Gul Saeed Khattak

NVIDIA Open-Sources Auto Ontology, a Semantic Layer for AI Agents Your boss asks the AI agent: "What was revenue last quarter?" It answers in 3 seconds. Clean number. Nice chart. But it used the wrong "revenue" column. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗻𝗼𝘁𝗶𝗰𝗲𝘀. That's the problem NVIDIA's new open-source library, Auto Ontology, goes after. Its README puts it plainly: "…𝘢 𝘳𝘦𝘷𝘦𝘯𝘶𝘦 𝘮𝘦𝘵𝘳𝘪𝘤 𝘢𝘯𝘴𝘸𝘦𝘳𝘦𝘥 𝘧𝘢𝘴𝘵𝘦𝘳 𝘪𝘴 𝘸𝘰𝘳𝘵𝘩 𝘯𝘰𝘵𝘩𝘪𝘯𝘨 𝘪𝘧 𝘵𝘩𝘦 𝘯𝘶𝘮𝘣𝘦𝘳 𝘪𝘴 𝘸𝘳𝘰𝘯𝘨." 𝗪𝗵𝗮𝘁 𝗶𝘁 𝗶𝘀: A "context layer" that sits between an AI agent and your database. It tells the agent where the data lives, what each field means, and whether to use it. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀, 𝘀𝘁𝗲𝗽 𝗯𝘆 𝘀𝘁𝗲𝗽: • 𝗗𝗶𝘀𝗰𝗼𝘃𝗲𝗿: it reads your database metadata and drafts an ontology, a map of business terms like "active customer" • 𝗚𝗿𝗼𝘂𝗻𝗱: agents use those terms instead of guessing from raw table names • 𝗩𝗮𝗹𝗶𝗱𝗮𝘁𝗲: it checks generated SQL against certified definitions • 𝗘𝘅𝗽𝗼𝘀𝗲: agents reach it through MCP, a CLI or an API • 𝗟𝗲𝗮𝗿𝗻: certified corrections go back into the map 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀: Ask a question, get 𝘁𝗵𝗲 𝗮𝗻𝘀𝘄𝗲𝗿, 𝘁𝗵𝗲 𝗦𝗤𝗟 𝗶𝘁 𝗿𝗮𝗻, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗿𝗼𝘄𝘀. You can check the work, not just trust it. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀: • Apache 2.0, runs with Docker Compose or Helm • Works with Databricks, Postgres, Snowflake and DuckDB • Defaults to NVIDIA's Nemotron models, but 𝗮𝗻𝘆 𝗢𝗽𝗲𝗻𝗔𝗜-𝗰𝗼𝗺𝗽𝗮𝘁𝗶𝗯𝗹𝗲 𝗲𝗻𝗱𝗽𝗼𝗶𝗻𝘁 𝘄𝗼𝗿𝗸𝘀 • The MCP server has 𝟵 𝗿𝗲𝗮𝗱-𝗼𝗻𝗹𝘆 𝘁𝗼𝗼𝗹𝘀. There is deliberately no tool to browse raw tables, so agents must go through the business terms 𝗜𝘁 𝗯𝘂𝗶𝗹𝗱𝘀 𝗼𝗻 𝗲𝗮𝗿𝗹𝗶𝗲𝗿 𝘄𝗼𝗿𝗸: • Cube and dbt MetricFlow: semantic layers you write by hand in code • Wren AI: an open semantic layer for AI agents, with a context layer built partly by agents • Snowflake's Semantic View Autopilot: also drafts semantic models with AI, but only for Snowflake What's different here: 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲, 𝘄𝗼𝗿𝗸𝘀 𝗮𝗰𝗿𝗼𝘀𝘀 𝗳𝗼𝘂𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀, 𝗮𝗻𝗱 𝗯𝘂𝗶𝗹𝘁 𝗳𝗼𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗵𝗿𝗼𝘂𝗴𝗵 𝗠𝗖𝗣. 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁: • 𝗧𝗵𝗲 "𝟵𝟬% 𝗳𝗮𝘀𝘁𝗲𝗿" 𝗰𝗹𝗮𝗶𝗺 𝗶𝘀 𝗡𝗩𝗜𝗗𝗜𝗔'𝘀 𝗼𝘄𝗻 𝘁𝗲𝘀𝘁𝗶𝗻𝗴, with no baseline published • 𝗡𝗼 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆 𝘀𝗰𝗼𝗿𝗲𝘀 𝗮𝗿𝗲 𝗽𝘂𝗯𝗹𝗶𝘀𝗵𝗲𝗱. NVIDIA ships an eval harness (BIRD, 1,534 questions) but no results • That harness was tested with 𝗖𝗹𝗮𝘂𝗱𝗲 𝗢𝗽𝘂𝘀 𝟰.𝟴, 𝗻𝗼𝘁 𝗡𝗲𝗺𝗼𝘁𝗿𝗼𝗻 • 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗛𝗮𝘀𝗵𝗶𝗖𝗼𝗿𝗽 𝗩𝗮𝘂𝗹𝘁, 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲 𝗽𝗮𝘀𝘀𝘄𝗼𝗿𝗱𝘀 𝗮𝗿𝗲 𝘀𝘁𝗼𝗿𝗲𝗱 𝘂𝗻𝗲𝗻𝗰𝗿𝘆𝗽𝘁𝗲𝗱 in Postgres (per the README) • It's young: 𝗮𝗯𝗼𝘂𝘁 𝟳𝟱𝟬 𝗚𝗶𝘁𝗛𝘂𝗯 𝘀𝘁𝗮𝗿𝘀, and NVIDIA isn't accepting outside contributions 𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗼𝗶𝗻𝘁: the hard part of AI analytics isn't writing SQL. 𝗜𝘁'𝘀 𝗮𝗴𝗿𝗲𝗲𝗶𝗻𝗴 𝗼𝗻 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲 𝘄𝗼𝗿𝗱𝘀 𝗺𝗲𝗮𝗻. GitHub: github.com/NVIDIA/auto-ontol… Links in reply 👇

