this.”
- Concerns that steering agents and reviewing walls of prose/code is cognitively harder than writing well‑structured code directly.
Work intensity, labor, and responsibility
- Multiple comments argue AI increases work intensity: more PRs, more context switching, more artifacts (plans, docs, stories) generated by AI to vet.
- The “human in the loop” is reframed as “human on the hook”: responsible when things go wrong, with little credit when they go right.
- Fears that engineering may shift toward a lower‑skill, factory‑like role (guiding tools rather than deeply understanding systems), with potential downward pressure on wages.
- Skepticism that productivity gains will translate into shorter workweeks; instead, some foresee layoffs plus overwork for those remaining.
Diverging developer archetypes and reactions
- Commenters note that LLMs amplify differences:
- Those who love fast shipping and high‑level system thinking often enjoy LLMs.
- Those who love deep understanding, careful architecture, and “flow” in writing code often feel alienated or exhausted.
- Some find solitary LLM‑driven building nostalgic and joyful; others experience loneliness, sameness, and a sense that the internet and codebases are being homogenized.