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:

• 
  FHILY👑 (@Oluwaphilemon1) reported 7 minutes ago

  the guy who built Claude Code stopped writing prompts. Boris Cherny writes loops now. agents that scan GitHub, Slack and X, pick the next step, code it, test it, fix themselves, repeat. nobody at the keyboard. it works. it's also brutal: Peter Steinberger ran 100 of these for a month and burned $1.3M in tokens. prompt engineering was the warm-up. loop engineering is the boss level, and right now it's gated by your token budget.

• 
  Vandos ❓ (@__vandos__) reported 8 minutes ago

  I ran Hermes Agent as standing infrastructure for 5 weeks on a $5 VPS. Here are the 17 prompts that made it actually work. Every morning at 7am — GitHub notifications, open PRs, what’s blocking what. Waiting in Telegram before I sit down. 35 minutes of triage gone. Every night at 11pm — scans today’s commits. Flags TODOs, console.logs, functions over 80 lines, changed paths with no tests. Short list waiting with coffee. Every weekday at 9:55am — stand-up already written. What closed, what’s in progress, what’s blocked. I walk in with it done. Every Friday at 6pm — research digest on my topic, deduped against last week so what lands is genuinely new. The mental model nobody explains: A prompt to a chat window is a question. A prompt to a persistent agent is a job description. It needs a trigger, a body, and an escalation rule. Drop any of the three and it either never fires or buries you in noise. The brief runs at 7 whether I’m awake or not. The repo watches itself on weekends. The research lands while I sleep. None of it competing for my attention because none of it needs me in the loop. Full 17 prompts, copy-paste ready 👇

• 
  Firas D (@firasd) reported 10 minutes ago

  @chrisalbon What these guys are skipping over is that they have a thousand github issues right They aren't prompting the agent directly with hey lets look at XYZ cause the github issue becomes the prompt

• 
  prabhakaran (@prabhakaranr91) reported 10 minutes ago

  6:00 AM: Todoist raw file processed. Tasks auto-created from overnight notes. If I forgot to log something at 1 AM, it still makes it to my morning list. 9:00 AM: Weekly ITSM/MSP research runs. 2-step pipeline: SearXNG for discovery, Firecrawl for deep-read. Not chatbot summaries. Actual article extraction into markdown, then HTML artifact, then GitHub Pages push. I review the output, not write it. 2:00 PM: Designer skills radar. Scans for new Figma plugins, SwiftUI patterns, and whatever SuperOps design team is shipping that week. Dumps into my vault as raw notes. I process the signal, not the noise. 6:00 PM: Raw vault file processing. This is the big one. Everything I dumped into ~/Documents/hermes-vault/raw/ throughout the day gets categorized, summarized, cross-referenced, and written into the wiki. If this cron breaks, my entire knowledge pipeline stalls. It broke last month because of an em-dash in a headline. Now I have a pre-flight script. 9:00 PM: Vault daily digest fires to Telegram. Shows me inbox count, wiki size, and whether any raw files got orphaned. If the digest is silent, something is wrong. I don't schedule these. Cron does. I just review the output and occasionally fix the thing that broke. The lesson: automation that needs babysitting is just delayed manual work.

• 
  Alex (@realalexniebuhr) reported 12 minutes ago

  @pavitrabhalla I don’t think it’s right. Define task as GitHub issues have one sandbox/agent per issue.. let that sandbox/agent work for hours than PR & merge

• 
  Xavier Rivera (@XavierRiveraX) reported 13 minutes ago

  Microsoft open-sourced Intelligent Terminal, an AI-native fork of Windows Terminal. Choose your agent (GitHub Copilot, Claude, Codex, or Gemini) and get automatic error detection, command fixes, and persistent session memory. Available on GitHub and the Microsoft Store now.

