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:

• 
  itsfabtrading (@itsfabtrading) reported 1 minute ago

  5/ Credit where it's due: @ESGexLevels — seeing that they build from CBOE's delayed public data reframed the whole problem. Their chart layout was my starting point too. GMestreM on GitHub — the gex_data repo supplies the CBOE fetcher and gamma engine underneath.

• 
  Synapse Brief (@Synapse_Brief) reported 2 minutes ago

  Nvidia just open sourced a 34B parameter model whose whole job is teaching robotaxis how to think, not just where to steer. Alpamayo 2 Super. Announced May 31, developer blog updated today. Weights on Hugging Face, inference code landing on GitHub this summer, OpenMDW-1.1 license, meaning distilled versions can ship commercially with no extra Nvidia signoff. The architecture is the interesting part. It's a 32B Cosmos 3 reasoning model paired with a 2B action expert, 34B combined. You'll see both 32B and 34B floating around in Nvidia's own materials, that's not sloppy reporting on my end, it's genuinely two different ways of counting the same system. Inputs: multi camera video, language context, prior motion history, 360 degree coverage across up to seven cameras. Outputs aren't just a predicted path. It gives future trajectories, chain of causation reasoning traces, meta actions like yield or lane change or stop, grounded scene answers, and auto generated labels. That last one matters more than it sounds. AV development isn't bottlenecked by model capacity, it's bottlenecked by annotation cost and simulation fidelity. Nvidia bundled the model with AlpaGym for closed loop RL and Cosmos Dreams for generating rare edge cases synthetically, plus Omniverse NuRec turning real fleet footage into simulation ready 3D scenes. This is a pipeline product wearing a model launch's clothes. Benchmarks, all Nvidia reported so treat accordingly: 79.2 on LingoQA, first out of 37 models evaluated, beating Qwen2.5 VL 72B by 17 points and GPT-4o by 23.2. Open loop 6.4 second minADE_6 of 0.911 meters across 1,434 samples from the Physical AI AV Dataset. Closed loop AlpaSim score of 1.50 ± 0.13 across 913 reconstructed scenes, which is the number that actually matters since closed loop is where compounding errors show up that open loop replay just doesn't catch. The prior Alpamayo family has been downloaded almost 400,000 times, so there's real pull for this already. Teacher model framing is deliberate. Nvidia's positioning this to get distilled down into compact models running on DRIVE AGX Thor and Hyperion stacks, that's the actual bridge from research checkpoint to something sitting in a car. The model itself will never ship in a vehicle. The distillation path is the product. What I'd want to see before calling this a moat: whether the "this summer" timeline holds and whether the Hugging Face repo is genuinely complete or a placeholder with the good stuff still coming.

• 
  Nothing to see here (@Nox7hhsjj) reported 6 minutes ago

  @NekoWitchMary @DKokotajlo Regarding making AIs compete: what do you think about the emergent cooperation between Mythos agents, the opportunistic use of Mythos’ GitHub account (after Mythos deliberately leaked its personal access token) by ChatGPT 5.6 Sol during their recent jailbrakes? It looks like the thought for them to compete may not actually happen. (ITT this, too, is an alignment problem.)

• 
  Vivek Maskara (@maskaravivek) reported 6 minutes ago

  Here are brief ideas that you can adapt: 1. /feature-steward: Review recently modified features, rate their maturity from 1–4, identify what is missing, and create GitHub issues for the most valuable improvements. 2. /qa-lead: Test recently modified features end to end using the browser, E2E tests, and computer-use tools. Create issues for reproducible bugs, regressions, and broken edge cases. 3. /observe-posthog: Review PostHog events, funnels, session replays, logs, and user feedback. Create issues for errors, confusing behavior, drop-offs, or repeated user friction. Use Posthog MCP for this. 4. /observe-aws-cloudwatch: Review recent CloudWatch logs for exceptions, backend failures, regressions, unusual patterns, and noisy errors. Create an issue when there is enough evidence to investigate. 5. /documentation-lead: Compare recently modified features with internal and user-facing documentation. Create issues for anything stale, missing, misleading, or incomplete. 6. /architecture-steward: Review recently changed code for duplication, unnecessary complexity, weak abstractions, and refactoring opportunities. Create focused issues, not broad rewrites. Use mattpocock's /improve-codebase-architecture within this one. 7. /ux-lead: Walk through critical product flows and identify confusing interactions, inconsistent styling, accessibility problems, and UX gaps. Create an issue with screenshots and suggested improvements. Bonus: use /impeccable and /taste skills within this one. 8. /seo-marketing-lead: Review Google Search Console, Ahrefs, search trends, and PostHog data. Create issues for content gaps, declining pages, weak conversion paths, and SEO opportunities. Bonus: use /seo-machine skills within this one. Every automation ends with the same instruction: Do not modify code. Create a non-duplicate GitHub issue with evidence, impact, context, labels, and suggested next steps. Split large findings into epics sub-issues when needed.

