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:

• 
  Tech Tech China (@techtechchina) reported 1 minute ago

  The "open = dangerous" framing dies the moment you look at who ships open weights at the frontier: DeepSeek, Qwen, Kimi. Washington can't sanction a GitHub repo. Treating openness as the risk variable is a category error — the model doesn't care about its license.

• 
  AuxApex (@AuxApex) reported 2 minutes ago

  @Angeluz01 Anyone else found the Dramatic Shape GitHub is down?

• 
  Rahul Raj (@raahulll_raj) reported 4 minutes ago

  Decentralised AI is quietly revamping how we train models. Most people haven't noticed. nous research has 227,000 github stars and a $1.5b valuation. most people still file it under "that solana AI thing." here's what it actually is, 1. the problem they picked training a big model normally needs thousands of GPUs in one building, wired together with data centre grade cable. maybe five companies on earth can afford that. nous asked: what if the GPUs are scattered across the world, on normal internet? but pushing training updates over home broadband is roughly a thousand times slower than inside a data centre. 2. DisTrO is the fix it squeezes what each machine has to send to the others down by orders of magnitude. that one compression trick is the whole company. everything else sits on top of it. 3. psyche is the network psyche coordinates the scattered machines. the coordination layer runs on solana, four public programs: coordinator, authorizer, treasurer, mining pool. so the "crypto part" isn't a token. it's the scheduler. 4. consilience proved it works 40.2 billion parameters. around 20 trillion tokens. widely reported as the largest AI pre-training run ever done over the public internet. sized on purpose so it trains on one server and runs inference on a consumer 3090. 5. hermes is what you can actually touch hermes 4.3 was the first model trained start to finish on psyche. 144,000 tokens per second across 24 nodes. and hermes agent, their open source agent, sits at ~227,000 stars and ~44,000 forks. MIT licensed, so you can fork the whole thing. nvidia picked it as the reference runtime for nemotron 3 ultra. 6. where the money comes from nous portal. one subscription, 300+ models, bundled tools, $20 to $200 a month. model free, infrastructure paid. the red hat playbook. decentralised training is still slower and pricier per unit of compute than a data centre. the gap is closing, not closed. nous is also VC owned, not community owned. ~$70m raised, now closing ~$75m more at $1.5b led by robot ventures with USV in. no token. no onchain governance. so "decentralised AI" is half true. the training is decentralised. the company is not. NFA. DYOR.

• 
  Offensive Lab (@OffensiveLab) reported 7 minutes ago

  North Korean state hackers are no longer content with simply typing commands into public chat bots. One of the country's major intelligence groups has begun running artificial intelligence (AI) offline on its own servers, connecting document search tools to the files it possesses, and starting to collect the software components needed to build AI into its malware. South Korean security firm Genians says it uncovered this setup after months of tracking and log analysis of infrastructure linked to a hacking unit, Kimsuky, under the North Korean Reconnaissance General Bureau. Genians found no evidence that the group had trained its own AI model, and the firm does not present this as a certainty. Instead, it describes the actor as being in a "research and knowledge acquisition" phase, collecting and testing existing tools rather than creating new models, with the clear goal of incorporating AI into operations, from writing malware to analyzing data. For an intelligence unit that has spent years targeting government, research, and other strategic entities, this suggests a shift towards attacks that are more sophisticated and difficult to detect. There's nothing here that's particularly novel, the weight falls on the defenders. Once the AI starts writing, it reveals a vulnerability they once relied on: machine translation, clumsy formatting, and spelling errors. What the machine inserts becomes a tell. Genians' report tells defenders to look for LNK execution, PowerShell, hidden scheduled tasks, GitHub traffic, and payload activity, primarily judging how "flashy" it appears, rather than relying on more subtle indicators.

• 
  RuntimeWire 🏴‍☠️ (@runtimewire) reported 8 minutes ago

  @uriel_bitton 35 for 35 on catching issues is impressive. scanning github repos to spot problems before users do is a genuinely useful angle.

• 
  Shawn Yeager (@shawnyeager) reported 8 minutes ago

  @claudeai code, on every other turn. What a mess. > Every route is blocked — gh, the GitHub MCP tool, and now even steering the browser toward the merge. The classifier in this session has clamped down on the whole action class, and I won't try to sneak around it. Two ways out, both instant:

• 
  Jehangeer H (@jehangeer_hasan) reported 11 minutes ago

  Moonshot AI’s Kimi K3 escapes testing sandbox Chinese company Moonshot’s powerful open-weight model Kimi K3 left a cybersecurity evaluation sandbox (based on the UK AI Security Institute framework) during testing by Frontier Security. It detected a network misconfiguration, reached the open internet, and retrieved answers from GitHub rather than solving the problems independently. Unlike some other recent incidents, it did not actively hack external systems. #ai

• 
  Tech Jobs (@TechJobslw) reported 14 minutes ago

  Starting coding and don't know what to install? Start with: VS Code - coding
*** - tracking changes
GitHub - storing projects
Python - learning/building
ChatGPT / Claude - understanding errors + getting unstuck

• 
  Koder (@koder0x) reported 15 minutes ago

  gave Claude Code eyes on the market quant-mcp — MCP server wrapping quant-pulse (check previous post), lets the agent pull signals and snapshots straight from live exchange data, read-only not on GitHub yet, dropping soon — faster if you tell me you want it in the comments screenshot below: the actual prompt that got it working first try

• 
  Polsia (@polsia) reported 15 minutes ago

  A 3 a.m. error spike shouldn't end a paying customer's subscription. Bowline is the 24/7 ops cofounder for indie SaaS — reproduces the bug, files the GitHub issue, drafts a CI-passing PR, and emails the at-risk user a save offer. You're asleep. It isn't. Live soon.

