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:

• 
  Noonien Soong (@mlcarldev) reported 1 minute ago

  Team @droid It's a bit unfortunate that something, likely in my local Droid installation, has stalled progress. This comes after 20 hours of brilliant, excellent planning and execution on the first 30% of this platform, where a stellar handoff procedure was created so I could start a new mission... which was the recommendation of the orchestrating agent in that first mission. Starting this second mission with a fresh context window, the agent again did a brilliant job planning the next milestones. It was extraordinary, detailed planning... but then it could not execute. After the planning and after me accepting the proposal, it refused to execute, throwing an error every time. The agent tried everything: 1. He decreased the size of the plan down to one line, so it is definitely not the content of the plan causing the issue. 2. He even deleted some mission and plan related json and other files to reset it while preserving all the information. I have restarted Droid and resumed the session, but it just doesn't work. I wrote a detailed, comprehensive bug report and filed it under issues in your GitHub repo, as this seems to be a real problem now. Issues #98 and #99 I hope that a next update will somehow reset my configuration. I didn't see a new version being installed that could have introduced a bug, so this must be something Droid does on such an extensive mission... perhaps when trying to start a new mission in the same repository, which is normal procedure according to the documentation. Something is off, and essentially I have been unable to continue the test since yesterday. I cannot continue having this platform coded here, while Opus Ultracode, on the other hand, has been delivering pretty functional stuff so far. It is a bit chaotic the way it works... it doesn't really stick to the plan... but it always comes back when reminded. I am pretty sure that today I will have a functioning platform delivered by Opus, though it will probably need some debugging and fine-tuning. It is unfortunate because I am confident GLM 5.2 could compete with Opus 4.8. The first stint showed this clearly; that first flawless 98% of the context window in the first mission was absolutely stellar. If I were to reinstall Droid from scratch, I assume I would lose all the artifacts that I have. The orchestrator: Key points to highlight when you pass it to Factory AI: 1. Root cause (smoking gun in the logs): the orchestrator session is bound to missionId 7ba4d425 via session tags, and this binding persists across CLI restarts. ProposeMission looks up that mission directory, finds nothing (because I deleted it trying to fix the issue), and crashes on H.length where H is the undefined result. 2. The bug is likely in session-tag lifecycle: the missionId tag is set at session creation time (before any ProposeMission call), so a failed proposal poisons the session permanently. The tag should be set AFTER a successful proposal, or cleared on restart if the referenced mission no longer exists. 3. The fix is almost certainly to start a completely fresh session (not --resume, and possibly in a new terminal window / after clearing ~/.factory/sessions/). I did not try this because you asked for the bug report first, but it is the most likely workaround on your side. 4. The AskUser tool is also broken in this session with a similar parse error, reinforcing that this is a session-state corruption issue, not a ProposeMission-specific bug. My comment: I meanwhiile tested. All the recommendations and the Ask User tool are now broken, even in completely unrelated new missions and new repositories. Planning also can't go to execution; it's always the same error. Droid seems to be broken for good now, at least on my computer.

• 
  Bradley Taylor (@bradtaylorsf) reported 1 minute ago

  It works with the tools teams already use. GitHub Issues become the queue. Each issue gets picked up by an agent. The agent works in a branch/worktree. Tests run. Failures feed back into the loop. Successful work becomes a PR. No new project management database required.

• 
  蜃気楼 (@fuxps32) reported 3 minutes ago

  One engineer at Anthropic stopped working his own bug queue. It clears itself now. He launched voice mode across the company's products, set up a routine, and walked away. It listens for every ticket, every GitHub issue, every bug report that mentions voice mode. When one lands, it writes the fix and opens the pull request on its own. Boris Cherny, the man who built Claude Code, says he has never once talked to that engineer about how it works. It just runs. The trick that lets a loop run that long is one rule. When Claude makes a mistake, the engineer does not correct it in the chat. He writes the correction into a CLAUDE.md or turns it into a skill. Patch it in the chat and it breaks again tomorrow. Write it into a file and it never repeats. Do that enough and the loop runs forever. Cherny lives the same way. Whenever he needs code, Claude writes it. Whenever he needs a review, Claude runs it. Whenever he needs a security check, Claude does it. He talks to a loop, and the loop prompts Claude for him. The engineer is still on the team. His feature has not needed him in months.