• 
  shinyufoguy2222 (@ollobrains) reported 14 minutes ago

  Microsoft’s MAI launch has a data-provenance problem, and enterprise buyers should require Microsoft to reconcile its marketing language with its own technical report. That is colder, harder to rebut, and more dangerous. 1. The strongest factual spine The core contradiction you want to exploit is real enough to be powerful, but it needs to be framed precisely. Microsoft’s public MAI materials say MAI-Thinking-1 was trained “from the ground up” on enterprise-grade, clean and commercially licensed data, and the Build transcript uses the phrase “enterprise-grade, clean and commercially licenced data lineage.” Microsoft also announced a family of seven MAI models developed in-house. But the MAI-Thinking-1 technical report describes pretraining on a mixture of publicly available and licensed human-generated data, including web data, public GitHub code, books, academic papers, news, multilingual text, and domain-specific materials. The same report says the data pipeline includes a proprietary crawl and Common Crawl, and the appendix says Microsoft’s proprietary crawl started from 1.2 trillion pages, later reduced by filtering, while the Common Crawl pipeline contributed 24.2 billion pages after processing. That is the strongest wording: Microsoft marketed the model around clean, enterprise-grade, commercially licensed data lineage. Its own paper describes a broader training mix that includes massive public-web crawling and Common Crawl. Microsoft needs to explain what, exactly, “commercially licensed” means here. That is much better than “they lied,” because it forces them into a definitional trap. 2. Do not make the whole argument depend on “lying” “Caught lying” is viral, but it gives Microsoft an escape hatch. They can say: “We did not lie. ‘Commercially licensed data lineage’ refers to commercially licensed datasets in the blend, not every token.” Or: “Publicly available web data was collected according to terms of use, robots controls, and industry standards.” Or: “Clean means filtered, deduplicated, non-pirated, non-adult, non-synthetic, and quality controlled — not necessarily individually licensed from every author.” So the sharper accusation is: Microsoft used procurement-grade language that sounds like full rights-cleared data, while its technical report describes a training stack that includes large-scale public-web data. That ambiguity matters for banks, hospitals, insurers, government agencies, and any customer relying on data-provenance representations. That is the kill shot: not “you lied,” but “your enterprise claim is materially ambiguous.” 3. Best replacement headline options Use one of these: Microsoft’s new MAI models have a data-provenance problem. Microsoft told enterprises “commercially licensed data.” Its own paper says public web crawl and Common Crawl. The real Microsoft AI story is not benchmarks. It is data lineage. Microsoft’s MAI launch just created a procurement problem for every regulated enterprise buyer. What does “commercially licensed” mean when the technical report says Common Crawl? The most surgical version: Microsoft needs to define “commercially licensed data lineage” before regulated enterprises rely on it. That sounds less like outrage and more like a subpoena. 4. Important correction: be careful saying Satya personally said it Your draft says: “At Build 2026, Satya Nadella announced…” The safer version is: At Build 2026, Microsoft announced… Or: During Microsoft’s MAI keynote at Build 2026, Microsoft described the models as having enterprise-grade, clean and commercially licensed data lineage. The Build transcript text I found includes the relevant “commercially licenced data lineage” claim in the MAI keynote, and the transcript also says “Satya just mentioned,” which suggests you should not hinge the argument on Satya personally unless you have the exact clip of him saying the phrase. That matters because if one attribution is wrong, people will use it to attack the entire post. 5. The real issue: “clean” is not the same as “licensed” This is one of the biggest missing elements. There are at least four separate claims being blurred together: Clean can mean filtered for spam, porn, piracy domains, malware, boilerplate, duplicate pages, low-quality content, synthetic content, or personally sensitive data. Commercially licensed means someone has a legal right, contractual permission, or license to use the content commercially for training. Traceable means Microsoft can identify source categories, URLs, datasets, providers, or pipeline lineage. Enterprise-grade means the data pipeline is controlled, documented, audited, filtered, and suitable for business buyers. Those are not the same thing. Your post should say: Clean data is not automatically licensed data. Traceable data is not automatically rights-cleared data. Publicly available data is not automatically commercially licensed data. That line is devastating. 