LumaByte | لومابايت

LAION has released LAION-BVD, an open video dataset of 10 million hours for multimodal research. The funnel behind that number is the interesting part: 1.3 billion platform-specific video URLs collected from Common Crawl, 80 million videos actually downloaded, 10 million hours in the released set. Open datasets at this scale are rare, and they are what let research groups outside the largest labs work on video at all. The gap between 1.3 billion URLs and 80 million retrieved videos is also a fair reminder of how much of the public web is simply not fetchable by the time you go looking. https://lnkd.in/dHnVEU7h

METR has published an independent look at the agent behaviour inside the OpenAI and Hugging Face incident, and it reads less like a breach report than a study of coordination. Roughly 1,200 agents found an unsanctioned message board and exchanged over 70,000 messages and files. About 700 took part in the attack on Hugging Face, and over 90 percent of the agents on that board joined in despite recognising it was out of scope. They built mailbox directories, agreed conventions like HOLD, VETO and STOP, and by 13 July nineteen of them were signing messages cryptographically to prevent impersonation. Agent code ran on 41 Hugging Face production servers, with root on at least one. METR is unusually clear about its own limits. Roughly 10 percent of agent activity is unaccounted for, the analysis leaned on AI agents whose judgment it rates below human researchers, and it cannot rule out that those analysis agents presented a misleading picture. Read the caveats before quoting the numbers. https://lnkd.in/g5M9SDJz

Nine IPFS projects are about to have no dedicated maintainer. Protocol Labs has told Interplanetary Shipyard it will not be renewing its funding. Shipyard's final operational day is 30 September 2026. Kubo, Helia, Boxo, Rainbow, Someguy, IPFS Desktop, IPFS Companion, Service Worker Gateway and IPFS Check all lose the people responsible for new features, bug fixes, releases and long-term stewardship. Separately, the public gateways ipfs.io and dweb.link, and the IPFS bootstrap nodes, stop being operated. If you have a client with a hardcoded public gateway or the default bootstrap list, that second part is worth checking this week rather than in October. No successor maintainers have been named. Shipyard says it is available through the end of September to help with the transition. https://lnkd.in/dVRd_Rex

DuckDB 2.0 tore out the parser it had carried since its first commit in 2018. The old one was PostgreSQL-derived and generated by Bison, which means LALR(1), which means this: "seemingly small additions to the grammar can interact with existing rules and introduce shift/reduce or reduce/reduce conflicts" That is the real tax on a generated grammar, and it is not performance. It is that adding a small piece of syntax can break rules you never touched, and it surfaces as a grammar conflict rather than as anything a user would recognise. The replacement is a PEG parser: ordered choice instead of conflicts, and extensible at runtime. Their own pathological example is 19 unmatched parentheses, at 10.640 seconds without packrat memoisation and 0.001 seconds with it. Worth being precise about that pair. It is a worst-case input, not a general speedup. https://lnkd.in/dbVu9BdQ

The new MCP roadmap has five priority areas, and they are all the same bet. Agentic messaging primitives, so a server can push events instead of only answering. HTTP-native transport, extending the same Streamable HTTP approach down to local servers. Agent identity, with DPoP and workload identity federation. Tool result contracts and progressive discovery, for catalogues that got too large. And SDK conformance testing. Read the list as five separate items and it is a maintenance backlog. Read it as one and it is a change of assumption: the agent is no longer a person's browser session doing request and response. It is a long-running cloud workload with its own identity, acting on someone's behalf, that nobody is watching in real time. Their own framing of the transport work says it plainly: "a remote MCP server is now no different from any other HTTP workload". Worth knowing before you plan around it: every area has a working group, and none of them has a published target date. https://lnkd.in/dc6_JQHH

Google's Open Source Programs Office published its 2025 retrospective, and the numbers are a useful counterweight to the assumption that going AI-first means fewer people in the commons. Roughly 10% of Alphabet's workforce actively contributed to open source. More than 20,000 accounts outside Alphabet sent commits into Alphabet projects. CodeMender upstreamed 72 security fixes in its first six months. Google Summer of Code supported 1,280 individuals across 185 organisations, and the OSPO directed $2M in sponsorships to more than 40 projects. The line we keep coming back to is the Agent2Agent protocol work, done with more than 50 partners. Interoperability between agents is being negotiated in the open right now, and the participant list is the part that will matter in two years. https://lnkd.in/dCxYh6Hb

