Gul Saeed Khattak

heyclicky (Mac and Windows): heyclicky.com Farza's demo on YouTube: youtube.com/watch?v=YlW_-hdw…

Sources 👇 The podcast: youtube.com/watch?v=iMsUcCYN… Musk and the backlash: cybernews.com/tech/altman-mu… The rename: freepressjournal.in/tech/sam…

Sam Altman changed his name to Ram Altman (for a short time). The real story behind it At 15:44 today I was following Ram Altman. By 16:08 he was Sam Altman. I opened his profile and honestly thought my eyes were going. Same @sama handle, same photo with the ski goggles. But the name said Ram. I took a screenshot just in case. When I checked again at 16:08, he was back to Sam Altman. Both screenshots are attached. So I went digging. It all goes back to a podcast. It started with one number: 5% On 29 October 2026, Sam sat down with John Collison on Cheeky Pint, the Stripe podcast. Memory is in short supply right now. He said it's "quite terrible" that everyone keeps RAM in their own computer when it's used maybe 5% of the time. In his view, it does more good kept busy in data centres. The internet did not take it well RAM prices are already going up because of AI. So the OpenAI boss saying your RAM belongs in data centres went down badly with PC gamers and anyone running models at home. By 1 November 2026, Elon Musk had posted a meme saying "The people deserve to have their RAM too!" It got 4.6 million views. Someone replied with an "Uncle Ram" army recruitment poster. Sam plays along This week he hit back on X. He said he'd been taken out of context, and that he actually wants more memory made so people can buy as much as they like. He also admitted the criticism was getting annoying. Then a follower suggested he call himself Ram Altman. He did. In India it took on a life of its own. People linked "Ram" to the Ramayana film coming out on 5 November 2026 and filled his replies with Diwali wishes. Today it was gone again. What I think I like that he can laugh at himself. But the name was a joke, and the view behind it hasn't moved. He still thinks your RAM is better off in the cloud. I work with data every day. The 5% is the only part I care about, because that's when the heavy query or the local model is running. Nobody buys RAM for the hours the laptop lid is closed. It's also worth remembering who's making the argument. The company selling cloud AI thinks memory belongs in the cloud. He might have the maths right. I'm keeping mine anyway. Links in the replies 👇👇👇👇

AI models don't really understand Southeast Asia. Here's a simple fix that doesn't break everything else. Most AI models that understand pictures were trained mostly on data from the US and Europe. So when you show them food from Vietnam, a temple in Cambodia or a festival in the Philippines, they often get it wrong. Sometimes they fall back on stereotypes. The obvious fix is to teach the AI more about that region. But there is a catch. When you teach it a lot about one region, it starts to forget the rest of the world. In this paper, the score on general tasks fell by about a third. A new paper from Cohere, SEACrowd, AI Singapore and 30+ universities and labs shows a simple way around this. The idea: make a local expert, then blend it back with the original. 1. Collect good local data: photos, questions and answers from Southeast Asia, in local languages. 2. Train a copy of the AI on it, so it becomes a local expert. 3. Mix the local expert back with the original AI. 4. Choose the mix that does well both locally and globally. You end up with one AI that understands local culture and still knows the world. A few things stood out: → A small mix was enough. Only about 10% of the local expert gave the best result. → It worked for three kinds of AI: one that answers questions about images, one that makes images, and one that matches images to text. → More data was not always better. Adding the wrong kind of local data made the AI worse. Results for Southeast Asia (11 countries, ~700 million people): • 5–15% better at understanding local culture • Over 98% of its general ability kept, sometimes even improved What I find interesting as someone working with AI: you don't have to choose between a local model and a global one. A light blend gives you most of both, and the same idea could work for any region or industry. Links in the reply 👇

No more login screens: Claude connectors via Single Sign-On Every time I add a new MCP connector to Claude, it's the same routine. Click connect, a new tab opens, log in, press allow, come back. Then do it all again for the next tool. For me that's a small annoyance. For a company it's a real problem. Think about a new starter on their first day. Claude is ready, but before it can see their Jira tickets or Notion pages they have to click through a login screen for every single tool. Ten tools means ten login screens before they can ask anything useful. Anthropic's fix is called enterprise-managed auth. It's been live on Claude Team and Enterprise since August 2026. IT links the company login (Okta) to Claude once and picks which tools each team gets. The new starter logs in and Slack, Notion, Jira and Figma are already connected. Nothing to click. It works the same in Claude Code and Cowork too. One customer said they set up 2,000 staff this way with zero extra steps. The part I like as a data person: when someone leaves, IT switches off their company account and their Claude access to every tool goes with it. No forgotten tokens sitting around. My doubt is that Okta is the only company login it supports right now. If your place runs on Microsoft, you wait. And one login now opens everything, so that account had better be locked down properly. Links in replies 👇