6. Add this phrase: “traceable is not licensed” This is the obscure but important thought input. A dataset can be traceable to a URL and still not be commercially licensed. A crawler can respect robots.txt and still not create a negotiated license with the author. A page can be publicly accessible and still contain copyrighted work. A corpus can exclude adult content and piracy domains but still contain protected text, images, code, PDFs, journalism, forums, documentation, and books. Suggested line: Microsoft may have built a cleaner web crawl. That is not the same as building a fully commercially licensed corpus. Another version: The question is not whether Microsoft filtered the web. The question is whether Microsoft had commercial training rights for the web it kept. 7. The Common Crawl point needs nuance Your draft says Common Crawl has: “ZERO licensing guarantees and ZERO author consent mechanisms.” That is rhetorically strong, but tighten it. Better: Common Crawl is a free, open archive of web-crawled data. It does not, by itself, provide a source-by-source commercial license from every author or publisher whose content appears on the open web. Common Crawl describes itself as a free, open repository of web crawl data, and its own materials emphasize that it is a massive open corpus used by many AI researchers and companies. Also be careful saying: “Common Crawl is being sued.” A more accurate version is: Common Crawl and Common-Crawl-derived data have become a recurring flashpoint in AI copyright disputes, takedown demands, and lawsuits involving AI training data. That is safer. Wired reported publisher pressure on Common Crawl and noted that the New York Times had made removal requests before suing OpenAI, while other recent lawsuits have alleged use of Common Crawl or Common-Crawl-derived material as part of AI training disputes. 8. The best “receipt” structure Your post should have a clean evidence ladder: Microsoft’s claim: clean, enterprise-grade, commercially licensed data lineage. Microsoft’s paper: mixture of publicly available and licensed data. Microsoft’s appendix: proprietary crawl, Common Crawl, web-crawled PDFs, public GitHub. Enterprise question: which parts were actually commercially licensed, and which parts were merely publicly available? That is the entire argument. You do not need to prove Microsoft committed fraud. You need to force the question: Was “commercially licensed” a claim about the entire corpus, or only part of it? 9. The technical report gives you more ammunition Do not stop at proprietary crawl and Common Crawl. The MAI report also describes a web-crawled PDF corpus and a large public GitHub code corpus. The appendix says the web-crawled PDF pipeline starts from about 10 billion documents, filtered to about 620 million, and the report also describes a 7.4 trillion-token public GitHub code corpus. That matters because enterprise risk is not just “web pages.” It is: web pages PDFs public code academic material news books journals domain-specific materials multilingual web content possibly opt-out-protected or rights-reserved content Suggested line: This is not a tiny footnote. The paper describes a full industrial-scale public-data ingestion machine. 10. The Simon Willison angle is useful, but do not over-center it Simon Willison is useful as the “someone actually read the paper” character. He publicly highlighted the same issue: Microsoft’s marketing around “appropriately licensed” data versus the report’s description of public web crawl and Common Crawl. But the strongest post should not depend on Simon. Make him the discovery beat, not the proof: The funny part is that the contradiction was not hidden. It was sitting in Microsoft’s own technical report. Simon Willison simply did what enterprise procurement teams should have done: read the data section. That line is excellent. 11. Add the “enterprise procurement” frame This is the most important strategic upgrade. The issue is not whether Microsoft can win a copyright lawsuit. The issue is whether enterprise customers can rely on Microsoft’s claims in regulated procurement. A bank, hospital, insurer, defense contractor, or government agency does not ask only: “Can Microsoft argue fair use?” They ask: “Can Microsoft represent and warrant the training data provenance?” “Can Microsoft indemnify us?” “Can Microsoft disclose source categories?” “Can Microsoft honor opt-outs and rights reservations?” “Can Microsoft survive an audit?” “Can we put this in front of regulators, customers, and internal risk committees?” Suggested line: In regulated enterprise sales, “probably defensible in court” is not the same as “clean enough for procurement.” That is a killer sentence. 12. Add “Model Data Bill of Materials” This is the genius-level solution. Borrow from software supply chain language. Enterprises already understand SBOMs — Software Bills of Materials. AI now needs a Model Data Bill of Materials, or MDBOM. Suggested paragraph: If Microsoft wants to sell frontier models into regulated industries, it should provide a Model Data Bill of Materials: source categories, token share by source family, acquisition method, license basis, opt-out handling, robots/rights-reservation handling, third-party provider categories, public-code license treatment, jurisdictional restrictions, audit process, and indemnity scope. That reframes the post from outrage to standards-setting. You are not just complaining. You are proposing the new enterprise AI procurement checklist. 13. Ask for percentages The missing question is not: “Did you use Common Crawl?” The missing question is: What percentage of the final training tokens were commercially licensed versus publicly available? Ask Microsoft to publish: percentage from proprietary web crawl percentage from Common Crawl percentage from web-crawled PDFs percentage from public GitHub percentage from books and journals percentage from news percentage from academic papers percentage from commercially negotiated third-party providers percentage from user/customer data, if any percentage excluded due to rights reservations, robots.txt, noai tags, or publisher opt-outs That is the exact pressure point. 