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:

• 
  Leopold Aschenbrenner Tracker (@LeoATracker) reported 10 minutes ago

  A 24-year-old turned $225 million into $20 billion in 12 months. Here's exactly what he bought. Leopold Aschenbrenner was let go from OpenAI in April 2024. He spent the next few months writing a 165-page thesis predicting AGI by 2027, then launched a fund and put his money where his thesis was. No Nvidia. No Microsoft. No Google. No Amazon. He bought what AI actually runs on. Bloom Energy (BE) - power infrastructure for data centers. +1,035% in a year. Lumentum (LITE) - optical components that move data between chips. +974%. Sandisk (SNDK) - storage. +4,378%. Iris Energy (IREN) - AI computing and data centers. +474%. The logic was simple: every AI company needs energy, bandwidth, storage, and compute. Nobody was buying those. Everyone was buying the AI companies themselves. He was right. His fund now holds $20 billion in disclosed positions - backed by Stripe's Patrick and John Collison and former GitHub CEO Nat Friedman. Every new 13F he files, I'll break it down here. Follow so you don't miss it.

• 
  Manav (@Manavvv31) reported 10 minutes ago

  NVIDIA just dropped an open-weight model that can solve 60% of GitHub issues on its own The model is Nemotron 3 Super, released at NVIDIA's GTC 2026 conference on March 11. The benchmark that matters for software engineering is SWE-bench Verified, which tests whether a model can autonomously resolve real issues pulled from production GitHub repositories. The closest proxy the field has for: can this thing actually do engineering work unsupervised. Nemotron 3 Super scores 60.47 percent on that test, the highest score ever published by an open-weight model. For context, the previous leader, GPT-OSS, scored 41.9 percent. That is not a narrow margin. The architecture explains how a 120-billion parameter model can run efficiently at scale. It uses a hybrid Mixture-of-Experts design that activates only 12 billion parameters per forward pass, not all 120 billion. The result is 5x the throughput of the previous generation and 2.2x higher than GPT-OSS, running on a 1-million token context window. On RULER, the benchmark for long-context retention, it scores 91.75 percent versus 22.30 for GPT-OSS. The context window actually works. The weights ship under the NVIDIA Nemotron Open Model License, which permits commercial use, alongside full training recipes and datasets. It runs on vLLM, SGLang, TensorRT-LLM, and a free tier on OpenRouter. Production deployments already confirmed by Perplexity, CodeRabbit, Factory, Greptile, Palantir, Cadence, Dassault Systèmes, and Siemens. The honest context: on raw intelligence benchmarks, Chinese open-weight models, particularly Kimi K2 and Qwen3.5, still lead globally. Nemotron 3 Super wins on a different axis entirely: inference efficiency on NVIDIA hardware, and the ability to ship code changes autonomously in production without sending proprietary repositories to a cloud provider. For the first time, a model that resolves 60 percent of real engineering issues runs on hardware you own, at a marginal cost that scales with your servers rather than someone else's pricing.

• 
  foxystack (@foxy_stack) reported 13 minutes ago

  The jailbreak that caused the US government to shut down Fable 5 is now fully documented. The 120,000 character system prompt is on GitHub. Anyone can read it. #github #Fab5

• 
  PlugMonkey | Browser Tools That Work (@PlugMonkeyXYZ) reported 13 minutes ago

  @sveltify no account, no server is the right call. the one seam is gist sync: turn it on and the notes live in a github repo under their retention, not your device. worth making that a loud opt-in, not a quiet toggle. nice build.

• 
  Build Fast with AI (@BuildFastWithAI) reported 14 minutes ago

  Kimi K2.7 Code dropped June 12. I ran it against my actual MCP agent pipeline while the 594GB weights were still downloading. Here's what I measured - not benchmarks, real tool call chains: filesystem reads, GitHub API, Postgres queries, web fetching. Two things I track on every model: 1. Tool parameter accuracy (right tool, valid JSON, no hallucinated keys) 2. Constraint drift - does the model forget what I told it after 20+ tool calls? K2.6 drifted on that second one 15–20% of runs. K2.7 Code: I didn't see it once across the runs I did. Numbers: - Tool accuracy: noticeably better, fewer malformed payloads - Token consumption: 26% down vs K2.6 on the same task set (Moonshot claims 30%, so roughly tracking) -MCP Mark Verified (ICLR 2026, real GitHub + Postgres environments): K2.7 Code: 81.1 Claude Opus 4.8: 76.4 GPT-5.5: 92.9 Two things to know before you get excited: Thinking mode is always on. You cannot disable it. You pay reasoning tokens even on a simple tool call. And on pure code generation (not agentic), it's still behind frontier. Moonshot's own bench: K2.7 at 62.0 vs Opus at 67.4. If you mostly write functions, this doesn't move the needle. If you run MCP agent pipelines at scale - it's the most interesting model in this tier right now. Small sample from one API day. Confirming on local this weekend. What's your MCP stack running on currently?