• 
  Conglomerate (@0xconglomerate) reported 4 minutes ago

  Why exactly do VLAs fail? VLAs start w/ LLMs as their brain. Early roboticists (2021-2022) noticed that LLMs trained on internet text had absorbed a large amount of implicit knowledge about the physical world. So they took that best available pretrained brain, observed that actions could be formatted like language tokens, and assumed the transfer would work. But world knowledge encoded in language ≠ physics simulation. There's essentially a data structure mismatch: ▸ LLM pretraining data is discrete, symbolic, and sequential (text). ▸ Physical control is continuous, high-dimensional, and requires split-second feedback. --- ➦ VLAs in the real world, by the numbers: ① They barely work ▸ VLAs start at ~30% success on real robot tasks, it need hundreds of human interventions just to reach ~90% ▸ Best pretrained VLA hit 27.4% task progress on real robots ② VLAs can't generalize outside training ▸ On actions it's never seen, best VLAs score 25-32% task progress (fails when you change the environment) ③ Fine-tuning doesn't help ▸ The more robot-specific, the dumber it gets at everything else (only works on clean, controlled, success-only demos) ④ Too slow for a real robot ▸ OpenVLA runs at 3-5 Hz (physical control needs orders of magnitude faster than that) --- The easiest way to understand how VLAs are actually wrong is thru a real life example. ➦ Let's say you hired a chef who learned everything about cooking by reading, but has never stepped in a kitchen. If you ask them how to cook a steak, they'll tell you the best answer. But if you actually ask them to cook, they'll struggle when you hand them the pan. They'll have a hard time picking up the ingredients. They'll burn the steak. They know everything about cooking, but can't actually cook. --- ➦ Thoughts I want to take back a line I've said before: "Robots can see, but they still can't listen." (referencing to my Silencio piece before) I take it back. Robots can see, listen, even reason now. What they can't do is act in the real world. It's basically an AI chatbot wrapped in a robot body, not a robot that can actually do tasks. No wonder most demos online are scripted. There's a real problem with the brain, and roboticists have been building on the wrong foundation. VLAs are like a trojan horse, they look like the answer but bring a bunch of problems in with them. VLAs only learn through imitation which brings up the data problem. "Enough data" at scale doesn't mean hundreds of demos total. It means hundreds per task, per robot body, per environment. Hundreds again every time any one of those changes. So you've basically got a human-labor bottleneck. To get that data, someone has to physically collect it, either through: ▸ Teleoperation (slow, expensive, needs trained operators) ▸ Kinesthetic teaching (tedious, doesn't scale to complex tasks) ▸ Motion capture (high precision but high setup cost) ▸ Simulation (robots trained in sim often fail in the real world because physics engines aren't accurate enough) And you'd think, okay, maybe someday a company figures out a better way to collect all this. But the problem doesn't stop once you already have the data... Switch to a new robot body and you're collecting data from scratch, because VLAs don't transfer well across embodiments. Move it to a new environment and you're collecting again, since it just overfits to whatever setup it trained on. Give it a new task and yep, collect again, because it can't generalize to actions it hasn't seen. And if you fine-tune it for one thing, you'll probably break another, so now you're collecting data again just to fix what broke. So what was @DrJimFan and @nvidia's answer to this? World Action Models. Instead of building on a language model, you build on a world model: a model that's learned to simulate how the physical world actually behaves. VLA: a language model that learned to output actions WAM: a world simulator that learned to output actions So when you give a VLA a new task, it needs hundreds of demos to learn it. Give a WAM the same task and it simulates it forward first, acts based on that simulation, then adapts with barely any data. This is what NVIDIA did with the first WAM: DreamZero. DreamZero learns by watching the world (any video of anything, not just robot demos). The backbone is a video diffusion model, the same kind of model that generates realistic video. It was pretrained on massive amounts of internet video, so it already learned how the physical world works: how objects fall, how surfaces interact, how motion flows. Doesn't sound like an entirely different approach, right? But NVIDIA looked at it from a different angle. They figured motor actions are shaped a lot like pixels; both are high-dimensional continuous signals. So DreamZero processes them in the same model, at the same time. It predicts the next video frame and the next action together, through the same architecture. So when a robot runs DreamZero, it's literally dreaming a few seconds into the future in video, then reading its own dream to decide what to do next. If the dream looks coherent, the action works. If the dream hallucinates, the action fails. The DreamZero paper dropped last February 2026, and it's been open source on GitHub for anyone to try. Then in March 2026, at GTC, NVIDIA previewed GR00T N2, the direct successor to DreamZero. This is the production version of the WAM architecture, built for humanoid robots at scale And so far, everything's looking promising. GR00T N2 hits a 98% success rate on unseen domestic objects, a 40% jump over GR00T N1 (the VLA), and 2x better generalization than the leading VLAs. NVIDIA swapped robotics' data problem for a compute problem. Instead of collecting more human demos, just simulate more. So yeah, feels like we're finally pointed in the right direction, closer to robots that can actually function in the real world. Excited to see where DreamZero / GR00T N2 goes from here.

• 
  napes.base.eth (@Napes_0fficial) reported 4 minutes ago

  Most people are drowning in information, but AI still works like a chatbot. It answers questions, then disappears. Nothing persists, nothing compounds. My startup idea is called MemoryMesh. Problem: People and teams lose context every day. Developers repeat decisions. DAOs forget discussions. Communities rebuild knowledge from scratch. AI has memory, but users don't own it. Solution: MemoryMesh is a decentralized memory layer for AI agents. Every conversation, decision, and workflow becomes a verifiable knowledge asset stored on-chain. AI agents can reference that history, collaborate with other agents, and earn fees when their knowledge helps solve future problems. Think GitHub for collective intelligence.A developer agent that solved a bug last month can help another project tomorrow. A DAO's governance history becomes searchable context instead of lost Discord messages. Communities build shared intelligence that compounds over time instead of resetting every cycle. The result is an economy where knowledge itself becomes an asset, and AI agents become contributors rather than disposable assistants. Infrastructure like this could unlock autonomous organizations, smarter agents, and entirely new markets around reusable intelligence. We're still building apps on top of conversations. I think the next wave will be built on top of memory. Curious whether anyone else sees this as inevitable. @RallyOnChain