𝗜𝗻𝘀𝘁𝗮𝗴𝗿𝗮𝗺 𝗜𝘀 𝗣𝗮𝘆𝗶𝗻𝗴 𝗬𝗼𝘂𝗧𝘂𝗯𝗲𝗿𝘀 𝗨𝗽 𝘁𝗼 $𝟱𝟬𝟬,𝟬𝟬𝟬 𝘁𝗼 𝗣𝗼𝘀𝘁 𝗟𝗼𝗻𝗴 𝗩𝗶𝗱𝗲𝗼𝘀 You finish a 40-minute video. You upload it to YouTube, like always. Now Instagram wants you to upload it there too, on the same day. And for a few creators, it's paying. 𝗪𝗵𝗮𝘁'𝘀 𝗵𝗮𝗽𝗽𝗲𝗻𝗶𝗻𝗴: Business Insider reports that Instagram has offered select YouTube creators $𝟱𝟬,𝟬𝟬𝟬 to post long videos to its new "Features" format. Some got six-figure offers, and some reached $𝟱𝟬𝟬,𝟬𝟬𝟬. 𝗪𝗵𝗮𝘁 𝗙𝗲𝗮𝘁𝘂𝗿𝗲𝘀 𝗶𝘀: • 𝗟𝗼𝗻𝗴: horizontal videos up to 𝟵𝟬 𝗺𝗶𝗻𝘂𝘁𝗲𝘀, inside the Instagram app. • 𝗡𝗲𝘄: announced on 26 October, tested in Brazil first, now with select US creators. • 𝗢𝗻 𝗧𝗩 𝘁𝗼𝗼: the Instagram for TV app on Apple TV will show it. 𝗧𝗵𝗲 𝗱𝗲𝗮𝗹 𝘁𝗲𝗿𝗺𝘀, 𝗮𝘀 𝗿𝗲𝗽𝗼𝗿𝘁𝗲𝗱: • 𝗡𝗼𝘁 𝗲𝘅𝗰𝗹𝘂𝘀𝗶𝘃𝗲: creators can keep posting on YouTube. • 𝗦𝗮𝗺𝗲-𝗱𝗮𝘆 𝗽𝗼𝘀𝘁𝗶𝗻𝗴: one source says creators are asked to put any long video they publish on YouTube onto Instagram that same day. Meta didn't confirm the amounts. A spokesperson said: "𝘛𝘩𝘦 𝘥𝘦𝘵𝘢𝘪𝘭𝘴 𝘰𝘧 𝘵𝘩𝘰𝘴𝘦 𝘥𝘦𝘢𝘭𝘴 𝘢𝘳𝘦 𝘤𝘰𝘯𝘧𝘪𝘥𝘦𝘯𝘵𝘪𝘢𝘭, 𝘣𝘶𝘵 𝘸𝘦'𝘳𝘦 𝘦𝘹𝘤𝘪𝘵𝘦𝘥 𝘵𝘰 𝘴𝘦𝘦 𝘵𝘩𝘦 𝘤𝘰𝘯𝘵𝘦𝘯𝘵 𝘨𝘰𝘪𝘯𝘨 𝘭𝘪𝘷𝘦." 𝗙𝗼𝗿 𝗰𝗿𝗲𝗮𝘁𝗼𝗿𝘀: a few get paid up front. Everyone else gets a new place to post, but 𝗻𝗼 𝗽𝘂𝗯𝗹𝗶𝗰 𝗮𝗱-𝗿𝗲𝘃𝗲𝗻𝘂𝗲 𝘀𝗵𝗮𝗿𝗲 𝘆𝗲𝘁. Adam Mosseri wrote that "𝘵𝘩𝘦 𝘦𝘤𝘰𝘯𝘰𝘮𝘪𝘤𝘴 𝘮𝘢𝘵𝘵𝘦𝘳 𝘢 𝘵𝘰𝘯" and that Instagram is exploring options. 𝗙𝗼𝗿 𝘃𝗶𝗲𝘄𝗲𝗿𝘀: more long videos in the app where you already scroll. 𝗪𝗲'𝘃𝗲 𝘀𝗲𝗲𝗻 𝘁𝗵𝗶𝘀 𝗯𝗲𝗳𝗼𝗿𝗲: Instagram launched IGTV in 2018 for long videos. It paid creators for shows in 2021 and even tried an ad-revenue share. IGTV was shut down in 2022. Meta has also paid creators for Reels through bonus programs. Those programs have shrunk since. 