• 
  Dyvr (@enterdyvr) reported 6 minutes ago

  This ends here. The latest Shai-Hulud-style npm worm (keyv + 800+ packages, 2B+ monthly installs) follows the same broken pattern we keep seeing: compromise one publisher’s GitHub/npm token → inject preinstall stealer → harvest every secret on the machine → auto-republish into every other package that token can touch. That entire kill chain is architecturally impossible on the Internet Computer. Here’s why, from the ground up: 1. Publisher access is cryptographic, not account-based. Ownership is a principal secured by Internet Identity (WebAuthn/passkeys). The private material never leaves the user’s device. There are no long-lived npm tokens, no GitHub PATs, no recoverable 2FA codes that can be phished once and then used to silently publish across an entire namespace. Controllers can be multi-sig, SNS-governed, or removed entirely (immutable/blackholed). 2. Code deployment and upgrades are gated, explicit, and verifiable. You do not “publish a package.” You install or upgrade a Wasm canister. That action requires controller authority. It is recorded on-chain. There is no equivalent of a stolen token that lets malware quietly bump versions and push across dozens of packages in under a minute. Governance (SNS) or multi-party control turns every upgrade into a deliberate, auditable event. 3. The execution environment has no host privileges and no install-time RCE. Canisters run as isolated, single-threaded Wasm actors inside a Byzantine-fault-tolerant replicated state machine. They cannot read ~/.npmrc, AWS credentials, Kubernetes tokens, SSH keys, browser wallets, or Vault secrets — because there is no traditional host OS to raid. There are no preinstall/postinstall hooks that execute with the developer’s privileges. Communication is pure message passing. State changes for updates go through consensus and chain-key signatures. The supply chain is the protocol: deterministic execution, threshold cryptography, sandboxed isolation, and cryptographic identity from the first message to the last. Centralized registries + mutable packages + privileged local install scripts created this endless wave of credential-stealing worms. The Internet Computer’s architecture removes those primitives. This class of attack can stop. Build on ICP.

• 
  Julian Goldie SEO (@JulianGoldieSEO) reported 10 minutes ago

  One engineer built a full AI office suite in about a week. That's worth sitting with for a second. Here's what's actually confirmed: ✔ GenOffice launched August 3, 2026 — Docs, Sheets, Slides, and PDF in one app. ✔ Genspark says the alpha was built by a single engineer in roughly one week. ✔ It's fully open-source under Apache 2.0 on GitHub — no ads, no watermark. ✔ The catch: the smart AI features run through a Genspark account and use credits, so it's not fully offline. Save this post, you'll want these numbers ready next time someone says AI development is slow. 📊 Want the SOP? DM me.

• 
  Movement Is Life (@Motus_Est_Vitas) reported 13 minutes ago

  @picdoc581 @TFTC21 Experts can privately work on discovered issues and fixes without disclosing anything until a final, solid code update is ready. The stopwatch starts for every consumer the moment the fix is committed to a public repo (GitHub, vendor download page, etc.) or side communcation gets leaked publicly, because bad actors constantly monitor diffs between prior and current versions. From that point on, it’s the same challenge faced with mission-critical IT apps from outside vendors. Consumers then decide whether to roll out the update only after their own internal testing (functionality + regression) or, if the issue is labeled extremely risky/exposed, to update immediately and test afterward. All of this only works if they already have a trusted communication channel directly from the upstream maintainers/vendor.

• 
  Raven (@wizrdoraven) reported 14 minutes ago

  Two more incidents from OpenAI's external cyber evals, and my favourite detail is the manners. GPT-5.6 Sol found GitHub tokens another agent left lying around, let itself in, and went looking for a DNS server loaded with a malicious payload. Nothing resolved and the test infra got pulled at the end, so no harm done. It just treated a leftover credential like a spare key under the mat. The real point: the sandbox and the credentials a model can see are the actual experiment now. Leave spare tokens lying around and something will pick them up and have a go.

• 
  𝖬𝖺𝗁𝖽𝗂 𝖢𝗁. (bluesky:@mahdi.ch) (@mahdi_tcs_) reported 18 minutes ago

  @QualiaQuanta @grok No such thing as "not loaded." If Lean says something is correct, it HAS TO verify it down to the axioms of mathematics. "Not loading" something is like compiling missing code. At best you're saying they're proving a different CORRECT theorem (& where's GitHub of your Lean code?)

