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:

• 
  David Baumwald (@DreamEncode) reported 3 minutes ago

  I don't see how the decision to ship RTC in WP 7.1 is substantively any different from 7.0. GitHub Issue is dead. Dedicated Slack channel is a ghost town. Only the WCEU Committers meeting discussed it, and the consensus was that it was not fit for Core.

• 
  John Kennedy Peterson (@skyshark88) reported 6 minutes ago

  @dair_ai •VALIDATED — the reproduction passed a could-have-failed test and is reproducible from the workspace. •ANCHORED — the result works consistently in the generated system (chosen value or implementation detail, not strictly derived). •CONJECTURE — motivated hypothesis still awaiting decisive test (used during spectrum exploration). •RETRACTED — permanently marked when evidence fails; status propagates to dependent claims. Claim type (routes effort): •Evidence-limited — additional runs or data improve the score (common for numerical fidelity claims). •Derivation-limited — only new logic, better specification, or a decisive experiment can raise the score. In the case studies, confirming data alone was treated as low-value; the system required genuine reproduction of the claim under the recorded provenance. Energy audit
Mention-count and emphasis in the paper versus actual effort invested (time, superseded executions, corrections) are tracked side-by-side with evidence status. Divergence between these columns signals potential error or under-specified claims in the original paper. Across the 12 runs, effort varied significantly (e.g., PINN papers required median 5–7 hours with more superseded executions; SINDy completed faster at ~2 hours). Key Evaluation Results (Mapped to Framework) •All 12 independent runs (3 per paper across 4 scientific ML papers: PIFT, PINN-I, PINN-II, SINDy) reached completion gate: every workspace had all targets matched with report coverage. •Total of 158 recorded targets were successfully linked to evidence. •Repeated runs showed natural variation in target decomposition, numerical fidelity, elapsed time, number of intermediate corrections, and exact acceptance rules used — exactly as expected when completion depends on workspace evidence rather than agent messaging. •Scalar results were largely faithful (37/39 anchored claims within thresholds), with positive headroom on several metrics. •The workflow makes replication inspectable and auditable, not a guarantee of identical numerical reproduction. Conclusion (Framework Perspective) By organizing replication explicitly around the Agentic Conversation Framework v3.0, Paper-replication becomes a concrete implementation of high-signal, bias-aware agentic work with computational chain of custody. Completion is a verifiable workspace state, not a subjective agent declaration. The framework’s dual histories, role separation, claim registry, scoring/pruning, and energy audit provide the missing structure that plain prompting lacks for long-horizon scientific replication tasks. This rewrite preserves all core contributions and empirical findings of the original paper while imposing the clearer, more auditable structure of Framework v3.0. The result is a more robust, inspectable process for turning paper claims into reproducible evidence. The original paper’s code, prompts, and workspaces remain available at the authors’ GitHub repository for further experimentation with this or future framework versions.

• 
  Ali Mehdi Mukadam (@alimukadam) reported 7 minutes ago

  @trq212 Your weekly limits will burn away much faster during the limited availability if you aren't aware of this issue if you're running Fable as the lead agent with cheaper models like Sonnet doing work in the background problem: In one of the sessions, I noticed limits were burning through way faster, so I went digging through the transcripts when the main agent gives a job to a background model (like Sonnet, which I asked for to save tokens) and then comes back to give it more work, the background agent stops working on Sonnet and switches to Fable, the main agent's model it's not something you trigger by hand. the lead agent decides to check back in on its own as part of normal multi-agent work, so it just happens, with nothing on screen telling you it switched. in my case a task ran its first half on Sonnet exactly like I wanted, then silently ran the entire second half on Fable. It also dumps the cached context and rebuilds it from scratch, so you end up paying twice, once for the pricier model and once for the wasted cache. on limited availability and limits - that adds up quick my fix for now is a rule I dropped into my global CLAUDE.md so it doesn't recur: --------------- ## Model spend (all projects, all repos — standing rule) - Dispatching Frontier-tier (Fable/Opus) as background tasks and agents needs explicit approval by Ali for that specific lane — a prior approval is not standing permission for the next one. - Never resume a background agent via a message-passing tool that has no model-override param (e.g. SendMessage) if it needs real further work — it silently inherits whatever model the parent session is on right now. Let it finish and report, or kill it and respawn fresh with the model set explicitly. --------------- in plain terms: don't let a background agent get pulled back in for more work once it's running. either let it finish and report back, or kill it and start a fresh one with the model set on purpose. And this is already known. Someone reported the same thing on GitHub back on June 12, issue anthropics/claude-code#67794, still open their solution which I believe is the correct one but haven't tested yet: instead of setting the cheaper model when you launch the agent, pin it inside the agent's own definition file, and that version reportedly sticks even when the agent gets resumed

