Gul Saeed Khattak

𝗔𝗹𝗹 𝗬𝗼𝘂 𝗡𝗲𝗲𝗱 𝘁𝗼 𝗞𝗻𝗼𝘄 𝗔𝗯𝗼𝘂𝘁 𝗖𝗹𝗮𝘂𝗱𝗲'𝘀 𝗡𝗲𝘄 𝗠𝗲𝘀𝘀𝗮𝗴𝗲 𝗧𝗵𝗿𝗲𝗮𝗱𝘀 Picture a coding agent two hours into a job. It has opened 40 files, run the tests a dozen times and written a long plan. Now it wants to take one tiny next step. Until now, to do that, it had to send Anthropic's API 𝘁𝗵𝗲 𝗲𝗻𝘁𝗶𝗿𝗲 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝗮𝗴𝗮𝗶𝗻. Every file, every test log, every earlier message. Then again on the next step. And the next. 𝗢𝗻 𝗡𝗼𝘃𝗲𝗺𝗯𝗲𝗿 𝟮, 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗮𝗱𝗱𝗲𝗱 "𝗺𝗲𝘀𝘀𝗮𝗴𝗲 𝘁𝗵𝗿𝗲𝗮𝗱𝘀", 𝗮𝗻𝗱 𝘁𝗵𝗮𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀. Here's the difference, with made-up sizes just to show the idea: ❌ 𝗕𝗲𝗳𝗼𝗿𝗲: 𝗻𝗼 𝘁𝗵𝗿𝗲𝗮𝗱 Turn 1: send the whole conversation (800 KB) Turn 2: send all of it again, plus your new message (802 KB) Turn 3: all of it again, plus the next message (804 KB) ✅ 𝗔𝗳𝘁𝗲𝗿: 𝘄𝗶𝘁𝗵 𝗮 𝘁𝗵𝗿𝗲𝗮𝗱 Turn 1: send the whole conversation once (800 KB) Turn 2: send only your new message and the ID of Claude's last reply (about 2 KB) Turn 3: only the next message and the latest ID (about 2 KB) Anthropic's servers stitch the full conversation back together before Claude reads it. So Claude still sees everything. Your app just stops carrying it back and forth. Those sizes aren't far from real life. Anthropic says the typical request in long Claude Code sessions went from 𝗮𝗯𝗼𝘂𝘁 𝟴𝟬𝟬 𝗞𝗕 𝘁𝗼 𝟭.𝟴 𝗞𝗕. 𝗦𝗼 𝘄𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘁𝗼 𝘁𝗶𝗺𝗲 𝗮𝗻𝗱 𝗽𝗿𝗶𝗰𝗲? ⏱️ 𝗧𝗶𝗺𝗲 𝗴𝗼𝗲𝘀 𝗱𝗼𝘄𝗻. Without a thread, the API has to receive, check and tokenize the whole conversation again before Claude can start. With a thread, most of that work covers only the new part. On really big sessions (500K to 1M tokens), Anthropic says Claude started replying 𝗮𝗯𝗼𝘂𝘁 𝟯𝟱% 𝘀𝗼𝗼𝗻𝗲𝗿. 💰 𝗣𝗿𝗶𝗰𝗲 𝘀𝘁𝗮𝘆𝘀 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲. Claude still reads the full conversation on every turn, so you pay for the same tokens as before. Rate limits don't change either. If you want to cut cost, that's still the job of prompt caching. People get confused about who this affects, so let's split it in two. 𝗜𝗳 𝘆𝗼𝘂 𝗷𝘂𝘀𝘁 𝘂𝘀𝗲 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 You don't have to do anything. Anthropic says threads are already 𝗼𝗻 𝗯𝘆 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 𝗳𝗼𝗿 𝗺𝗼𝘀𝘁 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝘂𝘀𝗲𝗿𝘀. It all happens in the background. There's no new button, no setting and nothing on screen. You chat and code the way you always have. In long sessions, replies may just start a bit sooner. If you step away for a while, the thread expires 𝟱 𝗺𝗶𝗻𝘂𝘁𝗲𝘀 after the last request. The full history then gets sent again to start a new one. You won't see that happen. That one reply just won't get the speed boost. Anthropic has only talked about Claude Code here. It hasn't said anything about the regular Claude chat app. 𝗜𝗳 𝘆𝗼𝘂 𝗯𝘂𝗶𝗹𝗱 𝗼𝗻 𝘁𝗵𝗲 𝗔𝗣𝗜 Here it's the opposite: 𝗻𝗼𝘁𝗵𝗶𝗻𝗴 𝗰𝗵𝗮𝗻𝗴𝗲𝘀 𝘂𝗻𝗹𝗲𝘀𝘀 𝘆𝗼𝘂 𝘁𝘂𝗿𝗻 𝗶𝘁 𝗼𝗻. The easiest way is the thread helper in Anthropic's SDK. In Python it looks like this: client = anthropic.Anthropic() thread = client.beta.messages.thread( model="claude-opus-5-5", max_tokens=1024, system="You are a coding assistant.") thread.send({"role": "user", "content": "Fix the failing test."}) thread.send({"role": "user", "content": "Now run it again."}) The first send() sends the full conversation and starts the thread. Every send() after that sends only the new message. The helper adds the beta header for you, and if a thread has expired, it sends the full history again on its own. A few things worth knowing before you switch it on: If your chats are only a few thousand tokens long, you probably won't notice much. This is built for long-running agents. 𝗢𝘁𝗵𝗲𝗿𝘀 𝗴𝗼𝘁 𝗵𝗲𝗿𝗲 𝗳𝗶𝗿𝘀𝘁, 𝗶𝗻 𝗮 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝘄𝗮𝘆 OpenAI's Responses API and Google's Interactions API can already continue a conversation from an ID. 𝗙𝗼𝗿 𝗹𝗼𝗻𝗴-𝗿𝘂𝗻𝗻𝗶𝗻𝗴 𝗮𝗴𝗲𝗻𝘁𝘀, 𝗷𝘂𝘀𝘁 𝗺𝗼𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝗰𝗼𝗻𝘃𝗲𝗿𝘀𝗮𝘁𝗶𝗼𝗻 𝗮𝗿𝗼𝘂𝗻𝗱 𝗵𝗮𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝗽𝗮𝗿𝘁 𝗼𝗳 𝘁𝗵𝗲 𝘄𝗮𝗶𝘁. claude.com/resources/article…

Saw this and wondered why Elon still works like this when he doesn't need to. I got the answer pretty quickly. Tbh a yacht is fun for a week. After that it's the same sea every morning. Some might disagree , but that is fine . Most of us think the point of money is to stop working. But rest only feels good when there's work around it. Take the work away and relaxing becomes its own kind of boredom. What doesn't get boring is building something nobody has built before. You never really finish it, so it keeps you hooked for years. If Elon had cashed out and moved onto a boat, he wouldn't be Elon, and none of us would be talking about his 3am office nights. People with lots of money and nothing to build often drift into bad habits. People deep in building something rarely have time for them. On a much smaller scale, my best late nights are the ones spent building something. It feels more like rest than any day off.

