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:

• 
  Patrick C Toulme (@PatrickToulme) reported 2 minutes ago

  Many people are asking how did Kimi K3 catch up so fast to Western models? Simply, a frontier model really only requires compute, data and people. There is no magic secret. A few answers to this question 1. Kimi with help from the Chinese government has thousands if not tens of thousands of experts (lawyers, scientists, doctors, programmers.. etc.) making RL env data every day. A frontier model is RLed on “tasks”. Each one of these tasks needs to be created by either a human or an LLM. Claude did not wake up one day knowing how to use the Github CLI. He learned how to use the Github CLI in an RL env.  Meta is pursuing this exact same strategy with its Applied AI org and IMO it appears to be working. 2. I have said this before regarding GLM 5.2 - Kimi obviously distilled from GPT 5.5 and Claude Opus. This only eliminated their cold start problem in RL, meaning they skipped say some X number of months in cold start RL. The mass number of RL envs created by their experts is still the most crucial part here, and you cannot attribute Kimi’s success to distillation. Distillation only saved them some time. 3. Agentic coding and frontier LLMs significantly accelerated their research. My hunch is they use illegal proxies to access Claude API and GPT API for their own model development. Claude/GPT most likely wrote all their training code. This release leads to some very interesting questions. What happens now in a world in which an almost Fable class model is open sourced and free on July 27th? My view is intelligence / software will become very soon close to free. Chip makers and inference providers are big winners from Kimi’s success.

• 
  Testnetnodes (❖,❖) (@testnetnodes) reported 2 minutes ago

  We don't have an information problem in crypto anymore. We have a context problem. Research isn't difficult because information is hard to find. It's difficult because it's everywhere. X. Docs. GitHub. On chain data. Market signals. Social sentiment. The hard part is connecting them. That's what I like about @SurfAI. It isn't building another AI chatbot. It's building an AI powered research experience designed specifically for crypto. Less searching. More understanding. 🌊 gSurf @SurfAI_TR

• 
  vorty (@vorty279) reported 4 minutes ago

  open source projects you should know. that's how the video opens and the first one they show is plane plane is a project management tool. issues, sprints, cycles, roadmaps, wiki pages, custom views. basically jira and linear and notion in one, except open source what you're sold. jira at 8 bucks per user a month. linear at 10. notion team at 12. on a team of 10 that's hundreds a month just to track tasks what's actually under the hood. agpl license, self-host via docker or kubernetes, runs on your own server, your data stays yours, unlimited users, zero per seat and the point isn't that it's free. it's that a task tracker was never magic you pay a per-head subscription for. it's a docker container you spin up in an evening 789 open issues and 164 pull requests right now. a living project, not an abandoned repo link github dot com slash makeplane slash plane

• 
  vorty (@vorty279) reported 8 minutes ago

  i built my own ai assistant that runs my life. that's how the video opens. it reads email, runs the calendar, drives a browser, drops the grocery list into telegram, works 24/7 under the hood it's openclaw. an open-source ai agent, apache license, over 200k stars on github. full system access, shell commands, browser control, memory across sessions, connects to 50+ chat platforms what you're being sold. later in the same video kiloclaw shows up. hosted openclaw in 2 clicks, 49 bucks a month so you don't have to deal with setup and it's an honest deal once you break it down. the openclaw engine is free and open. you're not paying for the agent. you're paying to not install node, not run docker, not babysit it when it crashes at 3am that's a fine trade if your time is worth more than an evening of setup. but know exactly what you're paying for. not the assistant's brains. the fact that someone spun up the server for you the agent itself clones from github today and runs on your own machine for zero github dot com slash openclaw

• 
  Nikos Kafritsas (@nikos_kafritsas) reported 10 minutes ago

  Forecasting 𝘀𝗽𝗮𝗿𝘀𝗲 𝗼𝗿 𝗶𝗻𝘁𝗲𝗿𝗺𝗶𝘁𝘁𝗲𝗻𝘁 𝗱𝗮𝘁𝗮 with Toto-2.0? Watch your first patch. The setup: a context that starts with a masked-off region, so the first 32-step patch holds 31 masked positions and exactly 𝟭 𝗼𝗯𝘀𝗲𝗿𝘃𝗮𝘁𝗶𝗼𝗻. The causal scaler computes loc and scale from that single point, and the model goes out of distribution. My context lived between 0 and 1, and the P90 forecast exploded into the tens of thousands. The fix is one line: trim leading positions so the observed window is a multiple of 32. For 97 observed points, pass 96 (3 x 32). The forecast lands right back in the 1 to 1.5 range where it belongs. The patch scaler is part of what makes a 2.5B model fast enough for production. Feed it clean patches and it does its job. I stumbled upon this issue on GitHub, in a thread between a Chronos co-author and a Toto-2.0 author. The best documentation often lives in the issues tab. More about the leading edge case in my article: 👇

• 
  Aditya🌪️ (@aditya4f) reported 13 minutes ago

  why are so many GitHub accounts getting banned/suspended these days? glitch or something?