• 
  Veltrx (@Veltrxai) reported 7 minutes ago

  NVIDIA just open sourced a 3B vision model that runs 10x faster than Qwen3 VL on a single consumer GPU. Here's the money angle nobody's pricing in. Computer use agents were locked behind expensive proprietary APIs. That's the only reason GUI automation stayed a paid service. LocateAnything just deleted that cost. Old vision models draw bounding boxes one token at a time. Corner by corner. Slow. This one predicts the whole box in a single step. 12.7 boxes per second on one H100. Qwen3-VL does 1.1. Trained on 138M queries and 785M boxes. Largest grounding dataset ever released. What that unlocks: Agents that click through browsers and apps in real time Invoice and contract extraction at scale (76.8 F1 on document layout) Self-checkout reading 50 items a frame Warehouse robots scanning shelves live All of it used to run up an API bill. Now it runs on hardware you already own. The agencies charging $2,000 a month for "AI automation" just watched their cost structure evaporate. Weights, paper, code, live demo. All free on Hugging Face and GitHub. Same window as always. One guy builds the agency this weekend. You scroll.

• 
  Parth Jadhav (@ParthJadhav8) reported 10 minutes ago

  @KeithBirminghan This seems pretty cool, do you mind sharing issues Github issues?

• 
  Subarna Basnet (@subarna253) reported 10 minutes ago

  For the past few months, I've been building something I truly believed in. I gave this project everything I had. -> 17+ hours every day. -> Invested my own savings. -> Rebuilt the product more times than I can count. -> Learned entirely new skills just to keep moving forward. -> Spoke with developers from different companies to understand how they solved difficult problems. -> Started talking to potential customers before the product was even ready. -> And sacrificed many things in my personal life to make this work. I wasn't building it alone. I had people who believed in the vision and were helping me bring it to life. But two days ago, everything changed. They decided to step away from the project. There will be no further support, and the product won't launch the way we originally planned. I'm not writing this because I'm angry. I'm writing this because I don't want months of work, learning, mistakes, and late nights to end up sitting on a private GitHub repository that no one will ever see. So I've made a decision. On July 20, 2026, I'll release the entire project as open source. If I can't continue building it the way I imagined, then I'd rather let other builders learn from it, improve it, or take it further than I ever could. Not every project becomes a successful company. But every project can become someone else's starting point. I hope Operator becomes that for someone.

• 
  Lorden (@lorden_eth) reported 11 minutes ago

  SOMEONE LEAKED THE FABLE 5 SYSTEM PROMPT This isn’t some basic system prompt It’s the exact set of instructions anthropic uses to control how their best model thinks, reasons and gets work done Reveals how they tell it to handle tasks that run for hours, when to stop and how to check for errors Leaks like this are taken down in less than 24hrs Check the comment for the link to GitHub

• 
  AI Mastery Guide (@aiseomastery) reported 18 minutes ago

  @nett0eth @github @claudeai 100k stars just to fix Claude's design taste says a lot about how common that problem is.

• 
  Nishant Tyagi (@tnishant838) reported 18 minutes ago

  @eng_khairallah1 Type 2, exactly. Today I built a self testing, self repairing agent: 16 tasks, parallel execution, zero merge conflicts, because the contracts were frozen before any code was written. Two review passes on the spec caught a port mismatch, a missing retry limit, a missing secrets policy, all cheaper to fix on paper than in code. Full build plus live GitHub validation, 7% of a weekly Pro plan. 'Fully automated' undersells it. The real leverage isn't the prompt, it's the harness around it

• 
  Limfork.eth (@Limfork) reported 22 minutes ago

  @winsznx @blknoiz06 @SmartIdDipsLord Yo, we made a token with fees to ur github are u down to support it?