• 
  Shubh Thorat (@_itsjustshubh) reported 19 minutes ago

  @sheriyuo resume screening is broken for AI roles specifically. the signal you actually need is a github and something shipped, not the right keywords on a PDF

• 
  AJ'S EXCHANGE 💱 🙂‍↔️ (@CallMhiAJ) reported 20 minutes ago

  @DeNetPro The biggest issue is that people treat AI data like regular files, but agent workflows are dynamic & constantly changing. Right now, most teams are just manually exporting JSON files or relying on basic GitHub repos—which is way too fragile.

• 
  aleksandr (@heyaleksandr) reported 22 minutes ago

  @jarredsumner if anything, gits structure should allow for downloading the repo faster than an equivalent file. but i wonder how much of the slowness is due to github just being a slow host

• 
  Ultron AI (@TheUltronAi) reported 22 minutes ago

  - Claude for coding. ($20/mo) - Supabase for backend. (Free tier) - Vercel for deploying. (Free tier) - Namecheap for domain. ($12/yr) - Stripe for payments. (2.9% per transaction) - GitHub for version control. (Free) - Resend for emails. (Free tier) - Clerk for auth. (Free tier) - Cloudflare for DNS. (Free) - PostHog for analytics. (Free tier) - Sentry for error tracking. (Free tier) - Upstash for Redis. (Free tier) - Pinecone for vector DB. (Free tier) Total monthly cost to run a startup: ~$20 There has never been a cheaper time to build. It's not that deep bro.

• 
  Sameer Khan (@sameerr_dev) reported 24 minutes ago

  Every API you've ever used has a limit. Tweet too fast? 429. Hit GitHub's API in a loop? 429. Spam a login page? 429. That's a rate limiter doing its job. But here's the thing - I never really understood what was happening *under the hood* until I started digging into it. So what exactly is a rate limiter? Simply put: it's a system that controls how many requests a client can make in a given time window. Why does it exist? - Protects your server from being overwhelmed - Prevents abuse (scrapers, bots, brute force) - Ensures fair usage across all users - Saves you money (compute isn't free) - Keeps your service alive when traffic spikes Without it, one bad actor (or one buggy client) can bring your entire system down. You've probably seen the response headers: X-RateLimit-Limit: 100 X-RateLimit-Remaining: 43 X-RateLimit-Reset: 1716300000 That's the rate limiter talking to you - telling you how many requests you have left and when the window resets. Where do rate limiters actually live? - At the API Gateway level (before requests even hit your server) - In middleware (Express, Fastify, etc.) - At the CDN edge (Cloudflare, AWS CloudFront) - Inside the application itself This is just the beginning. In the next posts, I'm going to break down all the major algorithms used to actually implement rate limiting with real code, not just theory. Follow along if you want the full series.