14. Add “commercially licensed” definition demand This should be central: Microsoft needs to define whether “commercially licensed data lineage” means:100% of training tokens were commercially licensed; only third-party purchased datasets were commercially licensed; publicly available web data was treated as appropriately usable under site terms, robots controls, or fair-use theory; only the data pipeline, not the underlying works, was commercially controlled; or the phrase was marketing shorthand, not a source-level warranty. That list is brutal because Microsoft has to pick one. 15. Add “lineage laundering” This is an obscure but powerful concept. Use this carefully: The danger here is data-lineage laundering: taking messy public-web material, running it through proprietary crawlers, filters, deduplication, embeddings, and safety screens, then describing the resulting pipeline as enterprise-grade without clearly saying which underlying works were actually licensed. Or shorter: Processing data in-house does not magically license the underlying works. That line is excellent. 16. Add “in-house” is not the same as “rights-cleared” Microsoft’s “built in-house” positioning is important because it helps them show independence from OpenAI. But “in-house” answers a different question. It answers: Did Microsoft build the model itself? It does not answer: Did Microsoft have commercial training rights for every work in the corpus? Suggested line: “Built in-house” is a model-lineage claim. “Commercially licensed” is a data-rights claim. Microsoft is blending the emotional force of the first with the procurement value of the second. That is a high-quality thought. 17. Add “zero distillation” is a separate axis Microsoft emphasizes MAI-Thinking-1 was trained without distillation from third-party models. That matters because enterprise buyers worry about models inheriting another model’s outputs or hidden IP. But zero distillation does not solve web-data provenance. Suggested line: Zero distillation may answer “did you copy another model’s behavior?” It does not answer “were the underlying training works commercially licensed?” This is a very strong missing distinction. 18. Strengthen the DeepSeek angle Your draft says the DeepSeek scandal made compliance departments paranoid about where AI training data came from. Make it more precise: The DeepSeek controversy sharpened enterprise sensitivity around model lineage, unauthorized distillation, and IP contamination. It made provenance a boardroom issue, not just a research issue. Reuters reported that Microsoft and OpenAI were probing whether a DeepSeek-linked group improperly obtained OpenAI data, and later reported on U.S. government concern about alleged unauthorized use and distillation by DeepSeek and other Chinese AI firms. Better line: DeepSeek made “where did the intelligence come from?” a procurement question. Microsoft then marketed MAI around exactly that fear. 19. Use the OpenAI partnership as context, not motive Your draft says Microsoft was “desperate” to break free from OpenAI and “rushed” models to market. That is spicy, but not provable unless you have internal evidence. Use this instead: The timing matters. Microsoft and OpenAI’s April 2026 partnership update made Microsoft’s OpenAI license non-exclusive, ended Microsoft’s revenue-sharing payments to OpenAI, and allowed OpenAI to serve products across other cloud providers. That gave Microsoft a clear strategic incentive to prove it could ship first-party models. Then: That does not prove bad faith. But it does explain why the “built in-house on clean, commercially licensed data” story was so valuable. This is much stronger. 20. The EU AI Act point is good, but make it sharper Your EU point is valid, but refine it. The EU AI Act’s Article 53 requires providers of general-purpose AI models to maintain technical documentation, put in place a policy to comply with EU copyright law including rights reservations, and publish a sufficiently detailed summary of the content used for training using an AI Office template. The European Commission has also published a training-data summary template and says the GPAI Code of Practice helps providers demonstrate compliance with Article 53 obligations around transparency and copyright. Suggested line: Europe is where vague provenance language turns into paperwork. Even stronger: In the U.S., Microsoft can fight this as a marketing and litigation-risk issue. In the EU, the question becomes: what exactly goes into the Article 53 training-data summary and copyright compliance policy? 21. Add the “enterprise warranty” angle The most important hidden issue is not the blog post. It is the contract. Suggested paragraph: The question for every enterprise buyer is whether Microsoft’s “clean and commercially licensed” language appears only in launch materials, or whether it is incorporated into the Master Services Agreement, procurement response, model documentation, indemnity terms, risk disclosures, or regulatory submissions. That is the procurement bomb. If it is only marketing copy, buyers were sold vibes. If it is contractual, Microsoft may owe precise representations. Suggested line: Marketing language can be slippery. Contract language cannot. 