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗽𝗹𝗮𝘆𝗲𝗱 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝗴𝗮𝗺𝗲, 𝗶𝗻 𝗿𝗲𝘃𝗲𝗿𝘀𝗲: When short vertical video took off, YouTube launched Shorts, its answer to TikTok and Instagram Reels. To get creators posting there, it paid them. The $𝟭𝟬𝟬 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 Shorts Fund paid creators in 2021 and 2022. Then it switched to a lasting deal: from February 2023, Shorts creators get 𝟰𝟱% 𝗼𝗳 𝘁𝗵𝗲𝗶𝗿 𝘀𝗵𝗮𝗿𝗲 𝗼𝗳 𝗮𝗱 𝗿𝗲𝘃𝗲𝗻𝘂𝗲. I think Instagram is now doing to YouTube 𝘄𝗵𝗮𝘁 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗱𝗶𝗱 𝘁𝗼 𝗜𝗻𝘀𝘁𝗮𝗴𝗿𝗮𝗺. YouTube came for short video. Instagram is coming for long video. Same playbook: 𝗰𝗼𝗽𝘆 𝘁𝗵𝗲 𝗳𝗼𝗿𝗺𝗮𝘁, 𝗽𝗮𝘆 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗰𝗿𝗲𝗮𝘁𝗼𝗿𝘀, 𝘁𝗵𝗲𝗻 𝗵𝗼𝗽𝗲 𝗮 𝗿𝗲𝘃𝗲𝗻𝘂𝗲 𝘀𝗵𝗮𝗿𝗲 𝗸𝗲𝗲𝗽𝘀 𝘁𝗵𝗲𝗺. YouTube finished that playbook. Instagram is still on step two. 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁: • The dollar figures come from 𝗽𝗲𝗼𝗽𝗹𝗲 𝗳𝗮𝗺𝗶𝗹𝗶𝗮𝗿 𝘄𝗶𝘁𝗵 𝘁𝗵𝗲 𝗱𝗲𝗮𝗹𝘀, not from Meta. • We don't know 𝗵𝗼𝘄 𝗺𝗮𝗻𝘆 𝗰𝗿𝗲𝗮𝘁𝗼𝗿𝘀 got offers. • YouTube pays eligible creators 𝟱𝟱% 𝗼𝗳 𝗮𝗱 𝗿𝗲𝘃𝗲𝗻𝘂𝗲 on long videos. Instagram has nothing public to match that yet. • Both platforms borrowed short vertical video from TikTok, so neither invented the format it's now fighting over. 𝘗𝘢𝘺𝘪𝘯𝘨 𝘵𝘩𝘦 𝘵𝘰𝘱 𝘤𝘳𝘦𝘢𝘵𝘰𝘳𝘴 𝘨𝘦𝘵𝘴 𝘵𝘩𝘦 𝘧𝘪𝘳𝘴𝘵 𝘷𝘪𝘥𝘦𝘰𝘴. 𝘗𝘢𝘺𝘪𝘯𝘨 𝘦𝘷𝘦𝘳𝘺𝘰𝘯𝘦 𝘦𝘭𝘴𝘦 𝘪𝘴 𝘸𝘩𝘢𝘵 𝘮𝘢𝘬𝘦𝘴 𝘵𝘩𝘦𝘮 𝘴𝘵𝘢𝘺. Links in reply 👇 socialmediatoday.com/news/in…

