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:

• 
  kiryl.ziusko (@ziusko) reported 27 seconds ago

  @rizzrark Oh no, it should always give a correct result. Don't you mind opening an issue on GitHub? I would love to understand the issue better 👀

• 
  Peter Steinberger 🦞 (@steipete) reported 1 minute ago

  @EndGovTyranny Please file a github issue with more infos - with that alone we can't help. That's likely a weird model edge case. If you want a fast fix, use one of the top-gen models (OAI, Anthropic)

• 
  Sean Keenan (@sean9keenan) reported 3 minutes ago

  @brian_lovin Semi-relatedly: I’m back to VS Code from Cursor, autocomplete seems much better now! (Not that I’m crafting code by hand much) But importantly, the… basics seem much more stable (Cmd+f, and saving have been pretty broken in Cursor recently) Curious how GitHub Copilot feels!

• 
  AtomicNodes (@AtomicNodes) reported 5 minutes ago

  Hermes Agent vs OpenClaw on Qwen 3.6 Local Model We asked agents to scrape GitHub star history for both tools, find what caused the growth spikes, build a live dashboard in the browser. MacBook Pro M5 Max 64Gb. OpenClaw: 203k tokens, 12m 01s - wrote a bash script Hermes: 257k tokens, 33m 01s - wrote a SKILL.md OpenClaw: hit GitHub API, got truncated responses, paginated through contributors, pulled star-history JSON, found a security incident in OpenClaw's history, fetched SVGs, fixed broken HTML from trimming, rewrote it clean. Hermes: parallel tool calls across GitHub API, web search, and browser. Hit Google rate limit, auto-switched to DuckDuckGo. Fetched article contents, mapped viral moments, then built the dashboard. Both shipped a live dashboard with star growth charts and spike annotations

• 
  chandog (@thechandog) reported 11 minutes ago

  @kevinrose @digg how are you constructing novelty? stars are 40c on the dollar and a terrible way to measure anything on github.

• 
  Ryan Djurovich 😎 (@ryan0x44) reported 18 minutes ago

  @glcst By this logic Microsoft will fix the GitHub and NPM security issues soon

• 
  Lakshmi Tanmay (@lakshmitanmay) reported 18 minutes ago

  @ThePrimeagen Github is the only platform/service I believe that genuinely needs a proper rewrite… clearly something fundamental is broken.

• 
  Priyanshu (@iproductAI) reported 20 minutes ago

  For example: "Sign up for early GitHub Copilot Desktop access." Please fix this issue.

• 
  Neo (@neodevils_) reported 23 minutes ago

  @codeblue87 @diabrowser Hey, A few weeks ago, I tried to refresh to view all of my PRs in tabs with new GitHub Pull Request preview. But it was not working. I know this is a beta feature in GitHub and they might publish it sooner. Will Dia work on that before it is late?

• 
  Péter Szilágyi (@peter_szilagyi) reported 24 minutes ago

  @josefprusa @Mojee3d @Prusa3D I have a Prusa, across the parts and kits spent probably over 2K EUR on it. The multi-material printer fails incredibly often, software issues / hangs, random overvoltage errors, ignored github issues, etc. It’s not only about hw price, support is also very lacking, unfortunately.

• 
  Hackscorpio (@hackscorpio) reported 25 minutes ago

  @thsottiaux Codex review is not working right. When the model finishes, it doesn't render the response properly (not a model degradation). Seems like a regression in Codex application. I have no idea where to report that. I was reporting errors for cli and VSC extension on github.

• 
  Babyface4Lyfe | VTuber (@Babayface4Lyfe) reported 29 minutes ago

  I’m seeing a lot of people be concerned about custom chat elements if stream elements goes down. I may look into developing a local file system that doesn’t rely on stream elements, that I’ll then put on github or something for the community. Currently, the stream elements thing is still just rumour and speculation so I won’t make any promises until I hear otherwise.

• 
  Salt (@XMonetizationC_) reported 30 minutes ago

  🔥 Linus Torvalds has just made it clear that Linux will not become a dumping ground for AI-generated code. After months of internal debate, the Linux community has published its official rules on the use of tools like GitHub Copilot. The verdict: You can use AI to program, but the “slop”—that low-quality code spat out without thinking—does not pass the filter. The phrase that sums it all up: “Humans assume the errors.” You can rely on Copilot, Claude, or whatever you want. But if that code makes it into the Linux kernel, you are responsible. You verify it. You fix the bugs. You guarantee it meets the standards. This is the most mature stance I’ve seen in the open-source ecosystem regarding AI: neither hysteria nor blind adoption—just clear responsibility. The kernel has 30 years of history. They’re not going to ruin it to save 20 minutes with an autocomplete.

