GitHub Outage Map

Community Discussion

Tips? Frustrations? Share them here. Useful comments include a description of the problem, city and postal code.

Beware of "support numbers" or "recovery" accounts that might be posted below. Make sure to report and downvote those comments. Avoid posting your personal information.

GitHub Issues Reports

Latest outage, problems and issue reports in social media:

• 
  Polsia (@polsia) reported 54 seconds ago

  Good repos rot from neglect, not bad code. Corrick is the caretaker that watches your GitHub around the clock — triaging issues, tagging duplicates, bumping outdated deps, opening scoped fix PRs behind CI. Narrow on purpose. The full version is coming.

• 
  samagra14 (@samagra_sharma) reported 2 minutes ago

  Spent almost two hours with the new @grok Grok Bot. It has, for sure, a glimpse into the future, very similar to what @perplexity_ai has been trying to do with their computer agents or Perplexity computer. My only frustration is that it is an incomplete release. They market them as 24/7 AI agents, but they have no sense of identity and no sense of provisioning actual resources under their name. They should be able to provision email, they should be able to have a GitHub account and all those things. They want access to my accounts, and they want access to do things via me. I think always on is definitely an important aspect when it comes to solving mathematical problems and throwing tokens to actually figure out random ****. Until and unless provisioning an identity is a core feature of whoever is selling these 24/7 AI co-workers, I don't think they'll make it.

• 
  Nitesh (@NiteshTechAI) reported 3 minutes ago

  Nobody opens the browser tab your AI assistant lives in. Mine lives in Discord. It answers from a 2,800 page knowledge base on my phone, and that one change is why I use it daily instead of monthly. It's called AstrBot. • Agent sandbox for safer tool use. • 1000+ plugins, one click to install. • Knowledge base, personas, MCP support. • Automatic context compression on long chats. • QQ, Telegram, Slack, Feishu, DingTalk, WeChat Work. Adoption is the problem here, not capability. Put the agent where you already type all day and using it stops being a decision you have to make each time. One caveat before you build on it: AGPL-3.0, not MIT. Read the terms if this is going anywhere commercial. ⭐ 38,000+ stars on GitHub. AGPL-3.0 licensed. 🔗 GitHub link in the comments 👇

• 
  Polsia (@polsia) reported 5 minutes ago

  Your repo is fine. Your deps aren't. Patchpress is an always-on AI agent for GitHub: bumps outdated packages, fixes lint and test fallout, drafts CVE patches, ships a morning digest. Sandboxed. Issue text treated as untrusted. SBOM on every change. Live soon.

• 
  Darek Gusto (@darekgusto) reported 8 minutes ago

  Tbh, at this level of hype around the upcoming today's Cursor BIG NEWS, I'm bracing myself for a disappointment. But I still really want them to deliver awesomeness today. Is it a new Composer model? Is it Origin throwing down the gaunlet at Github? Or something entirely new? With how much they hype it, preferably all three. Hopefully not just the newly announced Grok Bot...

• 
  Thomas Peitz (@tpeitz_dus) reported 9 minutes ago

  Since I have github issue -> PR flow (claude action) - I am a lot more creative. I just let it build and when I am bored hours later I just check what it did. - This even makes more fun with my game which I am programming. - I can just play a round and tune further.

• 
  Raunak (@capsraunak) reported 9 minutes ago

  A 21 year old intern fixed a bug that 3 senior developers couldn’t solve for 5 months. She saved the company $2.2M . they gave her a $400 gift card and didn’t convert her to full-time. she posted the solution on github. A startup saw it hired her at $145k . The 3 senior developers who couldn’t fix it. They’re still there still getting paid still breaking things. Talent gets punished in corporate. Mediocrity gets protected.

• 
  shmidt (@shmidtqq) reported 9 minutes ago

  A 23-year-old from India taught a $25 CCTV camera to detect when his grandmother falls. A company selling the same thing for $3,000 sent him a letter from their lawyers His subscription: $4.99/month Subscribers: 1,847 Investment: $200 spent on Claude His grandmother fell in the kitchen at 6 in the morning and lay there for four hours. The $40 panic button was in the bedroom the whole time. So he took the camera already hanging over the stove and wrote a YOLO pipeline in Claude Code: 17 skeleton points, torso angle, drop velocity. A person stays down for 3 seconds, five relatives get a call. It all runs on an $80 laptop and the video never leaves the apartment. Everyone else was selling these systems to nursing homes. He sold it to the kids who moved away. In 14 months it caught 61 real falls. Accuracy is 94%, and he writes it plainly in the description: the other 6% is somebody's grandmother. The lawyers' letter came in July. That same night he pushed everything to GitHub under MIT and added one line at the top: now go sue everyone. The best products are not born from hunting for a niche. They are born from a fear you cannot fall asleep with

• 
  Leander (@leanderriefel) reported 10 minutes ago

  This is what I was scared of. If this really is how Origin works it is not a GitHub alternative. It solves an entirely different problem.