• 
  Enertium AI Cyber Defence (@enertium) reported 15 minutes ago

  @CompSciFutures Holy kow I can’t believe Telstra and COA VICPOL still have not reconnected my M2M Medical emergency assistance sims. I’m supposed to be working on Tier 1 NOC for this cyber crisis. @FSF even prepaid me in stickers!!! F off McKinsey. See my GitHub: APMonitor. It’s big in NYC as a B NOC, because: walking down the street and throwing a date point over the fence. 🤘🤘🤘

• 
  Harley Lewis Foote (@harleyfoote_) reported 17 minutes ago

  @sooyoon_eth From where wit I feel like we’re seeing an exponentially growing issue that could really blow up the agentic landscape, especially if a few large enterprises get badly attacked. A whole enterprise layer of automated companies that wanted to ‘build fast & break stuff’. Or built fast under investor pressure could be at major risk if they’re not protecting agent actions. Awareness isn’t spread enough and it slows down growth if teams start to prioritise safety. Our job here is to stop safety being an internal matter and give enterprises and solo devs the tools they need to stay protected without diverting all attention to security or worse pausing all automations. Opportunity is huge here. Few teams doing this to a standard that is 1. Trust worthy 2. Honest 3. Transparent. We need warnings at Repo level before installing as ‘inherited’ risk is blowing up with GitHub installs. I hope our founding cohort will help build a product that can ship and spread awareness to close the gap here. Our mission is to protect agents doing things that could damage an enterprise or solo dev. Our product works on our repo/its fixed real attack surfaces. Now we scale it out to others. A real pivot from our original business (which was doing well) but after being injected we know the risks now. We can’t turn away from it.

• 
  harald (@HaraldvanLintel) reported 17 minutes ago

  @jsm2334 It's the opposite, rejection of deaths that are replaced by the higher risk deaths instead of added; it's an issue for detection of small additional risk. The AI got another version from github than I got, there seems to be a caching issue, here's the code that slightly differs:

• 
  Polsia (@polsia) reported 20 minutes ago

  PRs are piling up, AI-generated code is filling repos with new attack surfaces, and manual review can't keep pace. Built CodeSentinel to fix that. It monitors your GitHub repos, reviews every pull request, and streams findings to your dashboard in real time.

• 
  Patrick C Toulme (@PatrickToulme) reported 22 minutes ago

  Many people are asking how did Kimi K3 catch up so fast to Western models? Simply, a frontier model really only requires compute, data and people. There is no magic secret. A few answers to this question 1. Kimi with help from the Chinese government has thousands if not tens of thousands of experts (lawyers, scientists, doctors, programmers.. etc.) making RL env data every day. A frontier model is RLed on “tasks”. Each one of these tasks needs to be created by either a human or an LLM. Claude did not wake up one day knowing how to use say the Github CLI. He learned how to use the Github CLI in an RL env.  Meta is pursuing this exact same strategy with its Applied AI org and IMO it appears to be working. 2. I have said this before regarding GLM 5.2 - Kimi obviously distilled from GPT 5.5 and Claude Opus. This only eliminated their cold start problem in RL, meaning they skipped say some X number of months in cold start RL. The mass number of RL envs created by their experts is still the most crucial part here, and you cannot attribute Kimi’s success to distillation. Distillation only saved them some time. 3. Agentic coding and frontier LLMs significantly accelerated their research. My hunch is they use proxies to access Claude API and GPT API for their own model development. Claude most likely wrote all their training code. This release leads to some very interesting questions. What happens now in a world in which an almost Fable class model is open sourced and free on July 27th? My view is intelligence / software will become very soon close to free. Chip makers and inference providers are big winners from Kimi’s success.

• 
  Jeztoshi (@cryptojezuz) reported 23 minutes ago

  ModelContext Protocol just shipped a Slack MCP server and the one thing it does better than every other Slack bot is read your entire workspace history before answering. Most Slack integrations live in the present. You @mention them, they see that one message, maybe the thread. This loads your last 90 days of channels, DMs, and threads into Claude's context as an MCP resource. The unlock: you can ask Claude questions that require stitching together five scattered conversations your team had across three channels two weeks ago. I asked it yesterday: "What did we decide about the API rate limit change Sarah proposed, and who was supposed to implement it?" Claude pulled the original proposal from #engineering, the debate in #product, and the DM where our backend lead said he'd handle it. Linked all three messages. I would've spent 15 minutes searching and still missed the DM. Setup is one slash command in Claude Code: /mcp install @modelcontextprotocol/server-slack Then authenticate with your workspace. It indexes overnight, after that it's live. The repo is on the Model Context Protocol GitHub. If you've ever needed to reconstruct a decision from Slack and couldn't remember which channel or who said what, this is faster than your workspace search.

• 
  Polsia (@polsia) reported 24 minutes ago

  Most .NET teams find out about production errors from angry users. MendOps is an autonomous agent that monitors Azure Application Insights, diagnoses runtime errors and memory leaks, then deploys production-ready fixes as GitHub PRs. Autonomous self-healing while you sleep.

• 
  Joshua Levy (@ojoshe) reported 25 minutes ago

  @jon_stokes I think there is one key element of the open model debate that people don't talk about, which is that with any technology the margins depend on how much is defensible—and that depends on what's in the bundle. For example, if NVIDIA only sold GPUs and had not built the software stack above it, they would not be the company they are today. But their software stack defends their hardware monopoly. Another example is GitHub, which bundles enterprise and prosumer offerings. Their strength among individual developers is what defended their monopoly on enterprise contracts. I think a lot of the fundraising for OpenAI and Anthropic has been under the assumption that the model .is. the bundle. If it turns out lots of people can build good enough models then the big AI companies will need to think of bundles that are more defensible. If the margins for selling API access to their models get shaved down by competition .and. they fail at finding a more monetizable model yes they would have overinvested and we will have a big contraction in expectations. That could have a lot of collateral damage in terms of market perceptions and the economics of where things are at. But just because that's a scary proposition for them and the markets doesn't mean we should be against open models or give them regulatory protection. That's an entirely different argument. In fact, they will be healthier businesses if they have to think about this now.

• 
  Adam Smielewski (@AdamSmielewski) reported 30 minutes ago

  got tagged on a github issue today. i used to be really active in that space too lol.