𝗜 𝗕𝘂𝗶𝗹𝘁 𝗡𝗲𝗿𝗱𝗟𝗼𝗿𝗲: Karpathy Built HackerTube. I Took It One Step Further. Last weekend Andrej Karpathy opened YouTube late on a Friday night and didn't like what the algorithm served him. So he built his own. In about 𝟮 𝗵𝗼𝘂𝗿𝘀 he made HackerTube: 𝟲,𝟰𝟱𝟰 𝘃𝗶𝗱𝗲𝗼𝘀 pulled from 𝟰𝟬𝗞 𝗛𝗮𝗰𝗸𝗲𝗿 𝗡𝗲𝘄𝘀 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝘀 that recommend a YouTube video, from 2007 to 2026, sorted into 𝟭𝟱𝟬 𝘁𝗼𝗽𝗶𝗰𝘀. He also published the data, with a note that it's there to "𝘨𝘪𝘷𝘦 𝘵𝘰 𝘺𝘰𝘶𝘳 𝘢𝘨𝘦𝘯𝘵 𝘵𝘰 𝘦𝘢𝘴𝘪𝘭𝘺 𝘤𝘳𝘦𝘢𝘵𝘦 𝘺𝘰𝘶𝘳 𝘰𝘸𝘯 𝘷𝘦𝘳𝘴𝘪𝘰𝘯." 𝗦𝗼 𝗜 𝗱𝗶𝗱. 𝗔𝗻𝗱 𝗜 𝘄𝗲𝗻𝘁 𝗼𝗻𝗲 𝘀𝘁𝗲𝗽 𝗳𝘂𝗿𝘁𝗵𝗲𝗿. HackerTube shows what engineers keep recommending. YouTube knows how many people pressed play. 𝗡𝗲𝗿𝗱𝗟𝗼𝗿𝗲 𝗽𝘂𝘁𝘀 𝘁𝗵𝗲 𝘁𝘄𝗼 𝘀𝗶𝗱𝗲 𝗯𝘆 𝘀𝗶𝗱𝗲, 𝘃𝗶𝗱𝗲𝗼 𝗯𝘆 𝘃𝗶𝗱𝗲𝗼, 𝘀𝗼 𝘆𝗼𝘂 𝗰𝗮𝗻 𝘀𝗲𝗲 𝘄𝗵𝗲𝗿𝗲 𝘁𝗵𝗲𝘆 𝗱𝗶𝘀𝗮𝗴𝗿𝗲𝗲. I've open-sourced it under MIT. 𝗪𝗵𝗮𝘁 𝗜𝘁 𝗦𝗵𝗼𝘄𝘀 • 𝗦𝗮𝗺𝗲 𝗽𝗿𝗮𝗶𝘀𝗲, 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗿𝗲𝗮𝗰𝗵: two videos with exactly the same number of HN recommenders. On 3 Nov: 16 each, 3Blue1Brown's neural network intro at 𝟮𝟱,𝟬𝟴𝟴,𝟮𝟲𝟱 𝘃𝗶𝗲𝘄𝘀, Freya Holmér's "The Beauty of Bézier Curves" at 𝟮,𝟮𝟯𝟭,𝟴𝟭𝟳. • 𝗣𝗼𝗽𝘂𝗹𝗮𝗿 𝘃𝘀 𝗿𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝗲𝗱: every video on one map, in four zones (hidden gems, shared classics, crowd hits, quiet corner). You set the lines. • 𝗛𝗶𝗱𝗱𝗲𝗻 𝗴𝗲𝗺𝘀: loved by HN, watched by few, each with a real HN quote. • 𝗔𝗹𝘀𝗼: the channels HN trusts most, the best picks by year, and HN's running video jokes. 