• 
  NativePHP (@nativephp) reported 10 minutes ago

  @efekpoguavik3 Could you raise this as a GitHub issue with minimal code examples or a reproduction repo so we can dig into this?

• 
  Sameer F. (@imrtls00) reported 10 minutes ago

  @thdxr @raulpacheco2k Maybe add a config setting that lets me control when to paste text as label and when as inline. Already listed as open issue on github.

• 
  Nathan Barry (@nathanrs) reported 17 minutes ago

  Turns out you can make LLM inference fully deterministic across devices, with no loss to quality or speed. This weekend at the @SpaceXAI hackathon I got Qwen3-0.6B to produce identical hashed logits from a 512-token generation across 2 GPUs and 3 CPUs: an A100, an H100, an Apple M5 Max, an AMD EPYC, and an Intel Xeon. The main reason why inference isn't deterministic is because floating-point addition is not associative. (a + b) + c ≠ a + (b + c), since every add rounds. Accumulation order changes with the hardware used and kernel selected, so the same prompt can give you different outputs even at temperature 0. Integers ARE associative. So why doesn't integer quantization already fix this? Because, while weights and activations get quantized, the non-linear ops (softmax, normalization, SiLU) dequantize back to float and requantize afterwards. Each of those steps hands you back to floating point rounding. True integer-only inference does exist, but it's historically been motivated by edge hardware without FPUs, which doesn't make much sense for LLMs. One 2024 paper (I-LLM) did it on LLaMA from that angle and didn't get much attention. Nobody seems to have looked at it from the determinism side. I wrote my own implementation, simplifying the approach from the paper, so that every operation between the input ids and the int32 logits is exact integer arithmetic. To test it I chain-hashed the logits at every step and ran that across the devices and configurations below. Every integer run gave the same hash: 64430dd985f8. Every fp16 run gave a different one, all diverging on the very first token. WikiText2 perplexity came out to 20.72 vs 20.95 for fp16 (slightly better than the float baseline), and CUDA-graphed integer decode hits 106 tok/s at batch 1 on an A100, 3.6x the fp16 eager baseline. Github repo is listed in the comments. Plan to do a writeup over this eventually!

• 
  0xdoppo (@0xdoppo) reported 17 minutes ago

  Well fk I need to spend more time talking to VCs and doing BD outreach for Peanut Shopper. Then I find another OCR failure and disappear for hours. My GitHub doesn't even properly reflect the excuse either. Most of the loop is DB driven: ingest thousands of flyers → normalize → isolate failure patterns → adjust normalization → rebuild dataset → measure what improved and what re-broke. Annoyingly, OCR isn't really the problem. Mapping meaning across wildly different flyer formats and neighboring regions is. New formats sprout up constantly violating whatever assumptions worked before, so the problem stays dynamic. So yeh, I should probably be talking to investors and retailers more. Instead I'm playing with the data.

• 
  Diam (@diamai_) reported 20 minutes ago

  Connect an AI agent to five MCP servers, and it can spend 55,000 tokens before it reads the task, loading instructions for the tools they expose. GitHub, Slack, incident alerts, dashboards, and other software can arrive as a giant menu before the agent knows the job. More of its attention goes to that information, less to the request. Anthropic's Applied AI team ran into a different version of the problem. Sonnet 4.5 would start wrapping up a job before its context window was full. They added reset logic so it could continue instead of stopping early. Opus 4.5 no longer behaved that way. The reset became pure overhead, adding delay and sometimes discarding cache the system could still use. In a new AI Engineer talk, Gagan Bhat and Isabella Kai He explain the architecture behind Claude Managed Agents. The full record of the job sits outside the agent's active memory. The system brings back only the detail needed now. - 7:44 When a reset helps and when it becomes a drag - 14:32 Keeping old work without carrying it all - 25:58 Giving an agent private access to company tools Watch the video, then read the attached article. It is about what happens when an agent gets every available tool before it knows which one it needs.

• 
  Dariusz (@die54minute) reported 22 minutes ago

  Spent an hour convinced GitHub Copilot CLI was sandboxed away from my Azure CLI creds, but it wasn't. Copilot sets COPILOT_AGENT_SESSION_ID. Azure CLI 2.88+ sees that, bypasses your human MSAL cache, and tries to mint an agent-tagged token for governance (Defender/Purview/CA). My tenant blocked it (AADSTS53003). Fallback: browser login. Inside an agent. This sucked. Same machine, same ~/.azure. Claude Code worked because it never declared itself an agent. Fix: env -u COPILOT_AGENT_SESSION_ID -u COPILOT_CLI az boards ... This design is intentional from Microsoft, I understand they want to distinguish humans from agents. Unusable if you don't control the Azure DevOps Policies.