Elevated error rate across multiple models

Original Article ↗ Hacker News Discussion ↗

Outage symptoms and status page

  • Users report widespread 500, 503, and 529 (“overloaded”) errors from Claude Code and claude.ai; some sessions work while others time out, suggesting uneven infra impact.
  • Status page shows many red/orange incidents (“Christmas ornaments” / “full three pepper blend”), with dynamic time windows changing reported uptime based on viewport width.

Uptime metrics and perceived impact

  • Published uptimes: ~99.1% for claude.ai, ~99.3%+ for “Claude for Government”, lower for Claude Code.
  • Several argue that <99.9% is poor for a core work tool, especially when outages cluster in US/EU work hours; others say even with downtime, the productivity boost is worth it.
  • Some want “uptime during working hours in my timezone” instead of aggregate SLAs.

Reliance on AI tools vs baseline skills

  • Many admit strong dependency on AI/Internet for real-world software work; others insist they can still develop offline, just slower.
  • Concern that future devs may lose ability to code or maintain large doc/codebases without LLMs.
  • Counterpoint: when tools are down, teams “fall back to baseline,” so long-run productivity is still net positive.

Alternatives, harnesses, and ecosystem complexity

  • Multiple suggestions to try pi, OpenCode, Codex, GLM‑5.2, DeepSeek, Zed, pool, oh-my-pi, openrouter, etc.
  • Experiences are mixed: some see pi as flexible and extensible; others report huge token burn and latency compared to other harnesses on the same model.
  • Many now juggle several models/harnesses plus emerging standards (MCP, ACP), noting growing complexity and fast-moving tooling.

Security and curl | sh installer debate

  • Long subthread debates piping installer scripts from the web into a shell.
  • Critics call it irresponsible and a supply-chain risk; proponents say it’s ubiquitous, convenient, and often just a wrapper around package managers.
  • Discussion touches on trust models, package managers vs ad‑hoc scripts, and whether this practice “trains” users into unsafe behavior.

Root causes: infra vs “AI-coded” systems

  • Some blame Anthropic’s heavy internal use of agentic loops and AI-authored code, arguing outages expose limits of “Claude-oriented programming.”
  • Others reject this as a “blame AI for everything” fallacy, pointing out that large-scale infra has always been hard and could be failing for many non-AI reasons.
  • Debate over whether these incidents show cultural/engineering problems or simply rapid growth and immature infra; no clear evidence either way.

Nature and limits of LLMs

  • Several developers describe LLMs as “probabilistic random tables on steroids”: great for ideas, debugging help, and rubber-ducking; unreliable as autonomous coders.
  • Examples given of absurd-but-confident code (string-based hex checks instead of integer comparisons), reinforcing the need for strict human review and guardrails.
  • Some argue non-determinism and hallucination are inherent; workflows must use narrow prompts, short horizons, and layered checks to be reliable.

Pricing, lock‑in, and future scenarios

  • Speculation about vendors raising prices once companies are dependent, with worries about multi‑year enterprise lock‑in.
  • Others note competition (OpenAI, open models, multiple harnesses) and expect sensible orgs to multi‑home rather than depend on a single provider.
  • Some foresee high per‑seat AI costs still being cheaper than human developers for many tasks, but doubt current tools fully replace engineers.

Cultural and humorous reactions

  • Many treat the outage as an enforced break: “touch grass,” holidays, doing planning, or coding “like in the 90s.”
  • Jokes about “Claude for Government” having better uptime, AI data centers heating the planet, internal “backup Claude” humans, and Anthropic using Gemini or DeepSeek to fix Claude.
  • Underneath the humor is anxiety about growing dependence on brittle cloud AI services.