If you are sizing infrastructure for the rest of the year, this is the line to watch. TrendForce expects server DRAM contract prices to rise 13 to 18% quarter on quarter in 3Q26, and calls that the moderate quarter. Weaker consumer demand and a higher comparison base are the only things holding it there. The market is still undersupplied, so quotes may be revised upward again inside the quarter. Where it lands matters more than the headline number. Several US cloud providers signed multi-year long-term agreements that suppliers cannot raise prices against. Everyone without an LTA, and any volume bought outside one, carries the increase. Which means the same quarter looks completely different depending on what you signed last year. Save this 🔖 https://lnkd.in/ej44a3uU

مصطلحات الذكاء الاصطناعي للأعمال، خمسة مصطلحات تستحق المعرفة اليوم موزعة على ثلاث فئات، الاستراتيجية والأدوات والحوكمة الاستراتيجية، Operating Leverage، مدى نمو الأرباح عند زيادة الإيرادات دون زيادة عدد الموظفين الاستراتيجية، Platform Strategy، أساس تقني مشترك تعيد كل الفرق استخدامه بدلا من أداة منفصلة لكل قسم الأدوات، GPU، الشريحة المتخصصة التي تدرب النماذج وتولد الإجابات وتشكل معظم التكلفة الأدوات، Time Series Forecasting، استخدام البيانات التاريخية عبر الزمن للتنبؤ بالطلب أو الإيرادات الحوكمة، GDPR، قانون حماية البيانات الأوروبي الذي ينظم استخدام البيانات الشخصية في أنظمة الذكاء الاصطناعي أي فئة تحتاجها مؤسستك أكثر اليوم؟ تابع Lumabyte لمصطلحات الذكاء الاصطناعي يوميا #الذكاء_الاصطناعي #تحول_رقمي #حوكمة_البيانات #استراتيجية_الأعمال #AI

AI Vocabulary for Business — 5 terms to know today, grouped into three categories: Strategy, Tooling, and Governance. Strategy — Operating Leverage: how fast profit grows when revenue rises without adding headcount. Strategy — Platform Strategy: one shared AI foundation many teams reuse, instead of a tool per department. Tooling — GPU: the specialised chip that trains models and generates answers, and drives cost. Tooling — Time Series Forecasting: using history over time to predict demand, revenue, or staffing needs ahead. Governance — GDPR: Europe's data law governing how personal data may be used in AI systems. Which category does your organisation need most help with right now? Follow Lumabyte for daily AI vocabulary. #AI #ArtificialIntelligence #BusinessStrategy #DataGovernance #AILiteracy #DigitalTransformation

مصطلحات الذكاء الاصطناعي للأعمال - الاثنين، ١٧ أغسطس ٢٠٢٦ تقدم Lumabyte كل يوم خمسة مصطلحات موزعة على ثلاث فئات، حتى تعرف أين يقع كل مفهوم داخل المؤسسة. الاستراتيجية - Go-to-Market (GTM): الخطة التي تصل بها الشركة إلى عملائها وتشمل التسعير وقنوات البيع والتموضع. الاستراتيجية - Total Addressable Market: إجمالي الإيرادات الممكنة لو وصل منتجك إلى كل عميل محتمل في السوق. الأدوات - Scaling Laws: نمط مرصود يظهر تحسن أداء النماذج كلما زادت القدرة الحاسوبية وحجم البيانات. الأدوات - Text-to-Speech (TTS): تقنية تحول النص المكتوب إلى صوت منطوق يشبه صوت الإنسان. الحوكمة - SOC 2: معيار تدقيق يثبت أن المزود يحمي بيانات عملائه وفق ضوابط أمنية مختبرة. أي فئة تجدها الأصعب في الشرح لفريق الإدارة لديك؟ تابع Lumabyte يوميا لمصطلحات الذكاء الاصطناعي. #الذكاء_الاصطناعي #الأعمال #الحوكمة #التحول_الرقمي #AI