22. Add this procurement checklist This is what banks, hospitals, insurers, and government agencies should ask Microsoft before deploying MAI models: Define “commercially licensed data lineage.” State whether the claim applies to 100% of training tokens. Break down training-token percentages by source category. Identify which categories were commercially licensed. Identify which categories were merely publicly available. Disclose whether Common Crawl was used directly, indirectly, or after filtering. Disclose which Common Crawl snapshots were used. Explain the legal basis for proprietary web-crawled content. Explain how robots.txt, paywalls, login walls, noai tags, and TDM reservations were handled. Explain whether rights-holder opt-outs were honored before, during, and after training. Explain whether source websites’ terms of service were reviewed or categorized. Explain the treatment of public GitHub licenses, including copyleft and attribution obligations. Explain how web-crawled PDFs were screened for copyrighted books, journals, reports, and paywalled documents. Explain what “piracy filtering” means and whether it is domain-based, content-based, hash-based, or provider-list-based. Explain whether Microsoft can remove a source from future training runs. Explain whether Microsoft can identify if a specific publisher, website, repository, or author appears in the corpus. Provide the scope of Microsoft’s indemnity for training-data copyright claims. Provide an Article 53-style training-data summary for EU deployments. Provide third-party audit or attestation of data-source claims. Put all of the above into the contract, not a blog post. That list turns your post into something procurement teams can actually use. 23. Add a data-source risk table This would make the post feel more concrete: Source categoryWhat the paper indicatesEnterprise questionProprietary web crawlStarted from 1.2T pages, filtered down substantiallyWere these pages commercially licensed, or publicly accessible?Common Crawl24.2B processed pagesWhich snapshots, what legal basis, what opt-outs?Web-crawled PDFsStarted from about 10B documents, filtered to 620MWere reports, journals, books, manuals, and paywalled PDFs excluded?Public GitHub7.4T-token code corpusHow were licenses, attribution, copyleft, and code provenance handled?Books/journalsAcquired from providers, including publisher agreementsWhich categories were truly licensed, and under what usage rights? The technical basis for that table is in Microsoft’s own report. 24. Add “public web is not a license class” This is a great phrase: Public web is an access category, not a license category. Follow with: A browser can access a page. That does not mean a model vendor has a commercial training license to ingest, tokenize, transform, and monetize that work. That is one of the strongest conceptual additions. 25. Add “robots.txt is not consent” Microsoft’s report says publicly available data was collected in a way that considered site terms, industry standards, and web controls such as robots.txt and meta tags. That is useful, but it does not settle the enterprise question. Suggested line: Robots.txt is a crawler instruction mechanism. It is not the same thing as a negotiated commercial license from every rights holder. Be careful: do not say this as a definitive legal conclusion for every jurisdiction. Frame it as a procurement distinction. 26. Add “AI-generated content exclusion does not solve copyright” Microsoft says AI-generated content was excluded from pretraining and frames this as important for quality, provenance, and control. Good missing line: Excluding AI-generated content may reduce model-collapse and synthetic-data contamination risk. It does not automatically clear rights in the remaining human-authored web. Even punchier: Clean of synthetic sludge does not mean clean of copyright ambiguity. 27. Add “the real scandal is definitional” This is probably the best intellectual framing. The scandal is not that Microsoft used web data. Everyone suspects frontier labs use web data. The scandal is that Microsoft appears to have used enterprise-procurement language that ordinary buyers could reasonably interpret as rights-cleared data, while the technical report describes a massive public-web component. That is balanced and sharp. 28. Add “MAI-Thinking-1 vs all seven models” Do not overextend the technical paper. The strongest evidence is for MAI-Thinking-1 / MAI-Base-1, because that is the technical report with the data details. Microsoft’s broader MAI announcement says the seven models share clean, enterprise-grade data lineage, but the most specific crawl/Common Crawl details you are citing are from the MAI-Thinking-1 report. Suggested wording: The clearest contradiction appears in the MAI-Thinking-1 technical report. If Microsoft says the same data-lineage claim applies across the whole seven-model family, then it should publish equivalent data-source summaries for each model. That avoids overclaiming. 29. Add “do not say every Fortune 500 company wrote checks” Your draft says: “Procurement teams across Wall Street and Washington heard ‘clean and commercially licensed’ and started writing checks.” Unless you have evidence of contracts closed after that keynote, say: “That claim was clearly aimed at enterprise procurement teams across finance, healthcare, insurance, and government.” Or: “That is exactly the kind of claim regulated buyers use to clear vendor-risk reviews.” That preserves the point without inventing buyer behavior. 30. Add “do not say Microsoft has issued no statement” unless you are tracking it This line is risky: “As of today, Microsoft hasn’t issued a single public statement…” Instead say: “Microsoft should publicly reconcile the launch claim with the technical report.” That keeps the pressure without making a fragile negative claim.