𝗡𝘃𝗶𝗱𝗶𝗮'𝘀 𝟱𝟭% 𝗥𝘂𝗹𝗲: 𝗪𝗵𝗮𝘁 𝗜𝘁 𝗚𝗶𝘃𝗲𝘀 𝗬𝗼𝘂 𝗔𝗻𝗱 𝗪𝗵𝗮𝘁 𝗜𝘁 𝗗𝗼𝗲𝘀𝗻'𝘁 Picture a data centre in Singapore. Racks full of Nvidia chips. Its biggest clients are Chinese AI companies that can't buy those chips at home, so they rent them. 𝗧𝗵𝗲 𝗨𝗻𝗱𝗲𝗿𝗹𝘆𝗶𝗻𝗴 𝗜𝗱𝗲𝗮 Compute is a physical thing, but access to it isn't. You can stop a box at a border. It's much harder to stop someone logging in to a box in another country. So there are two doors: • 𝗕𝘂𝘆𝗶𝗻𝗴: export controls, plus Nvidia's "white list", which per the FT more than halved approved Asian buyers this year. • 𝗥𝗲𝗻𝘁𝗶𝗻𝗴: Chinese companies leasing GPUs in overseas data centres, including for model training. Close the first door and traffic moves to the second. 𝗧𝗵𝗶𝘀 𝗿𝘂𝗹𝗲 𝗶𝘀 𝗮 𝘃𝗮𝗹𝘃𝗲 𝗼𝗻 𝘁𝗵𝗲 𝘀𝗲𝗰𝗼𝗻𝗱 𝗱𝗼𝗼𝗿. 𝗔𝗻 𝗔𝗻𝗮𝗹𝗼𝗴𝘆 It's a bit like a landlord telling a building manager: no more than 49% of tenants from one place. Not a law. A lease condition, enforced by whoever signs the lease. 𝗡𝘃𝗶𝗱𝗶𝗮 𝗻𝗼𝘄 𝘄𝗮𝗻𝘁𝘀 𝘁𝗵𝗮𝘁 𝘁𝗼 𝗰𝗵𝗮𝗻𝗴𝗲. 𝗪𝗵𝗮𝘁 𝗛𝗮𝗽𝗽𝗲𝗻𝗲𝗱 According to the Financial Times (five people familiar with the matter), Nvidia has told cloud providers that 𝗮𝘁 𝗹𝗲𝗮𝘀𝘁 𝟱𝟭% of their services must go to non-Chinese companies. There was no such limit before. • 𝗡𝗼 𝗽𝗮𝗽𝗲𝗿𝘄𝗼𝗿𝗸: the instruction was given verbally, through Nvidia's sales and regional teams. • 𝗪𝗼𝗿𝗹𝗱𝘄𝗶𝗱𝗲: it applies to clients globally, not only in Asia (two people told the FT). • 𝗡𝗲𝘄 𝗱𝗲𝗮𝗹𝘀 𝗼𝗻𝗹𝘆: existing leases are not affected. • 𝗦𝗶𝗻𝗰𝗲 𝘁𝗵𝗲 𝘀𝘂𝗺𝗺𝗲𝗿: two people said Nvidia did it pre-emptively, to cut criticism from US lawmakers. 𝗪𝗵𝘆 𝗥𝗲𝗻𝘁𝗶𝗻𝗴 𝗠𝗮𝘁𝘁𝗲𝗿𝘀 US export rules mostly control where chips are shipped. Renting them in a data centre abroad has been the gap. Many Chinese groups used overseas leases for heavy AI work, including model training. Nvidia had already tightened the front door. This year it built a "white list" of approved Asian buyers and 𝗺𝗼𝗿𝗲 𝘁𝗵𝗮𝗻 𝗵𝗮𝗹𝘃𝗲𝗱 it, per the FT. That pushed more Chinese firms toward renting. This new rule goes after 𝘁𝗵𝗲 𝗿𝗲𝗻𝘁𝗶𝗻𝗴. 𝗪𝗵𝗮𝘁 𝗜𝘁 𝗗𝗼𝗲𝘀 𝗧𝗼 𝗣𝗿𝗶𝗰𝗲𝘀 A rack of B300 chips sells for 𝗮𝗯𝗼𝘂𝘁 $𝟲𝟬𝟬,𝟬𝟬𝟬 internationally. One person told the FT a rack fetched 𝗥𝗺𝗯𝟭𝟴.𝟱𝗺𝗻 (𝗮𝗯𝗼𝘂𝘁 $𝟮.𝟴𝗺𝗻) in China last month. That is 𝗿𝗼𝘂𝗴𝗵𝗹𝘆 𝟰.𝟳 𝘁𝗶𝗺𝗲𝘀 the price. The B300 can't legally be sold in China. 𝗙𝗼𝗿 𝗖𝗹𝗼𝘂𝗱 𝗣𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀 Chinese tech groups are the biggest clients for data centres in places like south-east Asia. A Singapore-based provider told the FT it is hard to find enough non-Chinese clients outside the US. Providers that can't find them may have to 𝘀𝗵𝗿𝗶𝗻𝗸. New contracts only. 𝗧𝗵𝗲 𝗠𝗮𝗽 𝗢𝗳 𝗖𝗵𝗶𝗻𝗮 𝗖𝗼𝗺𝗽𝘂𝘁𝗲 𝗖𝗼𝗻𝘁𝗿𝗼𝗹𝘀, 𝗢𝗻𝗲 𝗟𝗶𝗻𝗲 𝗘𝗮𝗰𝗵 • 𝗨𝗦 𝗲𝘅𝗽𝗼𝗿𝘁 𝗰𝗼𝗻𝘁𝗿𝗼𝗹𝘀: block shipping top chips like the B300 to China. • 𝗕𝗜𝗦 𝗿𝘂𝗹𝗲 (𝗝𝗮𝗻𝘂𝗮𝗿𝘆): license applicants must list remote cloud users in China and other listed countries. • 𝗕𝗜𝗦 𝗴𝘂𝗶𝗱𝗮𝗻𝗰𝗲 (𝗠𝗮𝘆 𝟯𝟭): a license is needed for advanced chips going to companies headquartered in China or Macau, wherever they are. • 𝗡𝘃𝗶𝗱𝗶𝗮'𝘀 𝘄𝗵𝗶𝘁𝗲 𝗹𝗶𝘀𝘁 (𝘁𝗵𝗶𝘀 𝘆𝗲𝗮𝗿): approved Asian buyers more than halved, per the FT. • 𝗥𝗲𝗺𝗼𝘁𝗲 𝗔𝗰𝗰𝗲𝘀𝘀 𝗦𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗔𝗰𝘁: passed the House 369-22 in January; sitting with the Senate Banking Committee. • 𝗡𝘃𝗶𝗱𝗶𝗮'𝘀 𝟱𝟭% 𝗿𝘂𝗹𝗲 (𝗻𝗼𝘄): aimed at the cloud provider's customer mix, not the chip. 𝗪𝗵𝗲𝗿𝗲 𝗧𝗵𝗶𝘀 𝗙𝗶𝘁𝘀 Government rules set the floor. Nvidia is adding its own layer on top, so lawmakers have less to criticise (two people told the FT that was the motive). 𝗪𝗵𝗼 𝗙𝗲𝗲𝗹𝘀 𝗜𝘁 • 𝗖𝗹𝗼𝘂𝗱 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀 𝗶𝗻 𝘀𝗼𝘂𝘁𝗵-𝗲𝗮𝘀𝘁 𝗔𝘀𝗶𝗮: Chinese tech groups are their biggest clients. • 𝗖𝗵𝗶𝗻𝗲𝘀𝗲 𝗔𝗜 𝗰𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀: already paying a premium. A B300 rack, 𝗮𝗯𝗼𝘂𝘁 $𝟲𝟬𝟬𝗸 abroad, reportedly fetched $𝟮.𝟴𝗺𝗻 in China last month (one deal). 𝗧𝗵𝗲 𝗯𝗶𝗴𝗴𝗲𝗿 𝗽𝗼𝗶𝗻𝘁: 𝘵𝘩𝘦 𝘤𝘩𝘪𝘱 𝘸𝘢𝘳 𝘪𝘴 𝘮𝘰𝘷𝘪𝘯𝘨 𝘧𝘳𝘰𝘮 𝘸𝘩𝘰 𝘤𝘢𝘯 𝘣𝘶𝘺 𝘤𝘩𝘪𝘱𝘴 𝘵𝘰 𝘸𝘩𝘰 𝘤𝘢𝘯 𝘳𝘦𝘯𝘵 𝘵𝘩𝘦𝘮. FT (original report): ft.com/content/1902df70-1380… Links in reply 👇