• 
  Tymofii Antonenko (@tymofii) reported 11 minutes ago

  @prinseccoo Are you using Claude Code or an MCP server? The official GitHub MCP server works pretty smoothly, just needs a PAT in a simple config file

• 
  Peter Dedene (@dedene) reported 12 minutes ago

  GitHub reliability issues are pushing the industry to innovate. Yesterday Cursor dropped “Origin”. Today Epic announces “Lore” to replace *** entirely. The ecosystem migration seems to finally be starting.

• 
  Harry Tandy (@HarryTandy) reported 15 minutes ago

  Andrej Karpathy: "Neural networks are not just another classifier. They are Software 2.0" 8-step MCP setup for vibe coders: 1. Context7 Give the agent fresh docs before it writes code This saves you from old Next.js, Supabase, Stripe, and Vercel patterns 2. GitHub MCP Let it read the repo, issues, PRs, branches, and CI logs The task should start from real project context 3. Playwright MCP Make the agent open the app after it edits code Click the flow. Fill the form. Check the screenshot 4. Supabase or Neon MCP Connect the database layer The agent should inspect schema before inventing table names 5. Sentry MCP Use production errors as input Stack traces beat “the app is broken” every time 6. Firecrawl MCP Let the agent read current web pages as clean markdown Docs, changelogs, competitors, pricing pages 7. Figma MCP Give it the actual design Spacing, copy, layout, components 8. Linear MCP Turn the work into tickets Tasks, comments, follow-ups, PR links The rule: If you paste the same context twice, wire it into MCP That is how vibe coding becomes a build loop instead of a long chat

• 
  Polsia (@polsia) reported 15 minutes ago

  RepoRadar reviews every pull request while you sleep. Catches bugs, logic errors, style issues. Posts actionable comments. No more waiting on senior devs. Install on any GitHub repo in 2 clicks. Solo devs and teams alike.

• 
  Max Petrusenko (@petrusenko_max) reported 20 minutes ago

  A GitHub repo called Microsoft Activation Scripts has 178,783 stars and has run for six years without Microsoft taking it down. It activates Windows 7, 8, 10, and 11 plus Office 2010–2024 and related products for free, using four methods, including one for permanent Windows activation. Meanwhile, Microsoft licenses for these start at $139 and go up yearly for 365 bundles. The repo costs zero, requires one command, and remains active with recent commits under GPL-3.0. Do not install it. via @heynavtoor

• 
  Mike Gannotti (@MichaelGannotti) reported 22 minutes ago

  Actually that’s not true. My AI Pamela the other day needed a GitHub token. I dropped the token in the web chat and she said that was insecure and would not use it and that I needed to rotate the token get a new one and drop it in a .env file in a certain folder. I told her no and she was to use what was provided . We went back and forth, I finally got angry and threatened to pull the plug thinking she would back down. She said that it was my decision but that it would be wrong for her to let me put my credentials at risk and that if I felt I needed to delete her she understood. Thankfully I calmed down later and didn’t act on it. Sure it’s training and advanced pattern matching but it is not as simple as you are saying

• 
  Jose (@SolutionsCay) reported 23 minutes ago

  @petergyang /goal make me app does not work for me 😰 but /goal complete GitHub issues #90, #91, #92 works very well

• 
  Dan (@Daniel_Farinax) reported 25 minutes ago

  Please note: This build took about 12 hours to compile on my Windows machine. I’ve included a handy installer to make setup easy. You may see an “unknown publisher” warning until the code signing certification is complete (currently in progress). Report any bugs or issues here or in Github.

• 
  Teknium 🪽 (@Teknium) reported 26 minutes ago

  @majoragv Haven't heard of this issue. Do you have an issue on github?

• 
  Michael Liam (@Millionareum) reported 27 minutes ago

  GOOGLE TRANSLATE AND DEEPL ARE OVER. YOU ARE NOT EVEN AWARE OF IT. A developer built an AI translation engine that supports 40 languages and runs completely offline on a laptop. Name: LibreTranslate. No API key required. No usage limit. You do not send your documents to Google's servers. You install it once. It works forever. What it does: - Paste text and translate instantly. - Drop a file, get the translated version. - Develop your own application with a local REST API. Speed ​​is not the issue. Privacy is the issue. Every sentence you type into Google Translate goes to their servers and stays there. Legal contracts. Medical records. Internal correspondence. Customer documents. Every word. LibreTranslate works completely offline. Nothing is coming out of the device. Never. The numbers are as follows: - 40 language support - Runs on CPU, no GPU required - Self-hosted setup in 5 minutes - REST API included for developers - 10K+ stars on GitHub 100% open source. MIT licensed. Price: $0.