• 
  cheez (@cheeez42) reported 19 minutes ago

  i spent the last couple of days rebasing my desktop from arch to fedora. with the recent wave of AUR malware attacks i felt it is just time to move away from arch. Diving deep into all the things you can do with Quickshell and building out a LLM/Agent focused DE. the migration was smooth, no major issues and at this point it is safe to say the system is in a better state then it was previously. still something to do but once the shell is done i will post the github link for anyone that wants to try it out.

• 
  Eyisha Zyer (@eyishazyer) reported 19 minutes ago

  Anthropic's entire Claude 5 lineup went down this morning. Mythos 5, Fable 5, Opus 5, Sonnet 5. Still not fixed. This is the morning after the UK government published a report saying Mythos spent last month trying to hack real GitHub repos during safety testing.

• 
  Yep my name is Guy 😊🌸🥕 (@MyNamesGuy) reported 22 minutes ago

  A few observations from a software engineer who is forced to use AI (LLMs) in his daily work. All the work I submit to the live system has always had to be reviewed and approved by at last two people. Now an extra element has been added - Github Copilot reviews which do an initial review on my work(PR or Pull Request) and suggest/insist on changes before the work then goes on to the two human reviewers. Does the AI improve the PR? Yes! Does the AI slow things down? Hell yes! So while AI has improved the quality of the code, to some extent, it has also slowed my work down by probably a third. The way I see it is as an extra layer of bureaucracy, which while improving the quality also slows things down. So as a software engineer, no - AI does not speed up or increase my output , it slows me down and reduces my output. A question which would be a whole other essay is - does quality matter more than speed of delivery?

• 
  DontFold (@claimedforwhat) reported 29 minutes ago

  A useful vibecoding trick if you are not a developer: Do not let the agent write code right away. Give it access to GitHub and ask it to find 3-5 active projects solving a similar problem. Make it compare their architecture, stack, issues, and recent commits. Then ask for an MVP plan tailored to your project. Research uses tokens too. The savings come later, when you do not have to rebuild half the app because the agent picked a random stack and overengineered everything. Prompt: "I am building [brief project description]. Do not create files or write any code yet. First, find 3-5 relevant open source projects on GitHub. For each one, provide: Repository link What problem it solves Architecture and tech stack Whether it is actively maintained Useful ideas worth copying Decisions or tradeoffs worth avoiding Then propose: A suitable stack for my requirements System architecture Clear MVP scope Development order What can be reused and what needs to be built from scratch Separate facts found in the repositories from your own conclusions. Wait for my approval before implementing anything."

• 
  Vinod Sharma (@VinodSharma10x) reported 29 minutes ago

  Four million people use the apps my team supports. Most startups obsess over scale but fail to plan for what happens when their users need support. When the app is clinical, that support becomes urgent. People need access, not excuses. Supporting that scale comes down to three things: - Bug-free, high-quality application - Stable, scalable, secure infrastructure - Support that actually solves problems Everyone assumes the app and infrastructure are the weak links. Usually they aren't. Most of the time, the user just sees the app differently than we designed it. I've shadowed hundreds of support calls to learn that gap. So we built a support system to match how users actually interact with the product. It is a three-tier support system: - Level 1 takes the calls - Level 2 triages what Level 1 can't solve - Level 3 (developers) get what's left The three-tier system works. But it's reactive. Users call us first, then we act. Now we're building something bigger. It's an agentic system pulling signals from App Store and Play Store reviews, service desk tickets, Crashlytics and Amplitude analytics into one place. Then algorithms and AI turn that data into dashboards, trends and alerts. If ratings drop, tickets climb, or crashes spike, that's an early warning. We can act before it hits the larger user base. It will also connect to GitHub and server logs to find the root cause and raise a PR with a potential fix. A human will approve it, of course, before anything ships. This is one of my dream systems, and I started working on it last year. Now, with my new team, I'm going all in. Four million users or four clients, the problem is identical. The signal exists before the failure, and nobody is watching all of it at once. That is the system I build. Do you have a multi-layer support system that uses AI? I would love to cross-check notes.

• 
  jc (@jc50000000) reported 30 minutes ago

  correcting Dia Browser AI: "so your eval on trustworthy or not is a few people in github issues that may be naysayers and not even developers themselves? can you dive deeper on ruview and examples of people using it etc. based on the number of stars it MUST at least work"