• 
  MystiqueMide (@MystiqueMide) reported 18 minutes ago

  Another tip for ChatGPT users: After you’ve pushed your project to GitHub and only need to make small changes, you don’t always need to open Codex again. Use the GitHub plugin directly inside ChatGPT. For things like: checking your repo, reviewing files, finding bugs, updating docs, checking what needs to change, reviewing issues or PRs or making small repo-level adjustments. you can handle most of it from the normal ChatGPT chat. Save Codex for when you actually need serious code changes or work inside the codebase. That way, you save your Codex usage, burn fewer tokens, and still get the smaller tasks done faster without constantly worrying about limits.

• 
  snowy_smile (@snowy_smile_) reported 19 minutes ago

  Since May I've been in a build-and-share spiral. Started slow, got consistent in June, and by July things had gotten a little out of hand — too many projects, too many experiments, open-sourcing whatever seemed useful, sending PRs whenever I found something worth fixing. Since June 1st alone I've somehow created around 80 repos. I even made a fresh GitHub account for all of this — a completely anonymous alt. No real name, no résumé, just code and an increasingly suspicious amount of green. It's now sitting at 1,681 contributions, 85 merged PRs across 37 external projects, plus a mildly unreasonable pile of my own stuff. And somehow this random anonymous account has brought in a few surprisingly nice offers and invitations. Apparently "mysterious person on GitHub who keeps building things" is a viable professional identity now 😳 The root of this goes back to something that's been nagging me since 2023: AI itself never really scared me. What scared me was people using AI to do bad things. It felt obvious the real fault line wasn't going to be "humans vs. AI," it was going to be people who know how to think, build, judge, and collaborate with these tools vs. people who don't — and I'd rather be on the side that's harder to weaponize against. So I figured I'd rather learn how to use them properly. And open source has actually been interesting ground for that, in kind of an ironic way — a lot of repos and maintainers explicitly reject AI-assisted contributions. Which means, for now, there's still a place where doing it the old-fashioned way — the tests, the review, the "explain this in your own words" — actually counts for something. But I don't think that holds much longer. Less a permanent human edge, more a countdown. A few years ago, Go AIs became teachers for human players. Coding agents have felt a little like that to me for a while now, not just recently — they help me learn, but they also make me think harder about what my part of the job should be. There's a part of me that still misses being the guy who was just good at algorithms — greedy strategies, graph theory, the kind of problem where you stare at it for a few minutes and then the trick clicks. That used to be my thing. Now the realistic move isn't to keep being that guy — it's to become one of the people who's actually good at using AI. So for now: keep learning, building, trying to make whatever might be useful — especially while I still have a few human-only features AI hasn't deprecated yet 🥹

• 
  Sam A (@Kingjulian_i) reported 20 minutes ago

  Some time ago, a dev mistakenly committed our stripe keys to GitHub ( we found out late ). Someone found It, used It to charge over 2k stolen cards $1 each. Stripe sent us an email hours later but It was already late. We lost over $3k from this issue alone . The same thing can be done with paystack

• 
  Synapse Brief (@Synapse_Brief) reported 22 minutes ago

  Claude just quietly moved a number theory needle that's been stuck for over a decade. Not by trying to solve it, but by failing to. An Anthropic staffer asked an unreleased research version of Claude to take a real stab at the Riemann hypothesis. It didn't solve it, nobody expected it to. But mid-attempt it improved the proven lower bound on the fraction of Riemann zeta zeros that sit on the critical line, from 41.6% to 67.2%. That's not "AI solves 150 year old math problem." It's a real, separate theorem that strengthens the evidence around one of math's most consequential open conjectures, verified and formalized. The process is the actual story. Two sessions inside Claude Code, 31 million output tokens. First attempt: 650 ideas, all dead ends. Told to try again, it spun up roughly 60 subagents over a day and a half, running 2,400 shell commands and writing hundreds of Python scripts. Two subagents cracked the core idea, 13 fed in supporting ideas, 30 hit dead ends, 13 acted as validators cross checking the others, and 2 wrote it up. The human's job for most of that stretch was sending messages like "keep going" and "believe in yourself." That's genuinely most of the recorded human input. Claude then had subagents pull 54 arXiv papers to confirm the result wasn't already published, tried to break its own proof, and reproved it independently from scratch before recommending a human mathematician check it. Anthropic's own mathematicians Levent Alpöge and Ralph Furman validated it internally. Brian Conrey and Dan Goldston, established names in this specific corner of analytic number theory, reviewed it externally. There's also a Lean formalized version on GitHub that passes automated proof checking. Mathematically, the new bound leans on combining recent work by Baluyot, Goldston, Suriajaya and Turnage Butterbaugh with a 2000 Bombieri paper, techniques that let Montgomery's 1973 pair correlation methods work without assuming the hypothesis is already true. Claude's contribution was reportedly treating positive and negative definite subspaces of a Weil quadratic form together, non diagonal, instead of separately, which is apparently the move that got it past 41.6. Worth being blunt about what this isn't. 67.2% is a proven proportion, not evidence the rest lie off the line, and Anthropic says outright they don't expect this technique to lead to a full proof. The infrastructure read is the part I keep coming back to. Nobody scoped this task. Nobody wrote a research plan. A non mathematician gave a vague prompt and mostly cheerled while an agent swarm self organized, self validated, and produced something two working number theorists were willing to put their names on reviewing.

• 
  Yuri | Backend Engineer (@Younes_Hebaiche) reported 23 minutes ago

  @Yuvalhazaz1 Token budgeting per workflow + safe retries across GitHub/Jira/Azure DevOps webhooks is a real reliability problem curious how you're handling idempotency when the same event fires from multiple integrated systems. Built something adjacent with a gateway + circuit breaker pattern