Stack Overflow’s forum is dead but the company’s still kicking

Original Article ↗ Hacker News Discussion ↗

Stack Overflow’s collapse in new questions—from hundreds of thousands a month to just a few thousand—is widely attributed to a mix of years of heavy-handed moderation and the sudden convenience of AI coding tools. Commenters describe how strict duplicate policing, unfriendly gatekeeping and outdated top answers had already driven many users away before large language models finished the job. While LLMs now outperform SO for quick help, people worry about losing a public, community-vetted knowledge base—and about where future models will get high-quality, up-to-date programming data once sites like Stack Overflow and Reddit wither.

Stack Overflow’s Decline and Culture

  • Many participants report SO had become hostile and adversarial years before LLMs: nitpicking, downvotes, instant closures, and “XY problem” accusations.
  • Early SO is remembered as fun, friendly, and helpful; later it’s described as draconian, over‑moderated, and optimized for “tidiness” and Google, not for helping askers.
  • Strict duplicate-closing is a recurring complaint: questions closed as “dupes” of older, different-tech answers, often obsolete (e.g., old framework or Python 2 vs 3).
  • Others argue strict moderation is precisely what made SO high quality and prevented it from becoming a “dumpster fire.”

Moderators, Gamification, and User Experience

  • Gamification and moderator power are blamed for attracting rule‑obsessed users who edited or closed posts harshly, sometimes even rewriting others’ wording.
  • Answerers describe burnout from wading through low-effort or duplicate questions; many simply stopped answering.
  • Some found the strict question template helpful as “rubber duck debugging,” but say the later culture made posting traumatic.

LLMs as Cause and Consequence

  • Data cited: questions per month fell from ~300k at peak (2020) to ~3k in 2026; some are shocked how close to zero it got.
  • Many now default to LLMs for coding help; even imperfect models are “good enough” and far less abrasive.
  • Multiple comments note that LLMs were trained heavily on SO (and similar sites), so the “less abrasive alternative” rests on that earlier human labor.
  • Concern: if public Q&A dries up, what will future models train on, especially for new technologies and undocumented “gotchas”?

Knowledge Quality, Trust, and Future Data

  • SO’s value: canonical questions, multiple competing answers, comments, and long-tail solutions that LLMs often miss or average away.
  • Worries that AI-generated docs and “slop” will fill the web, causing self‑reinforcing degradation when models train on their own output.
  • Some foresee agents learning from code, docs, usage telemetry, RL, and synthetic data; others doubt this replaces human-discovered edge cases.

Broader Ecosystem and What’s Lost

  • Other StackExchange sites (math, stats, smaller topics) are seen as friendlier but also in decline.
  • Reddit is cited as undergoing its own “near-death” via bots and low‑quality engagement.
  • Many feel we’ve lost a unique, public, community‑validated corpus and a human learning space, even if SO’s culture had become deeply flawed.