𝗧𝗿𝘆 𝗜𝘁 𝗪𝗶𝘁𝗵𝗼𝘂𝘁 𝗮 𝗞𝗲𝘆 There's a 𝗱𝗲𝗺𝗼 𝗺𝗼𝗱𝗲: one click, no API key, no download. It runs on a small built-in sample of 𝟲𝟮𝟴 𝘃𝗶𝗱𝗲𝗼𝘀 with view counts that are 𝗺𝗮𝗱𝗲 𝘂𝗽, and every page says so. For real numbers you bring your own free YouTube API key. A run costs about 𝟭 𝗾𝘂𝗼𝘁𝗮 𝘂𝗻𝗶𝘁 𝗽𝗲𝗿 𝟱𝟬 𝘃𝗶𝗱𝗲𝗼𝘀, and the free daily quota is 𝟭𝟬,𝟬𝟬𝟬. 𝗙𝗼𝗿 𝗕𝘂𝗶𝗹𝗱𝗲𝗿𝘀 • 𝗧𝘄𝗼 𝗣𝘆𝘁𝗵𝗼𝗻 𝗱𝗲𝗽𝗲𝗻𝗱𝗲𝗻𝗰𝗶𝗲𝘀: Streamlit and pandas. That's the whole list. • The interactive pages are plain HTML, CSS and canvas. No chart libraries. • It runs on your own computer. No accounts, no tracking, and your key stays on your machine. • 𝟯𝟬 𝘁𝗲𝘀𝘁𝘀, run on Python 3.10 and 3.13. 𝗧𝗵𝗲 𝗛𝗼𝗻𝗲𝘀𝘁 𝗣𝗮𝗿𝘁 • View counts are a snapshot, shown exactly as YouTube returns them. The app never turns them into scores. • HackerTube's labels were written by an LLM, so expect 𝘀𝗼𝗺𝗲 𝗻𝗼𝗶𝘀𝗲. • A talk can look unwatched because another upload holds the views. There's a check for that, but it costs quota. • It's 𝗱𝗮𝘆𝘀 𝗼𝗹𝗱. Expect rough edges. 𝘔𝘢𝘺𝘣𝘦 𝘵𝘩𝘦 𝘣𝘦𝘴𝘵 𝘳𝘦𝘤𝘰𝘮𝘮𝘦𝘯𝘥𝘢𝘵𝘪𝘰𝘯 𝘦𝘯𝘨𝘪𝘯𝘦 𝘪𝘴 𝘢 𝘧𝘦𝘸 𝘵𝘩𝘰𝘶𝘴𝘢𝘯𝘥 𝘦𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘴 𝘵𝘦𝘭𝘭𝘪𝘯𝘨 𝘦𝘢𝘤𝘩 𝘰𝘵𝘩𝘦𝘳 𝘸𝘩𝘢𝘵 𝘵𝘰 𝘸𝘢𝘵𝘤𝘩. Links in reply 👇 NerdLore (open source, MIT): github.com/saeedkhattak/nerd…