Qoduck free event: docs.qoder.com/events/qoduck… Download Qoder: qoder.com/download

𝗡𝗼𝘁 𝗶𝗻 𝘁𝗵𝗲 𝗤𝘄𝗲𝗻 𝟰 𝗘𝗮𝗿𝗹𝘆 𝗔𝗰𝗰𝗲𝘀𝘀? 𝗧𝗿𝘆 𝘁𝗵𝗲 Qwen 4 𝗦𝘁𝗲𝗮𝗹𝘁𝗵 𝗠𝗼𝗱𝗲𝗹 𝗜𝗻𝘀𝘁𝗲𝗮𝗱 🦆 Early testers are already posting what they say are Qwen 4 demos. Most of us don't have access yet. But there's something you 𝗰𝗮𝗻 try today. Qoder just launched "𝗤𝗼𝗱𝘂𝗰𝗸", a stealth model. Its maker hasn't been revealed. Qoder is Alibaba's coding tool, so many people are guessing it's an unreleased Qwen. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗵𝗮𝘀 𝗰𝗼𝗻𝗳𝗶𝗿𝗺𝗲𝗱 𝘁𝗵𝗮𝘁. 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂 𝗴𝗲𝘁: • 𝗔 𝗳𝗿𝗲𝗲 𝗱𝗮𝗶𝗹𝘆 𝗮𝗹𝗹𝗼𝘄𝗮𝗻𝗰𝗲 from 7 November • 𝟬 𝗰𝗿𝗲𝗱𝗶𝘁𝘀 𝗽𝗲𝗿 𝗰𝗮𝗹𝗹 within it, even at the highest thinking level • Free accounts are eligible too 𝗛𝗼𝘄 𝘁𝗼 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝘁𝗿𝘆 𝗶𝘁: • Make a Qoder account • Download the 𝗹𝗮𝘁𝗲𝘀𝘁 𝘃𝗲𝗿𝘀𝗶𝗼𝗻 of the app • Pick 𝗤𝗼𝗱𝘂𝗰𝗸 in the model picker 𝗢𝗻𝗲 𝗰𝗮𝘁𝗰𝗵: the event page says the Qoder IDE, JetBrains plugin and CLI "𝘤𝘶𝘳𝘳𝘦𝘯𝘵𝘭𝘺 𝘥𝘰 𝘯𝘰𝘵 𝘴𝘶𝘱𝘱𝘰𝘳𝘵 𝘵𝘩𝘦 𝘱𝘳𝘰𝘮𝘰𝘵𝘪𝘰𝘯", without saying which product does. Check the model picker in your app first. 𝗔𝗹𝘀𝗼 𝗴𝗼𝗼𝗱 𝘁𝗼 𝗸𝗻𝗼𝘄: • The daily limit isn't stated, and it may slow down at busy times • No end date yet • Teams and Enterprise accounts are excluded 𝘛𝘳𝘺 𝘪𝘵 𝘰𝘯 𝘺𝘰𝘶𝘳 𝘰𝘸𝘯 𝘵𝘢𝘴𝘬𝘴, 𝘵𝘩𝘦𝘯 𝘵𝘦𝘭𝘭 𝘮𝘦: 𝘸𝘩𝘰 𝘥𝘰 𝘺𝘰𝘶 𝘵𝘩𝘪𝘯𝘬 𝘮𝘢𝘥𝘦 𝘪𝘵? Links in reply 👇 Download Qoder: qoder.com/download

Qwen 4 demos, and Qoder has a mystery “stealth model” running for free. Same model? Still a question, not a claim.