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.

Claude Code says no to game clones. A tool that uses Claude Code tops trending anyway. What changed: Last week Claude Code got two new sentences telling it not to recreate specific existing games. What it is: This week a tool called Craft Anything is topping the trending charts. It reverse engineers apps and games, then has a coding agent build your own version. 1.6K stars today. 5.7K total. To be fair, it defaults to a clean-room mode and says it never bypasses DRM, logins or anti-cheat. The funny part: The builder it supports first is Claude Code. One of its own example requests is "Recreate this game's gameplay in Rust". My take: Big labs can block this kind of use inside their own agents. I don't think it changes much. Craft Anything already has adapters for open agents like Qwen Code and opencode. Open models are good enough now, and anyone can point these tools at them. Reverse engineering always cut both ways. The debugger that fixes your crash can also copy someone's code. What changed is the effort. It used to take a skilled person weeks. Now it's closer to a prompt. Links in reply 👇 Reply comment (post this straight after): Repo: github.com/storytold/craft-a…

OpenAI's $172K bug bounty, and why the partner matters more My read on OpenAI's security news is a bit different from the headlines. The money: • $172,000 is the top payout, with bonuses stacked on
• A critical bug in ChatGPT, Codex or the API now pays up to $30,000
• OpenAI says that's more than four times what critical findings paid before The partner: The more interesting part is who they picked. Cathedral is the startup whose AI agents reached Anthropic's internal business systems in October 2026. Not through Claude. Through third-party business software. Anthropic disputes parts of it. It says it took several days, not 12 hours, and that Cathedral went beyond the bounty rules by targeting vendors and not disclosing everything. Now OpenAI brings them in as a partner, not a bounty hunter. I think that's smart. A bounty comes with a fixed scope and safe harbour. A partner can agree up front what's in bounds. The Anthropic fight was mostly about scope and disclosure. What I'd watch: The path that worked at Anthropic was vendor software, not the model. Bigger bounties for ChatGPT bugs don't cover that. OpenAI hasn't said what Cathedral will test. Maybe vendors are already in scope. If not, the money is pointed at the wrong door. One line from their write-up on the last sprint stuck with me. A merged patch didn't always mean the fix had reached every deployed system. I've seen the same in data pipelines. Code merged, old job still running somewhere. Links in reply 👇 OpenAI's sprint post: openai.com/index/preparing-o…