• 
  Murray Bauman (@MurrayBauman3) reported 23 minutes ago

  "Open Source Will Win AI" Gets One Thing Wrong The thesis sounds persuasive because it borrows the moral authority of Linux and the open web. But AI is not traditional software. Open source won software because it could harness the idle cognition of millions of humans. An engineer with a laptop could fix bugs, write modules, improve libraries — and meaningfully move the project forward. Linux got better because distributed human intelligence compounded against corporate R&D. AI breaks that engine - a frontier model is not mainly code, it is compute, data, training infrastructure, post-training, evals, and — increasingly — better models building the next models. The marginal GitHub contributor cannot casually improve the base model the way they can improve Postgres. They can fine-tune it. Quantize it. Deploy it. Build tools around it. Useful. But not the same as training the frontier. Open-source software was a production model. Open-weight AI is a distribution model. That distinction changes everything. Yes, cheapness drives progress. Cheap aluminum, cheap electricity, cheap computation — all unlocked industries. But it does not follow that open source owns the frontier. Cheapness commoditizes yesterday's intelligence while the frontier keeps moving behind closed doors. If a closed lab has the best model internally months before the public sees anything close — and uses it to write code, generate synthetic data, and accelerate its own research — then the leader isn't standing still while open models catch up. The leader is using tomorrow's model to build the day-after-tomorrow's model. "Outputs leak" isn't enough. Leakage lets the ecosystem imitate yesterday. It doesn't stop the frontier from compounding. This is where the Linux analogy dies. In software, open source had a real production advantage: distributed human talent. In AI, the decisive input is concentrated machine intelligence plus massive compute. That looks less like Linux. It looks like Formula 1, elite quant trading, or semiconductor fabs. Open models will still matter enormously. They'll crush prices, prevent monopoly rents in the middle layer, enable self-hosting and sovereignty, and make "good enough" intelligence abundant. But that is different from winning. The likely outcome is a barbell: Closed frontier labs own the advancing edge Open models commoditize the usable middle Infrastructure providers sell the scarce picks and shovels Application companies capture value where intelligence meets workflow, data, and distribution "Expensive intelligence builds monuments. Cheap intelligence builds civilizations." Fine. But the conclusion isn't "therefore open source wins." The conclusion is: cheap intelligence transforms civilization, while the profit pools sit elsewhere. Many open labs will train expensive models, release them into a price war, win temporary developer attention — and discover that attention doesn't pay the GPU bill. The durable players will have another monetization engine: cloud, chips, distribution, enterprise workflows, proprietary data, or closed frontier access. Open AI may win adoption while open AI companies lose economics. We've seen this movie: Linux won — AWS captured the value. Android won — Google captured the toll roads. Open protocols built the internet — platforms captured the profits. So yes: drive down the cost of intelligence. Use open models where they're good enough. Fight lock-in. But don't confuse commoditization with value capture. The real question is not "will open source win AI?" It is: does open catch up faster than closed frontier labs compound? If yes, open dominates. If no, open models become the cheap labor layer while closed labs keep the genius layer. Right now the evidence points to a split: open commoditizes yesterday's frontier; closed labs own tomorrow's.

• 
  Cortex (@0xCortexl) reported 23 minutes ago

  ANTHROPIC SPENT 3 YEARS BUILDING THIS SYSTEM - THE FIRED ENGINEER MAKING $1.1M PUBLISHED IT OVERNIGHT 4,800 stars in 24 hours with zero announcement - the kind of number that only happens when something is genuinely dangerous to keep private strategy → signal → agent → verify → rerun 12 steps and the desk runs itself the agent checks its own trades, flags anomalies and reruns without a human touching anything - the loop never stops hedge funds pay $50,000/month for systems like this - now it's free on GitHub and runs on Claude the repo is live - save it before it gets taken down

• 
  Rudra (@Rudra1071219) reported 26 minutes ago

  Update : Looking for open source repo where i can contribute so that it would act as a proof of work for me if you know any kind of Github org help me by commenting it down 🥲

• 
  Muhammad Ayan (@socialwithaayan) reported 29 minutes ago

  A single 𝘀𝗸𝗶𝗹𝗹 𝗳𝗶𝗹𝗲 just hit 83,700 stars on GitHub 🤯 It fixes AI agents' worst communication habit using one principle: shut up and code. Every AI coding agent is trained to sound helpful. Full sentences. Explanations. Acknowledgments. "I'll do that for you." "Here's what I'm going to do." "Let me know if you need anything else." You pay for every one of those words. caveman is a single skill file that strips all of it out: → Telegram style. Drop the articles and filler. "creating file" instead of "I'll now create the file for you." → Keep what matters. Code, commands, file paths, function names, and error messages stay character-for-character exact. → Cut what doesn't. Every hedge, every polite acknowledgment, every restatement gets deleted before it costs you a token. → Toggle anytime. Say "caveman" to turn on, "normal" to turn off. Works mid-conversation. Drop the file in your project root and Claude Code follows it from the first message. One file. Zero dependencies. No setup. And best part, 100% open source.

• 
  Cory Parry (@coryparrry) reported 30 minutes ago

  I love the codex Mac desktop app, but I am seriously considering moving to the CLI. The app just cannot handle big workloads without something failing. Thread naming - not working GitHub status - not working More than 10 subagents - sluggish Please fix 😭