Early access eh

Qwen4 “Qoder just launched a free stealth model called ‘Qoduck’ 🦆. Identity: undisclosed. But: Qoder is Alibaba’s. Qwen teased ‘The becoming begins’ the night before. One tracker says its tokenizer matches the Qwen family. Is this Qwen 4 in disguise?”🦆 docs.qoder.com/events/qoduck…

Discounts Links forum.cursor.com/t/student-v… support.spotify.com/us/artic… oneplus.com/es/discount gemini.google/students/ support.google.com/youtube/a…

𝗪𝗵𝗮𝘁 𝗛𝗮𝗽𝗽𝗲𝗻𝘀 𝗪𝗵𝗲𝗻 𝗔𝗻𝘆𝗼𝗻𝗲 𝗖𝗮𝗻 𝗣𝗿𝗲𝘁𝗲𝗻𝗱 𝘁𝗼 𝗕𝗲 𝗮 𝗦𝘁𝘂𝗱𝗲𝗻𝘁? A free tool wants to make everyone a student. 𝗔𝘁 𝗹𝗲𝗮𝘀𝘁 𝗼𝗻 𝗽𝗮𝗽𝗲𝗿. Why? 𝗧𝗼 𝗰𝗹𝗮𝗶𝗺 𝘁𝗵𝗲 𝘀𝘁𝘂𝗱𝗲𝗻𝘁 𝗱𝗶𝘀𝗰𝗼𝘂𝗻𝘁. Many AI companies give students their best tools for free or cheap. It's a smart bet: win people while they're learning. And the list of student offers is long. Perplexity has offered students 𝘂𝗽 𝘁𝗼 𝟮 𝘆𝗲𝗮𝗿𝘀 𝗼𝗳 𝗣𝗿𝗼. Google has a 𝗳𝗿𝗲𝗲 𝗚𝗲𝗺𝗶𝗻𝗶 𝗽𝗹𝗮𝗻 for students. Spotify Premium Student and YouTube Premium give students 𝗱𝗶𝘀𝗰𝗼𝘂𝗻𝘁𝘀 𝗳𝗼𝗿 𝘂𝗽 𝘁𝗼 𝟰 𝘆𝗲𝗮𝗿𝘀. That's just a few. But it depends on one thing: 𝗸𝗻𝗼𝘄𝗶𝗻𝗴 𝘄𝗵𝗼 𝗶𝘀 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗮 𝘀𝘁𝘂𝗱𝗲𝗻𝘁. That job usually goes to a checker like SheerID. Perplexity, Spotify and YouTube all use