• 
  Louis Gleeson (@aigleeson) reported 30 minutes ago

  Grok runs the X algorithm. I just read the entire open-sourced codebase line by line. Here is exactly what makes a post go viral on X right now (save this): xAI quietly dropped the full For You algorithm on GitHub. 16,500 stars. Apache 2.0. Every Rust file, every Python script, every ranking signal. The first thing you need to understand is that there is no hand-engineered ranking anymore. None. xAI removed every single human-written rule from the system. The README states it directly. A Grok-based transformer does all the ranking now. That changes everything about how you should post. The transformer does not care about your follower count. It does not care about your blue check. It does not care about hashtags. It is looking at one thing. Your post's predicted engagement score across 15 specific actions. Here are the exact 15 actions the model is predicting for every post in your feed right now. Copied directly from the code: P(favorite). P(reply). P(repost). P(quote). P(click). P(profile_click). P(video_view). P(photo_expand). P(share). P(dwell). P(follow_author). P(not_interested). P(block_author). P(mute_author). P(report). The first eleven are positive. They push your post up. The last four are negative. They push it down. Your final score is the weighted sum of all fifteen. That is the formula. That is what every viral post is solving for whether the author knows it or not. Now look closer at the list. Eleven different ways to win. Most creators only optimize for likes and reposts. They are leaving nine signals on the table. The strongest signal in that list is dwell. Time spent on your post. The algorithm tracks how long someone stops scrolling to read what you wrote. A 400-word post that holds someone for 12 seconds beats a one-liner that gets 50 likes. The model has learned that dwell predicts every other engagement. This is why long posts are exploding right now. Not because X "promotes" them. Because they generate dwell, and dwell stacks on top of every other prediction the model is making. The second thing buried in the code that nobody is talking about is candidate sourcing. Your post enters the feed through two pipelines. Thunder serves your post to your followers. Phoenix serves your post to everyone else. Phoenix is the one that makes you go viral. Phoenix is a two-tower model. One tower encodes the user. The other tower encodes every post on the platform. It does similarity search using dot product matching against the global corpus. Then it pushes the top matches into feeds of people who have never followed you. This is exactly how a 12-follower account suddenly hits 800,000 views. Phoenix found a semantic match between the post and a user's engagement history, and the transformer scored it high on its 15 actions. Which means your post is not competing with your followers' posts. It is competing for embedding space. The way you win Phoenix is specificity. The two-tower model rewards posts that sit in a clear semantic neighborhood. Vague posts get vague embeddings and never get retrieved. Sharp posts about a specific topic with specific words get pulled into feeds of people obsessed with that topic. This is why "I built a SaaS" gets nothing and "I built a Postgres-to-Snowflake CDC pipeline in 4 hours using Estuary" goes viral. Same person. Same product. Completely different embedding. The third thing in the code is the Author Diversity Scorer. The model deliberately attenuates repeated author scores in the same feed. Translation: if your last three posts already got served to a user, the fourth post gets a penalty. This kills the "post 8 times a day for the algorithm" strategy. The algorithm is specifically engineered to dampen that. Better to post fewer times with stronger content than to flood and have your own posts compete with each other. The fourth thing is the filter list. Before any post gets scored, it has to pass through ten filters. The MutedKeywordFilter. The PreviouslySeenPostsFilter. The AuthorSocialgraphFilter. Plus a final VFFilter that removes anything classified as deleted, spam, violence, or gore. What kills your reach more than anything else is the PreviouslySeenPostsFilter. If a user has already seen your post once, you are filtered out completely from their feed. Forever. Which means every reply you make to a viral tweet that does not get visibility is permanently dead weight for that user. This is why the people who win at X reply only when their reply itself is good enough to be a standalone post. The last thing, and the one that should change how you write every single post: candidate isolation. During ranking, the transformer cannot let your post attend to other posts in the batch. It only attends to the user's engagement history. Your post is being scored alone. Against itself. Against what the user has previously engaged with. That is the entire game. Stop writing for the timeline. Write for the engagement history of the people you want to reach. Find the topics they already like, the accounts they already follow, the threads they already saved. Write into that semantic space. Phoenix will do the rest. The algorithm is no longer a mystery. It is sitting on GitHub at 16,500 stars. Apache 2.0. Anyone can read it. Almost nobody will. Link in comments.

• 
  Frooxius @ MFF - frooxius.bsky.social (@Frooxius) reported 31 minutes ago

  @MrRocketFX @ResoniteApp @unity They should be compressed on Resonite side? I'm not quite sure if I understand, it might be better to make GitHub issue for the request at the repo.