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:

• 
  SPARQIO (@sparqio) reported 55 seconds ago

  AI has moved from research curiosity to core infrastructure. Search engines, medical tools, financial platforms, enterprise software, all running on models that can sound completely confident while being completely wrong. That tension is the central problem nobody has fully solved yet. Before you can measure whether an AI is correct, you need to define what correctness actually means. It is not one thing. A response can be factually accurate but contextually useless. Logically coherent but dangerously incomplete. Precisely worded but subtly misleading. Practitioners who collapse all of this into a single quality score are building on sand. The more useful frame is five separate dimensions: factual accuracy, logical coherence, contextual relevance, completeness, and calibrated confidence. Each one requires a different evaluation approach. A model that scores well on fluency and coherence can still be catastrophically wrong on facts, and the score will never tell you. On the automated side, the oldest tools (BLEU, ROUGE, METEOR) measure lexical overlap against a reference answer. They have real uses in translation and summarization, but they are poor proxies for whether something is actually true. A model can paraphrase a wrong answer fluently and pass every metric. The field has moved toward embedding-based similarity and model-as-judge setups. BERTScore captures semantic equivalence rather than word matching. More recently, using a separate powerful model to score outputs against structured rubrics, assessing factuality, completeness, and reasoning quality, has become a serious evaluation paradigm. Benchmark datasets add another layer. TruthfulQA tests whether models give truthful answers to questions that humans typically get wrong due to common misconceptions. MMLU spans 57 academic domains. HaluEval is built specifically for hallucination detection. $AI-adjacent plays in the coding space might care about SWE-Bench, which evaluates code generation by running outputs against real test cases from actual GitHub issues. But generic benchmarks hide a serious trap. A model that performs well across general knowledge can still fail badly in specialized domains. Medical AI needs evaluation against clinical reasoning datasets like MedQA or PubMedQA. Legal AI needs BarExam-style benchmarks. Financial AI needs FinQA. Deploying a model because it passed a general benchmark, then using it in a high-stakes domain, is a risk management failure, not an engineering decision. Human evaluation still cannot be replaced, not fully. Automated systems miss errors of omission. They miss misleading framing. They miss the kind of subtle wrongness that a trained clinician, lawyer, or financial analyst would catch immediately. Structured annotation protocols with qualified reviewers remain the gold standard in any high-stakes deployment context. The honest takeaway: knowing when AI is telling the truth requires combining all of these layers. No single metric, benchmark, or review process is sufficient on its own. Organizations treating AI correctness as a solved problem are the ones most likely to discover otherwise at the worst possible time.

• 
  It's FOSS (@Itsfoss) reported 4 minutes ago

  Debian is voting on AI contributions. Codeberg banned vibe-coded projects entirely. Flathub already said no to AI generated apps. But blanket ban is not a solution...and to be honest, you cannot reliably detect AI-written code. And more importantly, almost every developer working these days is using AI assistance in some form. GitHub Copilot. Cursor. Claude. Codex. They use it to autocomplete functions, generate boilerplates, debug tricky logic, write tests. Does that make their code AI-generated? Or human-written with AI tools? The real problem Flathub and Codeberg are trying to solve is quality. Apps with no real functionality that add little value. And then the "vibe coded and abandoned" pattern that floods platforms with "abandonware" apps. That is a legitimate concern. But conflating it with "AI assistance" is going to create policies that punish thoughtful developers using AI responsibly, while doing nothing to stop bad actors who will just lie about it anyway. Debian's vote matters because it may set a precdence. The worst outcome is one that treats every developer who uses AI-assistance as a rule-breaker. So, where should we draw the line?

• 
  Suraj Sharma (@suraj_sharma14) reported 5 minutes ago

  So much of being a developer is overcoming fear. Fear of shipping imperfect code Fear of public failure on GitHub or X Fear of not knowing enough (imposter syndrome) Fear of building something nobody uses Fear of asking for help in public Fear of charging money for your work Fear of technical debt accumulating Fear of being judged by senior engineers Fear of pivoting from cool tech to boring problems Fear of production failures at 2 AM What are you avoiding because of fear? Do not let it control your craft. Ship the code. Ask the question. Charge the price. Most developers wait for permission. Builders ship and learn.