it. Think of it as 𝘁𝗵𝗲 𝗯𝗼𝘂𝗻𝗰𝗲𝗿 𝗰𝗵𝗲𝗰𝗸𝗶𝗻𝗴 𝗜𝗗𝘀 at the door of a student party. This week, a new open-source tool appeared that 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗲𝘀 𝘁𝗵𝗲 𝘁𝗿𝗶𝗽 𝗽𝗮𝘀𝘁 𝘁𝗵𝗲 𝗯𝗼𝘂𝗻𝗰𝗲𝗿. It queues up many attempts, uses a browser designed to avoid bot detection, and even types with random pauses to look human. In short, it tries to turn 𝘀𝘁𝘂𝗱𝗲𝗻𝘁 𝘃𝗲𝗿𝗶𝗳𝗶𝗰𝗮𝘁𝗶𝗼𝗻 𝗶𝗻𝘁𝗼 𝗮 𝘀𝗰𝗿𝗶𝗽𝘁 𝗷𝗼𝗯. It collected about 𝟴𝟴𝟬 𝘀𝘁𝗮𝗿𝘀 𝗮𝗻𝗱 𝟱𝟬𝟬+ 𝗳𝗼𝗿𝗸𝘀 𝗶𝗻 𝗮 𝗱𝗮𝘆. 𝗧𝗵𝗲 𝗰𝗮𝘀𝗲 𝗳𝗼𝗿 𝗰𝗼𝗻𝗰𝗲𝗿𝗻: • The features are built for 𝗱𝗼𝗶𝗻𝗴 𝗶𝘁 𝗺𝗮𝗻𝘆 𝘁𝗶𝗺𝗲𝘀, not once. • It builds school email addresses from a name and a school domain. • Abuse already existed: in 2019 SheerID said 𝟯𝟬–𝟯𝟱% of requests came from people pretending to qualify. 𝗧𝗵𝗲 𝗼𝘁𝗵𝗲𝗿 𝘀𝗶𝗱𝗲: • The author says it's "𝘧𝘰𝘳 𝘵𝘦𝘢𝘤𝘩𝘪𝘯𝘨 𝘢𝘯𝘥 𝘳𝘦𝘴𝘦𝘢𝘳𝘤𝘩" and forbids faking identities. • It doesn't create fake documents, and a person still has to pass the human check. • Nobody has shown that it works. 𝗪𝗵𝗮𝘁 𝘄𝗲 𝗮𝗰𝘁𝘂𝗮𝗹𝗹𝘆 𝗸𝗻𝗼𝘄: student offers are valuable, so they attract gaming, and gaming leads to stricter checks. 𝘐𝘧 𝘵𝘩𝘦 𝘣𝘰𝘶𝘯𝘤𝘦𝘳 𝘨𝘦𝘵𝘴 𝘴𝘵𝘳𝘪𝘤𝘵𝘦𝘳 𝘧𝘰𝘳 𝘦𝘷𝘦𝘳𝘺𝘰𝘯𝘦, 𝘸𝘩𝘰 𝘨𝘦𝘵𝘴 𝘭𝘦𝘧𝘵 𝘰𝘶𝘵𝘴𝘪𝘥𝘦: 𝘵𝘩𝘦 𝘤𝘩𝘦𝘢𝘵𝘦𝘳𝘴, 𝘰𝘳 𝘵𝘩𝘦 𝘳𝘦𝘢𝘭 𝘴𝘵𝘶𝘥𝘦𝘯𝘵𝘴 𝘸𝘪𝘵𝘩 𝘮𝘦𝘴𝘴𝘺 𝘱𝘢𝘱𝘦𝘳𝘸𝘰𝘳𝘬? Links in reply 👇