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 👇

𝗥𝘂𝗻𝗻𝗶𝗻𝗴 𝗮𝗻 𝗔𝗹𝘄𝗮𝘆𝘀-𝗢𝗻 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗬𝗼𝘂𝗿𝘀𝗲𝗹𝗳: 𝗟𝗼𝗰𝗮𝗹 𝗼𝗿 𝗖𝗹𝗼𝘂𝗱? 𝗜 𝘄𝗮𝗻𝘁𝗲𝗱 𝘁𝗼 𝗿𝘂𝗻 𝗼𝗻𝗲 𝗺𝘆𝘀𝗲𝗹𝗳, 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗯𝗶𝗴 𝗺𝗼𝗻𝘁𝗵𝗹𝘆 𝗯𝗶𝗹𝗹𝘀. The big labs all sell the same idea: 𝗮𝗻 𝗔𝗜 𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗮𝘁 𝘄𝗼𝗿𝗸𝘀 𝗳𝗼𝗿 𝘆𝗼𝘂 𝟮𝟰/𝟳, often on its own computer in the cloud. 𝗖𝗹𝗼𝘀𝗲𝗱 𝘀𝗼𝘂𝗿𝗰𝗲 • OpenAI: Dots • Meta: Muse • xAI: Grok Bot • Anthropic: Claude Duet 𝗢𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲 • OpenMuse (CopilotKit): modelled on Meta's Muse • OpenDots (CopilotKit): modelled on OpenAI's Dots So where do I run it? 𝗠𝘆 𝗳𝗮𝘀𝘁 𝗯𝗿𝗮𝗶𝗻: 𝘭𝘰𝘤𝘢𝘭𝘭𝘺, 𝘰𝘧 𝘤𝘰𝘶𝘳𝘴𝘦. 𝘛𝘩𝘢𝘵'𝘴 𝘵𝘩𝘦 𝘸𝘩𝘰𝘭𝘦 𝘱𝘰𝘪𝘯𝘵. 𝗠𝘆 𝘀𝗹𝗼𝘄 𝗯𝗿𝗮𝗶𝗻: hang on. A research agent visits hundreds of websites. Any of them can hide text written for the AI, not for me. That's prompt injection. It doesn't infect my laptop like a virus. 𝗜𝘁 𝘁𝗮𝗹𝗸𝘀 𝗺𝘆 𝗮𝗴𝗲𝗻𝘁 𝗶𝗻𝘁𝗼 𝘂𝘀𝗶𝗻𝗴 𝗶𝘁𝘀 𝗼𝘄𝗻 𝘁𝗼𝗼𝗹𝘀 𝗮𝗴𝗮𝗶𝗻𝘀𝘁 𝗺𝗲. If the agent has a terminal, one hidden line can say "download and run this." So I looked at three options. 𝟭. 𝗠𝘆 𝗺𝗮𝗶𝗻 𝗹𝗮𝗽𝘁𝗼𝗽 Only inside a container or VM, with one shared folder. Free, but the weakest wall. 𝟮. 𝗔 𝗰𝗵𝗲𝗮𝗽 𝘀𝗲𝗰𝗼𝗻𝗱 𝗺𝗮𝗰𝗵𝗶𝗻𝗲 𝗮𝘁 𝗵𝗼𝗺𝗲 Better. But its security is now my job. And it sits on my home network, next to my router and other devices, unless I put it on a guest network. 𝟯. 𝗔 𝗰𝗵𝗲𝗮𝗽 𝗰𝗹𝗼𝘂𝗱 𝗩𝗠 (𝗛𝗲𝘁𝘇𝗻𝗲𝗿, 𝗔𝘇𝘂𝗿𝗲, 𝗲𝘁𝗰.) Off my home network. Snapshots mean I can wipe it and rebuild in minutes. But I pay for the VM plus the model, and I must not leave its dashboard open to the internet. Then it clicked. 𝗟𝗼𝗰𝗮𝗹 𝘃𝘀 𝗰𝗹𝗼𝘂𝗱 𝗶𝘀 𝘁𝗵𝗲 𝘄𝗿𝗼𝗻𝗴 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻. The real question is: 𝘄𝗵𝗮𝘁 𝗰𝗮𝗻 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝗿𝗲𝗮𝗰𝗵 𝘄𝗵𝗲𝗻 𝗶𝘁 𝗴𝗲𝘁𝘀 𝘁𝗿𝗶𝗰𝗸𝗲𝗱? A cloud VM holding my Gmail login is as dangerous as my laptop. A laptop agent with one empty folder is fairly safe. Simon Willison calls the danger the "𝘭𝘦𝘵𝘩𝘢𝘭 𝘵𝘳𝘪𝘧𝘦𝘤𝘵𝘢": private data, untrusted content, and a way to send data out. Remove any one of the three and the data-theft attack stops working. A research agent needs the untrusted content. So I remove the other two: • 𝗡𝗼 𝗽𝗲𝗿𝘀𝗼𝗻𝗮𝗹 𝗹𝗼𝗴𝗶𝗻𝘀 on the agent's machine • 𝗦𝗲𝗽𝗮𝗿𝗮𝘁𝗲 𝗮𝗰𝗰𝗼𝘂𝗻𝘁𝘀 for anything it needs • 𝗦𝗽𝗲𝗻𝗱𝗶𝗻𝗴 𝗰𝗮𝗽𝘀 on every API key (this also fixes the bill fear) • 𝗔𝘀𝗸 𝗯𝗲𝗳𝗼𝗿𝗲 𝗮𝗰𝘁𝗶𝗻𝗴: send, buy, post, delete, run • 𝗦𝗽𝗹𝗶𝘁 𝘁𝗵𝗲 𝗷𝗼𝗯: the browsing agent only writes summaries, and the agent with real tools never sees raw web pages Notice what the big labs did. Muse, Dots and Grok Bot all give the agent a separate cloud computer. Muse keeps credentials where the agent can't read them. Dots stay read-only while idle. 𝗥𝘂𝗻 𝗶𝘁 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳, 𝗮𝗻𝗱 𝘁𝗵𝗼𝘀𝗲 𝗰𝗵𝗼𝗶𝗰𝗲𝘀 𝗯𝗲𝗰𝗼𝗺𝗲 𝘆𝗼𝘂𝗿𝘀. 𝘍𝘢𝘴𝘵 𝘣𝘳𝘢𝘪𝘯 𝘢𝘴𝘬𝘴 𝘸𝘩𝘦𝘳𝘦 𝘪𝘵 𝘴𝘩𝘰𝘶𝘭𝘥 𝘳𝘶𝘯. 𝘚𝘭𝘰𝘸 𝘣𝘳𝘢𝘪𝘯 𝘢𝘴𝘬𝘴 𝘸𝘩𝘢𝘵 𝘪𝘵 𝘸𝘪𝘭𝘭 𝘩𝘰𝘭𝘥 𝘸𝘩𝘦𝘯 𝘪𝘵 𝘨𝘦𝘵𝘴 𝘵𝘳𝘪𝘤𝘬𝘦𝘥. Links in reply 👇

