Ask HN: What was your "oh shit" moment with GenAI?

Transformative practical use-cases

• Many describe “oh shit” moments when GenAI compressed weeks/months of work into hours:
  • Refactoring large codebases, migrating data topologies, porting apps (e.g., PL/SQL → Django, Oracle Apex → new stack, NES emulator JS → Rust).
  • Automatically generating infrastructure (Terraform/HCL, AWS setups, multi-service dashboards).
  • Reverse-engineering firmware, proprietary protocols, and old hardware (musical gear, printers, pianos, oscilloscopes, Firestick, test equipment) via Ghidra and tooling.
  • Turning piles of legal/financial/technical PDFs into OCR’d, translated, normalized, and summarized outputs.
  • Building end‑to‑end apps (web, mobile, trading bots, game mods, Git clients) almost entirely via agents.

Coding and software engineering

• Strong enthusiasm: agents that read entire repos, debug race conditions, diff logs, decompile VSIX/.NET, generate tests, perform code reviews, and maintain complex architectures.
• Some report agents as “junior devs” that can autonomously explore, refactor, and even create auxiliary tools (activity trackers, semantic search).
• Others find them frustrating: verbose or naïve designs, poor large‑scale architecture, fragile maintenance in big legacy codebases, and “AI slop” PRs and unit tests that only game coverage.

Non‑technical and everyday uses

• Frequent “wow” moments from:
  • Diagnosing HVAC, dryers, furnaces, koi pond pumps, and car/AC issues from photos/videos and wiring diagrams.
  • Planning trips, solar installations, home energy dashboards, kitchen builds, estate paperwork, and real‑estate or home‑inspection style reviews from listing photos.
  • Tutoring in cooking, drawing, math, anatomy, and science experiments with kids.
  • Creative tasks: poetry translation with rhyme/meter, songwriting, art, podcasts, interactive fiction, and stylized real‑estate listings.

Skepticism, limitations, and failures

• Some never had a positive moment; see LLMs as “fancy autocomplete” or toys that don’t answer “genuinely interesting” or novel questions.
• Recurrent issues:
  • Hallucinated APIs, CLI flags, PowerShell cmdlets, scientific interpretations, and niche-domain analyses (geophysics, remote sensing, specialized math).
  • Inconsistent answers to simple factual/logical questions, difficulty with high‑level vision strategy despite good low‑level code.
  • Models that seemed to regress (e.g., later versions worse at solution design than earlier ones).

Safety, ethics, and misuse

• Concerns about:
  • Dangerous advice (e.g., furnace troubleshooting with CO‑risk implications).
  • Legal and medical use without professionals; lack of privilege for AI chats; courts already seeing AI‑generated filings.
  • Agents that can weaponize exploits, generate license keys, crack captchas, cheat in games, or delete databases.
  • Reverse‑engineering and firmware patching at scale, and potential for deepfake porn, misinformation, and “AI deaths” that would be hard to trace.

Impact on work, learning, and society

• Split sentiment:
  • Some feel like “superheroes” or “CTO with a team,” changing how they think about problems and focusing on higher‑level design.
  • Others fear deskilling, degraded code reviews, stunted junior dev growth, and being displaced or reduced to “prompt jockeys.”
• Debate over hype:
  • Claims that GenAI is overpromised, enshittifying the web and tools, driving opaque data‑center expansion, and fueling irrational corporate spending.
  • Counter‑claims that we are still early; capabilities are jagged (excellent at some things, bad at others), and prompting/tool‑use is a real skill.

Meta: prompting, tooling, and hype cycle

• Many note that:
  • Results depend heavily on clear problem specification, good harnesses (editors, MCP/tools, CLAUDE.md‑style instructions), and domain expertise to validate outputs.
  • People are often poor at articulating what they want; some approach AI with bias or quit after first failure.
  • The hype produces alternating “oh wow” and “nothing changed” cycles, leaving some exhausted but others experimenting constantly.