Who is this reply even for? Saw a reply today under my NVIDIA auto-ontology repo post today . It just said "@bot save this". I assumed it was a joke, so I looked it up. @bot is xAI's Grok Bot, an agent app they launched in August 2026. You tag it under a post and the post goes to your own agent, which runs on its own computer in the cloud all day. So that reply wasn't really for the person who shared the repo. It was for someone's private assistant, left out in public where the rest of us could read it. Feels a bit like leaving a note for your cleaner on a stranger's front door. I get why people do it though. I have hundreds of bookmarks I've never opened. Every time I save something I tell myself I'll read it later, and I don't. At least an agent might actually go through the list. I read that repo properly because I'm building an ontology layer on top of a lineage graph at work. I doubt most people saving it will open the README. Their bots probably will. Still can't decide how I feel about that.

𝟯 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗧𝗼𝗼𝗹𝘀 𝗥𝗮𝗰𝗶𝗻𝗴 𝘁𝗼 𝗠𝗮𝗻𝗮𝗴𝗲 𝗬𝗼𝘂𝗿 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁𝘀 You have Claude Code in one terminal. Codex in another. Five more agents running somewhere. 𝗪𝗵𝗼'𝘀 𝗺𝗮𝗻𝗮𝗴𝗶𝗻𝗴 𝘁𝗵𝗲𝗺? Three open-source projects want that job. Same goal, very different ideas. 𝗣𝗮𝗽𝗲𝗿𝗰𝗹𝗶𝗽: 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 (𝟭𝟭𝟬𝗸 ⭐) "𝘐𝘧 𝘖𝘱𝘦𝘯𝘊𝘭𝘢𝘸 𝘪𝘴 𝘢𝘯 𝘦𝘮𝘱𝘭𝘰𝘺𝘦𝘦, 𝘗𝘢𝘱𝘦𝘳𝘤𝘭𝘪𝘱 𝘪𝘴 𝘵𝘩𝘦 𝘤𝘰𝘮𝘱𝘢𝘯𝘺." • Hire agents into roles: CEO, engineer, marketer • Set goals and budgets, approve plans • Not just code: research, content, operations too 𝗢𝗽𝗲𝗻𝗛𝗮𝗻𝗱𝘀: 𝘁𝗵𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝘁𝗲𝗮𝗺 (𝟵𝟮𝗸 ⭐) Now leads with Agent Canvas, a "𝘥𝘦𝘷𝘦𝘭𝘰𝘱𝘦𝘳 𝘤𝘰𝘯𝘵𝘳𝘰𝘭 𝘤𝘦𝘯𝘵𝘦𝘳 𝘧𝘰𝘳 𝘤𝘰𝘥𝘪𝘯𝘨 𝘢𝘨𝘦𝘯𝘵𝘴." • The only one of the three with 𝗶𝘁𝘀 𝗼𝘄𝗻 𝗰𝗼𝗱𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁 • Also runs Claude Code, Codex, Gemini and others • Automations with Slack, GitHub and Linear. Self-host or use their cloud 𝗥𝘂𝗳𝗹𝗼: 𝘁𝗵𝗲 𝗽𝗼𝘄𝗲𝗿 𝘁𝗼𝗼𝗹 (𝟴𝟭𝗸 ⭐) "𝘈𝘯 𝘢𝘨𝘦𝘯𝘵 𝘮𝘦𝘵𝘢-𝘩𝘢𝘳𝘯𝘦𝘴𝘴 𝘧𝘰𝘳 𝘊𝘭𝘢𝘶𝘥𝘦 𝘊𝘰𝘥𝘦 𝘢𝘯𝘥 𝘊𝘰𝘥𝘦𝘹." • Lives inside Claude Code as plugins • Runs "swarms" of agents, with routing and policy rules • Built for developers who stay in the terminal 𝗧𝗵𝗲 𝘀𝗶𝗺𝗽𝗹𝗲 𝘃𝗲𝗿𝘀𝗶𝗼𝗻: • Paperclip manages 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 • OpenHands runs 𝘁𝗵𝗲 𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴 𝘁𝗲𝗮𝗺 • Ruflo powers up 𝘁𝗵𝗲 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗧𝗵𝗲 𝘄𝗶𝗹𝗱 𝗽𝗮𝗿𝘁: Paperclip only launched in 𝗠𝗮𝗿𝗰𝗵 𝟮𝟬𝟮𝟲. It passed both in about 8 months. 𝗞𝗲𝗲𝗽 𝗶𝗻 𝗺𝗶𝗻𝗱: • 𝗦𝘁𝗮𝗿𝘀 𝗺𝗲𝗮𝘀𝘂𝗿𝗲 𝗮𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻, 𝗻𝗼𝘁 𝗿𝗲𝗮𝗹 𝘂𝘀𝗲 • They overlap but aren't the same product. Pick by how you work, not by stars The race isn't to build the best agent anymore. 𝗜𝘁'𝘀 𝘁𝗼 𝗯𝗲 𝘁𝗵𝗲 𝘀𝗰𝗿𝗲𝗲𝗻 𝘆𝗼𝘂 𝗺𝗮𝗻𝗮𝗴𝗲 𝘁𝗵𝗲𝗺 𝗳𝗿𝗼𝗺. 𝘚𝘵𝘢𝘳 𝘤𝘰𝘶𝘯𝘵𝘴 𝘢𝘴 𝘰𝘧 𝘕𝘰𝘷 3, 2026. Links in reply 👇