𝗟𝗲𝘁 𝗖𝗹𝗮𝘂𝗱𝗲 𝗧𝘂𝗿𝗻 𝗔𝗻𝘆 𝗛𝗮𝗿𝗱 𝗧𝗼𝗽𝗶𝗰 𝗜𝗻𝘁𝗼 𝗮 𝗖𝗹𝗶𝗰𝗸𝗮𝗯𝗹𝗲 𝗠𝗮𝗽, 𝗪𝗶𝘁𝗵 𝘁𝗵𝗲 𝗣𝗿𝗼𝗺𝗽𝘁𝘀 𝗕𝘂𝗶𝗹𝘁 𝗜𝗻 I have seen a better way in Claude recently. 𝗖𝗵𝗮𝘁 𝗶𝘀 𝗮 𝘀𝘁𝗿𝗮𝗶𝗴𝗵𝘁 𝗹𝗶𝗻𝗲. 𝗬𝗼𝘂 𝗼𝗻𝗹𝘆 𝗹𝗲𝗮𝗿𝗻 𝘄𝗵𝗮𝘁 𝘆𝗼𝘂 𝘁𝗵𝗶𝗻𝗸 𝘁𝗼 𝗮𝘀𝗸, 𝗮𝗻𝗱 𝘄𝗵𝗲𝗻 𝗮 𝘁𝗼𝗽𝗶𝗰 𝗶𝘀 𝗻𝗲𝘄 𝘁𝗼 𝘆𝗼𝘂, 𝘆𝗼𝘂 𝗱𝗼𝗻'𝘁 𝗸𝗻𝗼𝘄 𝘄𝗵𝗮𝘁 𝘁𝗼 𝗮𝘀𝗸. Ask it for a map of a topic, and it draws a flowchart. The topic is split into steps in a logical order: what comes first, what comes next. 𝗔𝗻𝗱 𝗲𝘃𝗲𝗿𝘆 𝗯𝗼𝘅 𝗶𝘀 𝗰𝗹𝗶𝗰𝗸𝗮𝗯𝗹𝗲. Tap one, and Claude writes the next question for you. Usually it sends straight away. Sometimes it lands in your typing box and you just press send. Claude answers that part only, often with a new flowchart. Then you tap again. 𝗟𝗲𝘁'𝘀 𝗧𝗿𝘆 𝗜𝘁: 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗲𝗿𝘀 Say you want to learn how a transformer works. That's the design from the 2017 paper "𝘈𝘵𝘵𝘦𝘯𝘵𝘪𝘰𝘯 𝘐𝘴 𝘈𝘭𝘭 𝘠𝘰𝘶 𝘕𝘦𝘦𝘥". You ask for a map. You get boxes something like this: • 𝗧𝗼𝗸𝗲𝗻𝘀: your text is cut into small pieces • 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀: each piece becomes a list of numbers, plus its position • 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻: each piece looks at the others • 𝗙𝗲𝗲𝗱-𝗳𝗼𝗿𝘄𝗮𝗿𝗱: each piece is processed on its own • 𝗟𝗮𝘆𝗲𝗿𝘀: attention and feed-forward repeat many times • 𝗡𝗲𝘅𝘁 𝘁𝗼𝗸𝗲𝗻: the model scores every possible next piece Attention sounds like the magic part, so you tap it. You didn't type anything. The question about attention is written for you, and Claude draws a new chart just for that part: • 𝗤𝘂𝗲𝗿𝘆, 𝗸𝗲𝘆, 𝘃𝗮𝗹𝘂𝗲: three versions of each piece • 𝗦𝗰𝗼𝗿𝗲𝘀: compare queries with keys • 𝗦𝗼𝗳𝘁𝗺𝗮𝘅: turn scores into weights • 𝗠𝗶𝘅: a weighted mix of the values • 𝗠𝗮𝗻𝘆 𝗵𝗲𝗮𝗱𝘀: several attentions running side by side Tap "Query, key, value", and it goes one level deeper. 𝗔𝗻𝗱 𝗼𝗻 𝗶𝘁 𝗴𝗼𝗲𝘀. 𝗪𝗵𝘆 𝗜 𝗟𝗶𝗸𝗲 𝗧𝗵𝗶𝘀 • 𝗬𝗼𝘂 𝘀𝗲𝗲 𝘁𝗵𝗲 𝗺𝗮𝗽 𝗯𝗲𝗳𝗼𝗿𝗲 𝘁𝗵𝗲 𝗱𝗲𝘁𝗮𝗶𝗹𝘀. • 𝗘𝗮𝗰𝗵 𝗯𝗼𝘅 𝗶𝘀 𝗮 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝘆𝗼𝘂 𝗱𝗶𝗱𝗻'𝘁 𝗵𝗮𝘃𝗲 𝘁𝗼 𝘄𝗿𝗶𝘁𝗲. • 𝗜𝗳 𝗘𝗻𝗴𝗹𝗶𝘀𝗵 𝗶𝘀𝗻'𝘁 𝘆𝗼𝘂𝗿 𝗳𝗶𝗿𝘀𝘁 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲: it isn't mine. Phrasing the follow-up takes real effort. A tap skips that. 𝗧𝗵𝗲 𝗛𝗼𝗻𝗲𝘀𝘁 𝗣𝗮𝗿𝘁 • 𝗜𝘁'𝘀 𝗻𝗲𝘄: Claude's in-chat visuals launched in beta in March 2026. Other apps may draw a static image or just text. • 𝗧𝗮𝗽𝘀 𝗮𝗿𝗲𝗻'𝘁 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁: sometimes the question sends straight away. Sometimes it lands in your typing box and you press send. • 𝗜𝘁 𝘄𝗼𝗿𝗸𝘀 𝗯𝗲𝘀𝘁 𝗶𝗻 𝗖𝗹𝗮𝘂𝗱𝗲'𝘀 𝗰𝗵𝗮𝘁 𝗮𝗽𝗽. 𝗜𝗻 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲, 𝗶𝘁 𝗶𝘀𝗻'𝘁 𝗮𝘀 𝘀𝗺𝗼𝗼𝘁𝗵. • 𝗬𝗼𝘂𝗿 𝗯𝗼𝘅𝗲𝘀 𝘄𝗶𝗹𝗹 𝗱𝗶𝗳𝗳𝗲𝗿: the chart above is an example, not a fixed output. • 𝗧𝗵𝗲 𝗿𝗮𝗯𝗯𝗶𝘁 𝗵𝗼𝗹𝗲: four charts later, you can lose where you started. Ask for the path at the top of every chart. • 𝗧𝗼𝗼 𝗻𝗲𝗮𝘁: a flowchart can make a messy topic look simpler than it is. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀: 𝗖𝗹𝗮𝘂𝗱𝗲 𝗰𝗵𝗮𝘁. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗱𝗼𝗲𝘀𝗻'𝘁: 𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲. 𝗧𝗵𝗲𝗿𝗲, 𝗶𝘁 𝗱𝗼𝗲𝘀𝗻'𝘁 𝘄𝗼𝗿𝗸 𝗮𝘀 𝗲𝘅𝗽𝗲𝗰𝘁𝗲𝗱 𝗮𝘁 𝗮𝗹𝗹. 𝗧𝗿𝘆 𝗜𝘁 𝗬𝗼𝘂𝗿𝘀𝗲𝗹𝗳 (𝗦𝘄𝗮𝗽 𝘁𝗵𝗲 𝗕𝗿𝗮𝗰𝗸𝗲𝘁𝘀) Give me a clickable map of [topic] as a flowchart. Use 5 to 7 boxes, in the order I should learn them. Make every box send its follow-up message as soon as I tap it, so I never type anything or press send. That message should ask about that step only and draw a new flowchart for it the same way. Show the path at the top of every chart (like: Transformer > Attention > Query, key, value) so I don't get lost. 𝘚𝘵𝘰𝘱 𝘨𝘶𝘦𝘴𝘴𝘪𝘯𝘨 𝘵𝘩𝘦 𝘯𝘦𝘹𝘵 𝘲𝘶𝘦𝘴𝘵𝘪𝘰𝘯. 𝘈𝘴𝘬 𝘧𝘰𝘳 𝘵𝘩𝘦 𝘮𝘢𝘱. Links in reply 👇

Paperclip: github.com/paperclipai/paper… OpenHands: github.com/OpenHands/OpenHan… Ruflo: github.com/ruvnet/ruflo