Agree on the direction, but “never read the code” has a catch: abstractions leak. The person who can drop down a level when the agent fails quietly is the one you can’t replace. Maybe the red flag is reading every line, not reading any line.

AI launches used to be a blog post and a benchmark table. Now we get a 2,300-year-old Chinese poem and a blank space. Qwen, it worked. I’m awake.

Also up in anticipation here 👀 If there’s an ambassador group forming, count me in too. Two of us not sleeping now. Qwen, room for one more ambassador? 😄

The catch proved the physics and control systems can work. The bigger test is economics. Can you repeat it hundreds of times without the tower becoming a maintenance bottleneck or a single point of failure? That's where an engineering breakthrough becomes an industrial one.

Agreed as decision note beats Slack archaeology every time. The catch: most changes never get one. That’s where citations matter. If the agent has to show the Jira ticket, PR and thread it used, you can see when it guessed who decided.

𝗬𝗼𝘂𝗿 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗖𝗮𝗻'𝘁 𝗙𝗶𝗻𝗱 𝘁𝗵𝗲 𝗔𝗻𝘀𝘄𝗲𝗿 𝗕𝗲𝗰𝗮𝘂𝘀𝗲 𝗜𝘁'𝘀 𝗦𝗽𝗿𝗲𝗮𝗱 𝗔𝗰𝗿𝗼𝘀𝘀 𝗙𝗶𝘃𝗲 𝗔𝗽𝗽𝘀 Someone asks: "Why did we change the retry logic in the billing worker?" The answer exists. Part of it is in a Jira ticket. Part in a Slack thread. Part in a pull request. Part in a design doc on Drive. 𝗡𝗼 𝘀𝗶𝗻𝗴𝗹𝗲 𝗮𝗽𝗽 𝗵𝗮𝘀 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗮𝗻𝘀𝘄𝗲𝗿. So most AI assistants give you half of it. PipesHub is an open-source project built for exactly this. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀, 𝘀𝘁𝗲𝗽 𝗯𝘆 𝘀𝘁𝗲𝗽: • 𝗖𝗼𝗻𝗻𝗲𝗰𝘁 𝘆𝗼𝘂𝗿 𝗮𝗽𝗽𝘀: Slack, Google Drive, Jira, Confluence, GitHub, Microsoft 365, Notion and 50+ more. • 𝗜𝗻𝗱𝗲𝘅 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴 𝗼𝗻𝗰𝗲: it parses, chunks and embeds the content, and links people, tickets, PRs and docs in a knowledge graph. • 𝗔𝗻𝘀𝘄𝗲𝗿 𝘄𝗶𝘁𝗵 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗹𝗼𝗼𝗽: it searches by meaning and by keyword, greps for exact IDs and error strings, follows graph links, and opens the full document instead of trusting one snippet. • 𝗥𝗲𝘀𝗽𝗲𝗰𝘁 𝗽𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻𝘀: every search runs as the person asking, so you only see what you're already allowed to see. • 𝗖𝗶𝘁𝗲 𝗲𝘃𝗲𝗿𝘆𝘁𝗵𝗶𝗻𝗴: each answer points to the exact blocks it came from. 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀: one search box across your company's tools, with sources you can click. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀: an MCP server, so Claude Code, Cursor or Codex can use the same index. Plus SDKs for Python, TypeScript and Go. You bring the model, including local ones through Ollama. 𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀: On EnterpriseRAG-Bench, a public benchmark run by Onyx (500 questions over 500,000+ documents), PipesHub is #1 with 𝟵𝟯.