• 
  Rahul Verma (@RahulVerma989) reported 6 minutes ago

  @ClaudeDevs @AnthropicAI hey, any chance you guys can look into the MCP server stability? It’s been driving me crazy lately. 😐 I’ve got the server toggled on for my sessions, but the tools just refuse to show up-only the GitHub ones seem to work. Pretty sure the issue isn't on my end, so could you take a peek?

• 
  Glenski 📷🇨🇦 (@glenegrant) reported 6 minutes ago

  @araseb_ I actually have both working together: Codex on macOS, Claude Code on Windows PC they work through GitHub issues on a /loop as we port applicaitons from Win to macOS. Kind od wild to witness.

• 
  Lavanya | DeFiDecoded (@lavanyalakshma2) reported 7 minutes ago

  Most people are chasing shiny chatbots but they are missing the real change. What is happening right now? Early on, building a small application for my needs became so hard. No OpenAI tools could get me to make it fast. Data slips, models get tainted, trust fades. All those issues have to be faced. I need to learn Solidity. That changed when I started working with @CNPYNetwork . Linking an AI app to its own custom blockchain fixes the core problem. Talks stay private, models stay clean, and users keep total control, for handling all no tech boss required. Get faster, smoother, and full ownership on my networks. The biggest game-changer is how fast you can build now: * Before Canopy: Learn Solidity -> Build validators -> Hunt for funding * After Canopy: Choose a template -> Connect GitHub -> Launch Locked ledgers give AI real staying power. Five years from now, every major AI application will run on its own chain. I’m building for that future every day. Are u still watching or start building? @CNPYNetwork

• 
  Open-source Projects (@the_osps) reported 7 minutes ago

  • Terminal-native development with no context switching required • GitHub integration for repositories, issues, and pull requests via natural language • MCP-powered extensibility with custom MCP server support

• 
  BuBBliK (@k1rallik) reported 10 minutes ago

  THEY DELETED IT. THE INTERNET DIDN'T Anthropic pulled the leaked Claude share links from Google this weekend. Feels like it's over. It isn't - the internet remembers everything you make public, even for a second. - ChatGPT had the identical leak in July 2025, shared conversations indexed and searchable - Grok leaked hundreds of thousands of transcripts in August 2025, some of it graphic - Claude already leaked once before this, in September 2025, with 600 chats indexed - Deindexing from Google removes a page from search, not from the internet Someone scraped the leak before the fix and archived it publicly on GitHub - and it's not just Claude chats sitting there. Grok's leaked conversations are archived right alongside them. Link in the comments.

• 
  Julian (@orithellama) reported 10 minutes ago

  @akilress @ICPXProtocol @SumiroStudio Tbh I have had one error on the one-click deploy from Github, will post an update once it's resolved ;)

• 
  Xander Dunn (@xanderai) reported 11 minutes ago

  @philipliao_ Clause Tag is a severe yapper, I removed it from most of our channels. Same issue we had with Claude on GitHub in the early days. I can absolutely see the sparkling future where agents in Slack acting as humans do make sense, but this isn’t quite there.

• 
  Alessandro Frau (@alessandro_a0) reported 13 minutes ago

  @thiojoe @christitustech @codex I push daily with Codex, if you set up *** and login with your account or just install GitHub Desktop on your machine, it can do that on your behalf, leave comments, analyze comments left from other agents on PRs, etc.

• 
  tidux (@skibidiblazor) reported 16 minutes ago

  @prestonjbyrne It's because the "American" administrators of GitHub are also Indian. This is a demographic problem, not a legal theory problem.

• 
  Viber (@Avery_Coree) reported 17 minutes ago

  🚨 security camera firmware ships with a github admin token baked into the login page. this is a grep before you ship problem, not a hard problem, found stuff like this in an afternoon before

• 
  Ed Grosvenor (@MaybeEdward) reported 25 minutes ago

  @thdxr Same. I don't worry about frontend slop at all. I have a /deslop-frontend command that works great. It opens a GitHub issue and assigns it to someone who knows what they're doing.

• 
  Yogi Suria (@suriadesign) reported 25 minutes ago

  Me and Fable solved the NEET leak problem. Papers leak between "printed" and "opened" — we made that gap 0 seconds. The paper's born at exam time. Nobody's going to believe this, so I open-sourced the whole thing on GitHub. Here's how 👇