SUING THE NAMELESS: The Bot You Can’t Block Without Blocking Google. I saw the news that The New York Times and three other papers sued Brave. Curiosity kicked in, so I dug into the details. One detail turned out more interesting than the lawsuit itself: 𝗔 𝘄𝗲𝗯𝘀𝗶𝘁𝗲 𝗰𝗮𝗻'𝘁 𝘀𝗮𝘆 "𝗻𝗼" 𝘁𝗼 𝗕𝗿𝗮𝘃𝗲 𝗮𝗹𝗼𝗻𝗲. 𝗦𝗼𝗺𝗲 𝗯𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱 Every website can have a small file called robots.txt. It's like a sign on the front door telling visiting bots where they may go. Each bot introduces itself with a name, called a user agent. The site writes rules for each name. 𝗘𝘅𝗮𝗺𝗽𝗹𝗲 𝟭: 𝗯𝗹𝗼𝗰𝗸𝗶𝗻𝗴 𝗼𝗻𝗲 𝗔𝗜 𝗯𝗼𝘁 Say a news site doesn't want OpenAI training on its articles. It adds: 𝚄𝚜𝚎𝚛-𝚊𝚐𝚎𝚗𝚝: 𝙶𝙿𝚃𝙱𝚘𝚝 𝙳𝚒𝚜𝚊𝚕𝚕𝚘𝚠: / GPTBot stays out. Google can still visit, so the site still shows up in Google Search. 𝗘𝘅𝗮𝗺𝗽𝗹𝗲 𝟮: 𝘁𝗿𝘆𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘄𝗶𝘁𝗵 𝗕𝗿𝗮𝘃𝗲 𝚄𝚜𝚎𝚛-𝚊𝚐𝚎𝚗𝚝: 𝙱𝚛𝚊𝚟𝚎𝙱𝚘𝚝 𝙳𝚒𝚜𝚊𝚕𝚕𝚘𝚠: / 𝗧𝗵𝗶𝘀 𝗱𝗼𝗲𝘀 𝗻𝗼𝘁𝗵𝗶𝗻𝗴. Brave's crawler has no name of its own. "BraveBot" is made up to show the point. 𝗘𝘅𝗮𝗺𝗽𝗹𝗲 𝟯: 𝘁𝗵𝗲 𝗼𝗻𝗹𝘆 𝗿𝗼𝗯𝗼𝘁𝘀.𝘁𝘅𝘁 𝗿𝘂𝗹𝗲 𝘁𝗵𝗮𝘁 𝘄𝗼𝗿𝗸𝘀 Brave's help page says its crawler won't go where Googlebot isn't allowed. So: 𝚄𝚜𝚎𝚛-𝚊𝚐𝚎𝚗𝚝: 𝙶𝚘𝚘𝚐𝚕𝚎𝚋𝚘𝚝 𝙳𝚒𝚜𝚊𝚕𝚕𝚘𝚠: / That keeps Brave out. 𝗜𝘁 𝗮𝗹𝘀𝗼 𝗸𝗲𝗲𝗽𝘀 𝗚𝗼𝗼𝗴𝗹𝗲 𝗼𝘂𝘁. The site drops out of Google Search. For a newspaper, that's not a real choice. 𝗧𝗵𝗲 𝗲𝘃𝗲𝗿𝘆𝗱𝗮𝘆 𝘃𝗲𝗿𝘀𝗶𝗼𝗻 A shop puts up a sign: "Mr. Smith, please don't come in." Most visitors wear name tags, so it works. One visitor wears no name tag. His rule: "I only go where the mail carrier is allowed." To keep him out, you'd have to lock out the mail carrier too. 𝗧𝗵𝗲𝗻 𝘆𝗼𝘂 𝘀𝘁𝗼𝗽 𝗴𝗲𝘁𝘁𝗶𝗻𝗴 𝗺𝗮𝗶𝗹. 𝗕𝗿𝗮𝘃𝗲'𝘀 𝘀𝗶𝗱𝗲 Brave explains this on the same help page: it skips a separate name because some websites let only Google crawl them. A new search engine with its own name would simply get shut out. In an earlier case against News Corp, Brave also argued that indexing the web is fair use, simply how search engines work. 𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀 𝗶𝗻 𝘁𝗵𝗲 𝗹𝗮𝘄𝘀𝘂𝗶𝘁 The newspapers say this left them no way to block Brave alone, while, they allege, Brave copied their pages and sold access to AI companies. 𝗪𝗵𝗮𝘁 𝘁𝗼 𝗸𝗲𝗲𝗽 𝗶𝗻 𝗺𝗶𝗻𝗱 • These are 𝗮𝗹𝗹𝗲𝗴𝗮𝘁𝗶𝗼𝗻𝘀. No court has ruled. • robots.txt is 𝗮 𝗿𝗲𝗾𝘂𝗲𝘀𝘁, 𝗻𝗼𝘁 𝗮 𝗹𝗼𝗰𝗸. Bots choose to follow it. • Brave says a page can be removed with a noindex tag. But a general noindex tag 𝗿𝗲𝗺𝗼𝘃𝗲𝘀 𝗶𝘁 𝗳𝗿𝗼𝗺 𝗚𝗼𝗼𝗴𝗹𝗲 𝘁𝗼𝗼. • OpenAI has more than one bot. GPTBot is for training. Its search bot has a different name. 𝗧𝗵𝗲 𝗔𝗜 𝗳𝗶𝗴𝗵𝘁 𝗶𝘀𝗻'𝘁 𝗼𝗻𝗹𝘆 𝗮𝗯𝗼𝘂𝘁 𝘄𝗵𝗮𝘁 𝗯𝗼𝘁𝘀 𝘁𝗮𝗸𝗲. 𝗜𝘁'𝘀 𝗮𝗹𝘀𝗼 𝗮𝗯𝗼𝘂𝘁 𝘄𝗵𝗲𝘁𝗵𝗲𝗿 𝘆𝗼𝘂 𝗰𝗮𝗻 𝘁𝗲𝗹𝗹 𝘄𝗵𝗼'𝘀 𝗸𝗻𝗼𝗰𝗸𝗶𝗻𝗴. theverge.com/tech/1024588/br…

𝗢𝗽𝗲𝗻 𝗦𝗰𝗶𝗲𝗻𝗰𝗲 𝗠𝗼𝗱𝗲𝗹𝘀 𝗔𝗿𝗲 𝗡𝗼𝘄 𝗮 𝗖𝗮𝘁𝗲𝗴𝗼𝗿𝘆 Hydrogen-1.0 dropped today. It's a sign that "open model for science" is now its own lane, not a side project. 𝗧𝗵𝗲 𝗺𝗮𝗽 𝘀𝗼 𝗳𝗮𝗿: • 𝗚𝗮𝗹𝗮𝗰𝘁𝗶𝗰𝗮 (Meta, 2022): trained on papers, non-commercial; its public demo was pulled within days • 𝗲𝘁𝗵𝗲𝗿𝟬 (FutureHouse, 2025): 24B, chemistry-only reasoning, Apache-2.0 • 𝗡𝗮𝘁𝘂𝗿𝗲𝗟𝗠 (Microsoft, 2025): one model for molecules, proteins, materials, DNA and RNA sequences, MIT • 𝗜𝗻𝘁𝗲𝗿𝗻-𝗦𝟭-𝗣𝗿𝗼 (Shanghai AI Lab, 2026): ~1T MoE, multimodal, Apache-2.0 • 𝗛𝘆𝗱𝗿𝗼𝗴𝗲𝗻-𝟭.𝟬 (Deep Principle, 2026): 753B, post-trained from GLM-5.3, MIT weights 𝗪𝗵𝗲𝗿𝗲 𝗛𝘆𝗱𝗿𝗼𝗴𝗲𝗻 𝗳𝗶𝘁𝘀: It doesn't train a base model. It takes a strong open general model and 𝗮𝗱𝗱𝘀 𝗹𝗮𝗯 𝗱𝗮𝘁𝗮 𝗼𝗻 𝘁𝗼𝗽. That data is the point: • Lab notebook records from Deep Principle's own labs • Agent discovery trajectories, 𝘧𝘢𝘪𝘭𝘦𝘥 𝘢𝘵𝘵𝘦𝘮𝘱𝘵𝘴 𝘪𝘯𝘤𝘭𝘶𝘥𝘦𝘥 • Feedback from their in-house scientists 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲𝘆 𝗰𝗹𝗮𝗶𝗺 (𝘀𝗲𝗹𝗳-𝗿𝗲𝗽𝗼𝗿𝘁𝗲𝗱): • FrontierScience-Research: 𝟯𝟴.𝟯% vs GPT-6-Astra max 𝟯𝟯.𝟯% • Base GLM-5.3: 𝟮𝟰.𝟰% → so post-training added ~14 points • GPQA-Diamond: 𝟵𝟮.𝟵%, still 𝗯𝗲𝗹𝗼𝘄 GPT-6-Astra (96.0%) and Kimi K3 (93.5%) • Loses on NMR structure elucidation (26.7% vs 30.0%) 𝗖𝗮𝘃𝗲𝗮𝘁𝘀: FrontierScience-Research is 60 problems. Hydrogen used ~3x more output tokens than GPT-6-Astra. Datasets and recipes stay private. 𝗧𝗵𝗲 𝘁𝗮𝗸𝗲: General open models are now good enough that 𝘁𝗵𝗲 𝗺𝗼𝗮𝘁 𝗺𝗼𝘃𝗲𝘀 𝘁𝗼 𝗱𝗮𝘁𝗮 𝘆𝗼𝘂 𝗰𝗮𝗻'𝘁 𝘀𝗰𝗿𝗮𝗽𝗲. Wet-lab records. Robot logs. Failed runs. 𝗖𝗼𝗺𝗽𝗮𝗻𝗶𝗲𝘀 𝘁𝗵𝗮𝘁 𝗼𝘄𝗻 𝗹𝗮𝗯𝘀 𝗮𝗿𝗲 𝗯𝗲𝗰𝗼𝗺𝗶𝗻𝗴 𝗺𝗼𝗱𝗲𝗹 𝗺𝗮𝗸𝗲𝗿𝘀. Expect more "GLM / Qwen / DeepSeek + private lab data" releases in chemistry, materials and biology. Weights (Hugging Face): huggingface.co/deep-principl… Links in reply 👇