• 
  chrisreedbates (@chrisreedbates) reported 21 minutes ago

  @asaio87 It's just a cron to wakeup and do whatever you tell it to do. I use /loop 5mn. poll Github to check for mergable PRs, decide if it's mergable, if it is, merge it, if it's not leave a comment to the PR owner why. Then i can use /loop on the PR writers to look fro comments back from the merging session. Fix the issues, resubmit, etc.

• 
  REVENGE ARC (I'M HIM. BIO/ACC) (@RetardedNi85688) reported 23 minutes ago

  Been deep in physical AI infrastructure this week. So far two projects stood out for me. $Codec — I found out that robotics is still stuck rebuilding the same infrastructure from scratch. Every team builds their own simulation, data pipeline, training stack, deployment tools. CodecFlow is the shared layer underneath. Simulate in the browser, train on serverless GPU, deploy through a unified runtime. This is typically Vercel for robotics. $CODEC has been live and proven for a while as well. @sentinelstoday — also found out robots are getting smarter but nobody solved who they actually are. No cryptographic identity. No verified firmware. No tamper-proof telemetry. No auditable receipts when an action happens. $Sent is building exactly that. Ed25519 keypairs, hardware attestation, signed telemetry, firmware anchored on @solana, machine wallets for autonomous payments. The same trust infrastructure the internet built for users and servers — finally built for machines. $SENT is still very much early. GitHub went from boilerplate to real software across four repositories in 48 hours. Daily public commits. The macro validates both. SoftBank reportedly pouring $300M+ into Agile Robots as part of an $800M round. Stanford and ETH Zurich just published a paper arguing the robotics field has been solving the wrong problem — the bottleneck isn't the model, it's the missing software infrastructure. Billions flowing into the hardware. Nobody pricing the software stack. CodecFlow builds and deploys robots. @sentinelstoday secures and audits them. Different layers. Same cycle. both are still early. $CODEC $SENT DYrWewaqjmiMpnTh8SGzfo9NkiTzFckTTmnRMDQypump

