Major outages across ChatGPT and API

Original Article ↗ Hacker News Discussion ↗

A major ChatGPT and OpenAI API outage triggered wide concern about how quickly developers, companies, and individuals have come to depend on a single AI provider for everyday work. Commenters trade fallback options—from Azure OpenAI, Anthropic and Hugging Face models to fully local LLMs—and note practical issues like incompatible embeddings, lack of SLAs, and immature support. Many see this as a warning to design for redundancy and portability now, while others argue that the productivity gains from state‑of‑the‑art hosted models still outweigh the reliability and lock‑in risks.

Outage impact and reliance

  • Many commenters report being unable to work or significantly slowed, having replaced Google/Stack Overflow with ChatGPT for coding, scripting, documentation, and writing.
  • Some treat it humorously (“PTO day”, “training wheels/crutch”), but several admit they hadn’t realized how dependent they’d become.
  • Others say they’re unaffected because they don’t use ChatGPT or still “store knowledge in their head,” sometimes pushing back on over‑reliance.

Alternatives and failover strategies

  • People mention using: Azure OpenAI (largely unaffected), Anthropic Claude, Bard, Kagi’s GPT‑4 via Azure, Phind, You.com, Hugging Face Spaces (e.g., Zephyr), and local setups (Code Llama, Mistral, dolphin‑mistral, Phind‑CodeLlama).
  • For embeddings, suggestions include: Azure OpenAI, Amazon Bedrock, SBERT, Instructor, and storing multiple embedding types per document to enable switching.
  • Some products already failed over to Anthropic or other models; others note this is hard where OpenAI‑specific features (tools/function calling) are used.

Local and open‑source models

  • Strong interest in self‑hosting to avoid outages and platform risk.
  • Consensus that current open models are improving but still not at GPT‑4 quality; good enough for some tasks (summarization, simpler coding, RAG) but not as “general‑purpose.”
  • Hardware cost and scarcity (e.g., H100s) are major barriers; smaller models can run well on consumer GPUs or even CPU for some.

Reliability, SLAs, and enterprise concerns

  • OpenAI provides no meaningful SLA; several recount very poor support experiences and months‑long unresolved issues.
  • Some enterprises are moving to Azure OpenAI specifically for better reliability, SLAs, and support.
  • Others argue this is a normal fast‑growth phase; SLAs and robustness will improve, but lock‑in risk and outage planning are under‑appreciated.

Model quality, censorship, and behavior

  • Mixed views on new GPT‑4 Turbo: cheaper and faster but possibly slightly worse on some NLP tasks; some report variability even at low temperature.
  • Comparisons: GPT‑4 often seen as best overall; Bard perceived as weaker at coding and more aggressively filtered; Claude praised as a solid backup.
  • Concerns about over‑strict content filters (especially violence/war topics) and about hallucinations across all models.

Broader reflections

  • Many see the outage as a warning about centralizing critical workflows on a single AI API.
  • Some predict a future of embedded/on‑prem models and multi‑provider abstraction layers; others note the risk of long‑term dependence and skill atrophy, especially for junior developers.