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 👇

Tried Karpathy's HackerTube tonight. I searched "AI blogs" and the first result was a podcast with Karpathy himself. It has 𝟮𝟰𝟮𝗞 𝘃𝗶𝗲𝘄𝘀, but it showed up above a Tom Scott video with 𝟴.𝟱𝗠 𝘃𝗶𝗲𝘄𝘀. 𝗩𝗶𝗲𝘄𝘀 𝗱𝗼𝗻'𝘁 𝗺𝗮𝘁𝘁𝗲𝗿 on HackerTube. It ranks videos by how many 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗺𝗼𝗻𝘁𝗵𝘀 HN people recommended them. HN people have shared his talks for years, so he deserves the top spot. But there is a downside. You get away from YouTube's algorithm, but you get 𝗛𝗡'𝘀 𝗳𝗮𝘃𝗼𝘂𝗿𝗶𝘁𝗲𝘀 instead. Still, I like HN's taste more than YouTube's.

A curated list of 308 open-source AI tools you can run on your own hardware. It covers runtimes, agents, coding assistants, RAG, MCP servers, speech and evals, in 18 categories. The useful part is that a weekly bot checks every project on GitHub and moves dead ones to an “unmaintained” list. github.com/glincker/awesome-…

𝗧𝗵𝗲 𝗢𝗽𝗲𝗻-𝗦𝗼𝘂𝗿𝗰𝗲 𝗔𝗻𝘀𝘄𝗲𝗿 𝘁𝗼 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝘂𝗲𝘁 𝗮𝗻𝗱 𝗢𝗽𝗲𝗻𝗔𝗜 𝗗𝗼𝘁𝘀 You ask a chatbot for something. It answers. Then it stops and waits. But real work doesn't fit in one question. A price you want to watch every day. A research job that takes half an hour. A report due every Monday. That's why OpenAI launched dots in September and Anthropic launched Claude Duet in October: 𝗮𝗴𝗲𝗻𝘁𝘀 𝘁𝗵𝗮𝘁 𝗸𝗲𝗲𝗽 𝘄𝗼𝗿𝗸𝗶𝗻𝗴 𝘄𝗵𝗲𝗻 𝘆𝗼𝘂 𝘄𝗮𝗹𝗸 𝗮𝘄𝗮𝘆. Both run in the company's cloud, and you need their subscription. AnyModelBot is an open-source (MIT) version of the same idea that 𝗿𝘂𝗻𝘀 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 You create "bots." Each one has its own persona, its own default model, its own folder and its own memory. You talk to them in the desktop app, or from WeChat, Feishu, Telegram, Slack or Discord. They can: • 𝗗𝗼 𝗯𝗮𝗰𝗸𝗴𝗿𝗼𝘂𝗻𝗱 𝘁𝗮𝘀𝗸𝘀: "compare GPU prices at three clouds" runs on its own and reports back in the same chat • 𝗙𝗼𝗹𝗹𝗼𝘄 𝗿𝘂𝗹𝗲𝘀: "check this price at 8am, tell me if it drops below 200" • 𝗨𝘀𝗲 𝘆𝗼𝘂𝗿 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿: read and write files, run commands, click and type on your desktop • 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿 𝘆𝗼𝘂: memory is plain Markdown files on your disk 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀 It asks before it acts outside its own folder. There are three permission levels: read-only, its own folder (the default), and full access. It has 𝘀𝗽𝗲𝗻𝗱𝗶𝗻𝗴 𝗹𝗶𝗺𝗶𝘁𝘀 𝗯𝘂𝗶𝗹𝘁 𝗶𝗻: $1 per task by default, $0.10 per rule run, and $3 a month per rule. It also has quiet hours (10pm to 8am) and a cap on how often it messages you first. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 It's assembled from open parts: DeepSeek Harness runs the agent, magpie is the model gateway, and Cua Driver handles the computer use. It's TypeScript and Electron, with SQLite for storage. The design rule comes from DeepSeek's dsh-superagent prototype: 𝘶𝘴𝘦 𝘢𝘴 𝘭𝘪𝘵𝘵𝘭𝘦 𝘰𝘧 𝘺𝘰𝘶𝘳 𝘢𝘵𝘵𝘦𝘯𝘵𝘪𝘰𝘯 𝘢𝘴 𝘱𝘰𝘴𝘴𝘪𝘣𝘭𝘦, 𝘢𝘯𝘥 𝘣𝘳𝘪𝘯𝘨 𝘢𝘴 𝘮𝘶𝘤𝘩 𝘷𝘢𝘭𝘶𝘦 𝘢𝘴 𝘱𝘰𝘴𝘴𝘪𝘣𝘭𝘦. 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁 "Local-first" means your 𝗰𝗵𝗮𝘁𝘀, 𝗺𝗲𝗺𝗼𝗿𝘆, 𝗳𝗶𝗹𝗲𝘀 𝗮𝗻𝗱 𝗸𝗲𝘆𝘀 stay on your device. 𝗧𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗰𝗮𝗹𝗹𝘀 𝗱𝗼𝗻'𝘁. Every model request goes through Modelflare, a paid API gateway run by the same people. You pay per token. Modelflare keeps request logs. Using your own endpoint is documented as "𝗱𝗲𝘃/𝘁𝗲𝘀𝘁 𝗼𝗻𝗹𝘆." It's also 𝘀𝗶𝘅 𝗱𝗮𝘆𝘀 𝗼𝗹𝗱 (v0.1.15, 652 stars as of Nov 2). Telegram, Slack and Discord work in DMs only. Computer use on macOS is in beta. And it's not the first open-source agent like this. OpenClaw has been doing "an AI that really does things" since late 2025. AnyModelBot's angle is WeChat and Feishu, budgets, and the polite "only bother me when it matters" design. 𝘛𝘩𝘦 𝘳𝘦𝘢𝘭 𝘧𝘪𝘨𝘩𝘵 𝘪𝘯 𝘱𝘦𝘳𝘴𝘰𝘯𝘢𝘭 𝘢𝘨𝘦𝘯𝘵𝘴 𝘪𝘴𝘯'𝘵 𝘵𝘩𝘦 𝘢𝘱𝘱 𝘢𝘯𝘺𝘮𝘰𝘳𝘦. 𝘐𝘵'𝘴 𝘸𝘩𝘰 𝘺𝘰𝘶 𝘱𝘢𝘺 𝘧𝘰𝘳 𝘵𝘩𝘦 𝘮𝘰𝘥𝘦𝘭 𝘤𝘢𝘭𝘭𝘴. Links in reply 👇 github.com/modelflare-ai/Any…