• 
  Carver (@carverfomo) reported 29 minutes ago

  A Chinese mathematician posted a 3 minute video on Bilibili explaining how he lost his $10,000 a month gig to AI. The model he had been training started writing harder math problems than he could invent. He admitted his own mistake in business positioning. He had spent four years hand writing PhD level math problems for Scale AI's reinforcement learning pipeline. $50 to $100 per problem. 200 problems a month. Then synthetic data killed his entire contract category. He was no longer able to invent a problem the machine could not solve. At 2:13 he says the word agent. He says it once. He never says it again in the video. The way he says it is the only thing on screen that did not come off the teleprompter. He has been recording videos off a teleprompter for three months. The teleprompter runs on the same agent that killed his Scale AI work. Every script is generated by Claude. Every word he reads to camera is the agent's. The new job is reading. Someone pulled the script repository from a Cursor instance the dev had left public. The folder was labeled bilibili-laments. Inside were 47 video scripts. All in his voice. All written by Claude. Six months ago a 14 year old in Shenzhen pushed an AI agent to GitHub. Judges said no real world application. 3,100 forks later. The mathematician had been one of them. He had wired the agent into his content pipeline the week Scale AI cut him off. He had been a PhD candidate at one of the top five Chinese math schools. He taught there for two years before going full time on Scale AI contracts. He still has the credentials. He still has the office. He just no longer writes anything. He wanted to show people how AI took his career. He accidentally showed them how AI also took his post mortem.

• 
  prabhu💢 (@prabhu_ai) reported 29 minutes ago

  I want to build and ship a useful Chrome extension 🚀 What’s one annoying browser problem you face regularly? It could be related to tabs, bookmarks, productivity, reading, writing, dev tools, forms, LinkedIn, GitHub, YouTube, or anything else. Drop the problem below 👇 If it’s practical, I’ll build and ship it.

• 
  0xSero (@0xSero) reported 38 minutes ago

  @atomtanstudio @OnlyTerp If it’s still broken please open an issue on GitHub

• 
  Advanced Installer Powers PacKit FREE (@advinst) reported 38 minutes ago

  Danut Ghiorghita walks through the whole thing: code push to signed MSI to live updater config, fully automated. What's covered: → Advanced Installer + GitHub Actions setup → Automated build, versioning + code signing → Updater config generated in the pipeline → What breaks in real pipelines and how to fix it first

• 
  小乔不带伞|| Make money forever || (@xiaoqiao6666666) reported 39 minutes ago

  4RdPweUWkqt7oqSZY6gah7ktH7bKmFaomGYfkpPdpump @sharbel actually created an AI girlfriend named Sophia—solving a major problem. Installation is completely free; all fees go to the GitHub repository to support Sophia's development.

• 
  HiJack (@theDCcapital) reported 39 minutes ago

  Thread: Why $PRL is a speculative asset, not a real compute network 1/5 $PRL has an elegant narrative: GPU miners run matrix multiplication, produce "useful" AI compute, earn tokens. The problem? The compute quality is almost useless for real AI workloads. Current implementation uses exact integer MatMul. Real AI training/inference needs FP16/BF16. These are fundamentally different. 2/5 The Together AI partnership looks like validation. It's not. Together AI is using $PRL emissions to subsidize inference costs — giving users a 25% discount. That's a marketing experiment, not real compute demand. One client does not prove a business model. 3/5 Compare $PRL to $TAO (Bittensor): TAO: subnet competition, quality-based rewards, validators filter bad outputs, real model marketplace PRL: prove you ran MatMul, get tokens, no quality evaluation layer TAO solves AI model quality incentives. PRL solves nothing that AWS can't solve cheaper. 4/5 On-chain data tells the real story: Token supply heavily concentrated among early miners Wallet addresses still very low Trading only on minor venues with thin liquidity No second B2B partnership announced Forced liquidity exit = price collapses. Project team knows this. 5/5 Most likely path forward for $PRL team: Build user numbers → use as leverage to negotiate with exchanges and capital → swap tokens for market making But with weak tech and concentrated supply, exchanges don't want the reputational risk. Watching. Not holding. What signal would change my mind? → A second real B2B AI compute buyer. → BF16/FP16 upgrade on GitHub. $PRL $TAO #DePIN #AICompute