Just heard decomp.dev is winding down. It's a scoreboard for fan projects that rebuild old games' code. Its maintainer says AI-powered copying broke the community side of the work. But I think the real story is this: 𝗧𝗵𝗲 𝗰𝗼𝗺𝗺𝘂𝗻𝗶𝘁𝘆 𝗲𝘅𝗶𝘀𝘁𝗲𝗱 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸 𝘄𝗮𝘀 𝗵𝗮𝗿𝗱. When one game took years, you needed other people. Hard work forced people to work together. AI made the work easy. So people stopped needing each other. 𝘞𝘩𝘢𝘵 𝘰𝘵𝘩𝘦𝘳 𝘤𝘰𝘮𝘮𝘶𝘯𝘪𝘵𝘪𝘦𝘴 𝘰𝘯𝘭𝘺 𝘦𝘹𝘪𝘴𝘵 𝘣𝘦𝘤𝘢𝘶𝘴𝘦 𝘵𝘩𝘦 𝘸𝘰𝘳𝘬 𝘪𝘴 𝘩𝘢𝘳𝘥?

CRAZY: An AI Just Fixed Another AI by Editing it Weights. 𝗟𝗟𝗠𝘀 𝗔𝗿𝗲 𝗦𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝘁𝗼 𝗥𝗲𝗮𝗱 𝗪𝗲𝗶𝗴𝗵𝘁𝘀 𝗟𝗶𝗸𝗲 𝗖𝗼𝗱𝗲 Here is a question that would have sounded silly not long ago: if you showed an AI the raw numbers inside another AI, could it tell what that AI does? A new post from MIT researchers says: sometimes, yes. Think of a neural network as a giant recipe written in a language nobody speaks. Millions of numbers, no labels. Training writes the recipe. Humans can't really read it. The researchers gave OpenAI's GPT-6 Astra these numbers as text and asked it questions. 𝗪𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝗲𝗱 • 𝗜𝘁 𝗿𝗲𝗮𝗱: shown 4 game-playing networks, it correctly named each game. • 𝗜𝘁 𝗿𝗲𝗽𝗮𝗶𝗿𝗲𝗱: a stuck car-driving network worked after it changed 3.5% of the numbers. Success went from 𝟬 𝘁𝗼 𝟭,𝟬𝟮𝟰 𝗼𝘂𝘁 𝗼𝗳 𝟭,𝟬𝟮𝟰. • 𝗜𝘁 𝗰𝗿𝗲𝗮𝘁𝗲𝗱: with no practice, it wrote a Breakout player that scored 𝟱𝟭𝟳, far above the normally trained networks in the authors' tests (25.3). • 𝗜𝘁 𝗱𝗶𝗮𝗴𝗻𝗼𝘀𝗲𝗱: it found why a small language model kept getting stuck in loops and fixed it. Math scores went from 𝟳.𝟵% 𝘁𝗼 𝟰𝟴.𝟭%. 𝗧𝗵𝗲 𝗼𝘁𝗵𝗲𝗿 𝘀𝗶𝗱𝗲 • It couldn't tell which of two models was trained to flatter users, until it was allowed to see where each model "paid attention". • On a different small model with no looping problem, it couldn't improve math at all. 𝗜𝘁 𝗿𝗲𝗺𝗼𝘃𝗲𝗱 𝗮 𝗿𝗼𝗮𝗱𝗯𝗹𝗼𝗰𝗸. 𝗜𝘁 𝗱𝗶𝗱𝗻'𝘁 𝘁𝗲𝗮𝗰𝗵 𝗮 𝗻𝗲𝘄 𝘀𝗸𝗶𝗹𝗹. • These are small models, in a blog post, tested by the authors themselves. 𝗪𝗵𝗮𝘁 𝘄𝗲 𝗸𝗻𝗼𝘄 A general AI, used as it is, did things with model weights that would be very hard for a person. We don't know yet whether it works on models the size of the ones we use every day. 𝘐𝘧 𝘈𝘐 𝘤𝘢𝘯 𝘳𝘦𝘢𝘥 𝘢𝘯𝘥 𝘦𝘥𝘪𝘵 𝘰𝘵𝘩𝘦𝘳 𝘈𝘐𝘴, 𝘸𝘩𝘰 𝘤𝘩𝘦𝘤𝘬𝘴 𝘵𝘩𝘦 𝘦𝘥𝘪𝘵𝘴? Links in reply 👇