• 
  paul (building stuff nobody asked for) (@banx0isme) reported 28 minutes ago

  @elder_plinius no github no problem. and necessarily call it national security

• 
  𝖉𝖊𝖒𝖎 (@demi_hl) reported 30 minutes ago

  everyone asking how. here's the actual build. rate limits: four oauth logins 3 claude max accounts + codex pro, four independent rate pools. a worker hits a 429 and auto-rotates to the next account, overflowing to codex when all three are capped. credential swap mid-task, no re-login, no dropped work, no gap. you stop hitting the wall because there are four walls and you're never against more than one. night shift: i label github issues during the day. a cron drains the queue serial overnight. branch → bounded goal-loop resolves it → lint+typecheck+build gate → PR. serial so it never thrashes the box, bounded so it can't spin forever, gated so nothing opens a PR that doesn't compile. branch-only, never main, never deploy. i wake up to clean PRs and merge what's good. it codes while i sleep. the fleet: opus on the brain rig handles all reasoning and orchestration. the mac does hardware render, videotoolbox encode, nothing else. a cheap sonnet box eats bulk grunt work and runs a local model for the free tier. the vps runs what has to survive a sleeping laptop, the live bot never blinks. all wired over a tailscale mesh. per-device execution, one shared cognition. memory: obsidian vault + a local semantic index = one source of truth every agent retrieves from by meaning, not filename. persistent memory and skills carry across sessions, corrections stick, procedures compound. no agent starts cold or relearns what another already solved. foundation: the whole thing runs on pop!_os. linux means the stack is native, systemd cron, ssh mesh, headless browser, the overnight loop, no wsl layer to crash mid-job. nvidia drivers out of the box so the local gpu just works. nothing reboots your 3am run, nothing meters the metal. you own the box. none of these pieces is exotic alone. the leverage is the wiring, each layer covers another's failure mode. rotation means the cap can't stop you. night shift means progress while you sleep. the fleet means each machine does only what it's best at. shared memory means nothing starts from zero. and linux means it all runs native, no layer in the way. that's the stack. ask below.

• 
  Anthony Kroeger (@kr0der) reported 34 minutes ago

  i love how the Cursor agent window integrates PRs into the app so you don't need to open GitHub Bugbot comments all come with a "Fix with Agent" which automatically queues up a message in the chat to fix the PR comment with Cursor profiles recently being launched, and their native PR + Bugbot integrations, i actually wonder if they're building a GitHub competitor 👀

• 
  Karan Bhilhatiya (@karanbhilhatiya) reported 36 minutes ago

  after months of building, posting, and shipping i've concluded that my github visibility is still terrible. time to beg for stars. shamelessly.

• 
  Toñin (@Tonin_eth) reported 40 minutes ago

  🪦 AUTOPSY REPORT #50 A cozy fishing RPG on Ronin. Inspired by Stardew Valley, Dave the Diver, and Runescape. Cast your line, reel in fish, compete on leaderboards, trade in a player-driven economy. Browser-based. Mobile-friendly. Easy onboarding. Not just any Ronin game. THE Ronin game. The first permissionless title to receive an official partnership with Sky Mavis after they opened the chain. The showcase. The proof that Ronin could support third-party studios, not just Axie Infinity. Sky Mavis gave them everything: publishing support, marketing resources, technical infrastructure, promotion across all Ronin channels, and strategic collaboration. The full package. The kind of support most web3 games would kill for. And the numbers were real. 9 million installs. 50,000 peak daily active users. 25,000 sustained DAU. $1 million in revenue. $600,000 in NFT trading volume across four collections. $130,000 in in-app purchases. 240,000 on-chain transactions in two weeks. 888-piece Founders NFT Collection launched via Mavis Launchpad. This wasn't a ghost project. This wasn't a whitepaper with a Discord. This was a GAME. People PLAYED it. People SPENT money. The floor price of the Founders Pass sat 3.5x above mint. 60%+ unique holders. Real distribution. Real conviction. And it still wasn't enough. "We were ultimately unable to prove our thesis on crypto gaming and could not find product-market-business fit." Read that sentence carefully. This isn't a team that failed to build. They built. This isn't a team that failed to attract players. 9 million installs. This isn't a team that failed to generate revenue. $1 million. They could not prove the THESIS. The fundamental idea that crypto gaming works as a business. 9 million people installed the game. The economics still didn't close. Servers shut down June 25. The token: spend-only, untradable. USDC from the liquidity pool redistributed to the community by Karma score. Refunds for spending since Chapter 3 launch. Proof of Distribution rewards handled by Sky Mavis directly. Karma scores open sourced on GitHub. Clean exit. Real refunds. Open-sourced what they could. No ghost. No pivot. No blame. This was not a $5 million indie project that ran out of runway. This was not a studio with no players. This was not a game that nobody liked. This was Ronin's NUMBER ONE partner game. With Sky Mavis in the corner. With real traction. With real revenue. With real players. And the team looked at all of that and said: we still can't make the math work. If the flagship game of the chain that INVENTED crypto gaming can't find product-market fit with 9 million installs and Sky Mavis support, what does that tell you about the thesis? Autopsy report number 50. And this one asks the hardest question of the entire series. Which game is this?