𝟮𝟰. For comparison: OpenAI File Search 𝟲𝟭.𝟬𝟯, Amazon Q 𝟰𝟴.𝟵𝟲, Azure AI Search 𝟰𝟴.𝟰𝟮, Vertex AI Search 𝟰𝟭.𝟴𝟳. The team also tested Claude with and without the index on 100 work questions. Facts found: • 𝗛𝗮𝗶𝗸𝘂: 55% → 80% • 𝗦𝗼𝗻𝗻𝗲𝘁: 65% → 89% • 𝗢𝗽𝘂𝘀: 85% → 92% 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁: • The 100-question test is 𝘀𝗲𝗹𝗳-𝗿𝗲𝗽𝗼𝗿𝘁𝗲𝗱, on PipesHub's own company data. The questions are private, so nobody can reproduce it yet. They say a public version is coming. • The leaderboard top three are 𝘄𝗶𝘁𝗵𝗶𝗻 𝟬.𝟲 𝗽𝗼𝗶𝗻𝘁𝘀 of each other (93.24, 92.74, 92.64). "#1" is real, but it's a close race. • Some features, like an agent acting on behalf of a person, are in the paid Enterprise edition. The core is Apache 2.0. • It runs a real stack: Neo4j or ArangoDB, Qdrant, MongoDB, Redis. The installer is one command, but it's a real service to run and keep updated. 𝗧𝘄𝗼 𝗳𝗲𝗮𝘁𝘂𝗿𝗲𝘀 𝗜 𝗳𝗶𝗻𝗱 𝗴𝗲𝗻𝘂𝗶𝗻𝗲𝗹𝘆 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴: 1. 𝗣𝗲𝗿𝗺𝗶𝘀𝘀𝗶𝗼𝗻𝘀 𝗮𝗿𝗲 𝗰𝗵𝗲𝗰𝗸𝗲𝗱 𝗮𝘁 𝗾𝘂𝗲𝗿𝘆 𝘁𝗶𝗺𝗲. The README: access is "𝘳𝘦𝘴𝘰𝘭𝘷𝘦𝘥 𝘸𝘩𝘦𝘯 𝘵𝘩𝘦 𝘲𝘶𝘦𝘳𝘺 𝘳𝘶𝘯𝘴, 𝘢𝘨𝘢𝘪𝘯𝘴𝘵 𝘵𝘩𝘦 𝘴𝘰𝘶𝘳𝘤𝘦 𝘴𝘺𝘴𝘵𝘦𝘮'𝘴 𝘰𝘸𝘯 𝘱𝘦𝘳𝘮𝘪𝘴𝘴𝘪𝘰𝘯𝘴, 𝘪𝘯𝘴𝘵𝘦𝘢𝘥 𝘰𝘧 𝘣𝘦𝘪𝘯𝘨 𝘢𝘱𝘱𝘳𝘰𝘹𝘪𝘮𝘢𝘵𝘦𝘥 𝘢𝘵 𝘣𝘶𝘪𝘭𝘥 𝘵𝘪𝘮𝘦." An agent connects as a specific person via a Personal Access Token, not as an app with god-mode access. 2. 𝗜𝘁'𝘀 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝗹𝗼𝗼𝗽, 𝗻𝗼𝘁 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗥𝗔𝗚 𝗽𝗮𝘀𝘀. Tools: hybrid search, grep for exact identifiers, knowledge-graph navigation, and full-document reading. Coding agents get it over MCP. There are SDKs for Python, TypeScript and Go. 𝗦𝘁𝗮𝗰𝗸: Next.js, Node.js, Python services; Neo4j or ArangoDB, Qdrant, MongoDB, Redis, Kafka for bigger setups. Any LLM, including Ollama. 𝗧𝗵𝗲 𝗹𝗮𝗻𝗱𝘀𝗰𝗮𝗽𝗲, 𝗼𝗻𝗲 𝗹𝗶𝗻𝗲 𝗲𝗮𝗰𝗵: Glean made this category famous as a closed product. • 𝗚𝗹𝗲𝗮𝗻: the closed, enterprise-sales version of this idea. • 𝗢𝗻𝘆𝘅 (𝗳𝗼𝗿𝗺𝗲𝗿𝗹𝘆 𝗗𝗮𝗻𝘀𝘄𝗲𝗿): Onyx (formerly Danswer) is the best-known open-source take.; also runs the main benchmark. • 𝗥𝗔𝗚𝗙𝗹𝗼𝘄: open-source RAG engine with deep document parsing. • 𝗔𝗻𝘆𝘁𝗵𝗶𝗻𝗴𝗟𝗟𝗠: local-first, great for individuals and small teams. • 𝗔𝗺𝗮𝘇𝗼𝗻 𝗤, 𝗔𝘇𝘂𝗿𝗲 𝗔𝗜 𝗦𝗲𝗮𝗿𝗰𝗵, 𝗩𝗲𝗿𝘁𝗲𝘅 𝗔𝗜 𝗦𝗲𝗮𝗿𝗰𝗵, 𝗢𝗽𝗲𝗻𝗔𝗜 𝗙𝗶𝗹𝗲 𝗦𝗲𝗮𝗿𝗰𝗵: the cloud defaults. • 𝗣𝗶𝗽𝗲𝘀𝗛𝘂𝗯: open source (Apache 2.0), 50+ connectors, per-user permissions, an agent loop, and an MCP server so Claude Code and Cursor can use it. 𝗪𝗵𝗲𝗿𝗲 𝗣𝗶𝗽𝗲𝘀𝗛𝘂𝗯 𝗳𝗶𝘁𝘀: it's betting the main customer is 𝘵𝘩𝘦 𝘤𝘰𝘥𝘪𝘯𝘨 𝘢𝘨𝘦𝘯𝘵, not the person in a chat box. Repo: about 𝟰,𝟬𝟬𝟬 𝘀𝘁𝗮𝗿𝘀 and v1.0 shipped on 27 October. 𝘛𝘩𝘦 𝘯𝘦𝘹𝘵 𝘴𝘵𝘦𝘱 𝘧𝘰𝘳 𝘈𝘐 𝘢𝘵 𝘸𝘰𝘳𝘬 𝘪𝘴𝘯'𝘵 𝘢 𝘴𝘮𝘢𝘳𝘵𝘦𝘳 𝘮𝘰𝘥𝘦𝘭. 𝘐𝘵'𝘴 𝘢𝘯 𝘢𝘨𝘦𝘯𝘵 𝘵𝘩𝘢𝘵 𝘤𝘢𝘯 𝘧𝘪𝘯𝘥 𝘸𝘩𝘢𝘵 𝘺𝘰𝘶𝘳 𝘤𝘰𝘮𝘱𝘢𝘯𝘺 𝘢𝘭𝘳𝘦𝘢𝘥𝘺 𝘬𝘯𝘰𝘸𝘴. Links in reply 👇