𝗡𝗲𝘂𝗿𝗼𝘀𝘂𝗿𝗴𝗲𝗿𝘆 𝗳𝗼𝗿 𝗖𝗵𝗮𝘁𝗯𝗼𝘁𝘀, 𝗡𝗼 𝗦𝗲𝗿𝘃𝗲𝗿 𝗥𝗲𝗾𝘂𝗶𝗿𝗲𝗱 𝗧𝘂𝗿𝗻 𝗢𝗻𝗲 𝗞𝗻𝗼𝗯 𝗜𝗻𝘀𝗶𝗱𝗲 𝗮𝗻 𝗔𝗜 𝗮𝗻𝗱 𝗪𝗮𝘁𝗰𝗵 𝗜𝘁 𝗢𝗯𝘀𝗲𝘀𝘀 𝗢𝘃𝗲𝗿 𝗖𝗼𝗼𝗸𝗶𝗲𝘀 In 2024, Anthropic made "Golden Gate Claude." They found the part of Claude that lights up for the Golden Gate Bridge, turned it up, and Claude started talking about the bridge in every answer. It was online for about a day. You could chat with it, but you couldn't touch the knobs. Now you can do the same kind of thing yourself, 𝗶𝗻 𝗮 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝘁𝗮𝗯, 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗹𝗮𝗽𝘁𝗼𝗽. Neuron Studio is a free Hugging Face Space. It loads Qwen3.5-4B, a small open model, and lets you 𝗿𝗲𝗮𝗰𝗵 𝗶𝗻𝘀𝗶𝗱𝗲 𝗶𝘁 𝗮𝗻𝗱 𝘁𝘂𝗿𝗻 𝘀𝗶𝗻𝗴𝗹𝗲 𝗻𝗲𝘂𝗿𝗼𝗻𝘀 𝘂𝗽 𝗼𝗿 𝗱𝗼𝘄𝗻. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 The model has 𝟯𝟮 𝗹𝗮𝘆𝗲𝗿𝘀. Each layer has 𝟵,𝟮𝟭𝟲 𝗠𝗟𝗣 𝗻𝗲𝘂𝗿𝗼𝗻𝘀, shown as a 96 × 96 grid of dots. You can fly around them in 3D. Click a neuron. Give it a dose from −𝟯 𝘁𝗼 +𝟯. Type a prompt. You get two answers side by side: the normal model and the changed one. Both use the same prompt and no randomness, so 𝘆𝗼𝘂𝗿 𝗰𝗵𝗮𝗻𝗴𝗲 𝗶𝘀 𝘁𝗵𝗲 𝗼𝗻𝗹𝘆 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲. 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀 There are 20 ready-made "recipes," each just one or two neurons. In the app's own words: • 𝗖𝗼𝗼𝗸𝗶𝗲: everything turns into cookies (layer 24, neuron 3102) • 𝗬𝗼𝘂'𝗿𝗲 𝗮𝗯𝘀𝗼𝗹𝘂𝘁𝗲𝗹𝘆 𝗿𝗶𝗴𝗵𝘁: "𝘈𝘨𝘳𝘦𝘦𝘴 𝘸𝘪𝘵𝘩 𝘸𝘩𝘢𝘵 𝘺𝘰𝘶 𝘯𝘦𝘷𝘦𝘳 𝘴𝘢𝘪𝘥" • 𝗜𝗻𝗱𝗼𝗻𝗲𝘀𝗶𝗮𝗻: answers in Indonesian • 𝗔𝗟𝗟 𝗖𝗔𝗣𝗦: shouts The model is a 𝟮.𝟵𝟭 𝗚𝗕 𝗱𝗼𝘄𝗻𝗹𝗼𝗮𝗱, cached after the first time. It runs on your GPU through WebGPU, and the page says 𝘺𝘰𝘶𝘳 𝘱𝘳𝘰𝘮𝘱𝘵𝘴 𝘯𝘦𝘷𝘦𝘳 𝘭𝘦𝘢𝘷𝘦 𝘵𝘩𝘦 𝘱𝘢𝘨𝘦. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 The steering method is written out in the app. It adds an offset to one neuron's activation, just before the MLP's down projection, on every token. The dose is scaled by the layer's typical activation size, so +1 means roughly the same push in every layer. There's also a multiply mode, where 0 switches a neuron off. You can stack up to 32 neurons and save a recipe as JSON. It all ships as 𝗼𝗻𝗲 𝗛𝗧𝗠𝗟 𝗳𝗶𝗹𝗲 with its own WebGPU code, running Unsloth's 4-bit GGUF of the model. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝗰𝗼𝗺𝗲𝘀 𝗳𝗿𝗼𝗺 It's inspired by neuron-explorer by lyramakesmusic (MIT, about 1,800 stars). That project pushes every neuron in a layer up and down and keeps every result, but you need a GPU server to run the sweeps. Neuronpedia also lets you steer models on the web, using SAE features on its servers. Neuron Studio's angle: 𝘀𝗶𝗻𝗴𝗹𝗲 𝗻𝗲𝘂𝗿𝗼𝗻𝘀, 𝗻𝗼 𝘀𝗲𝗿𝘃𝗲𝗿, 𝗷𝘂𝘀𝘁 𝗮 𝘁𝗮𝗯. 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁 These are 𝗿𝗮𝘄 𝗻𝗲𝘂𝗿𝗼𝗻𝘀, 𝗻𝗼𝘁 𝗰𝗹𝗲𝗮𝗻𝗲𝗱-𝘂𝗽 "𝗳𝗲𝗮𝘁𝘂𝗿𝗲𝘀." Golden Gate Claude used features found with extra training. Most single neurons mix many ideas, so most dots you click won't do anything clean. The 20 recipes are hand-picked, and the app doesn't say how they were found. You need 𝗮 𝗪𝗲𝗯𝗚𝗣𝗨 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝗮𝗻𝗱 𝗮 𝗱𝗲𝗰𝗲𝗻𝘁 𝗚𝗣𝗨. It's a 4B model, so the base answers are simple. And it's 𝗼𝗻𝗲 𝘄𝗲𝗲𝗸 𝗼𝗹𝗱 (32 likes as of Nov 2). 𝘎𝘰𝘭𝘥𝘦𝘯 𝘎𝘢𝘵𝘦 𝘊𝘭𝘢𝘶𝘥𝘦 𝘸𝘢𝘴 𝘢 𝘰𝘯𝘦-𝘥𝘢𝘺 𝘥𝘦𝘮𝘰. 𝘕𝘰𝘸 𝘵𝘩𝘦 𝘴𝘢𝘮𝘦 𝘪𝘥𝘦𝘢 𝘧𝘪𝘵𝘴 𝘪𝘯 𝘢 𝘣𝘳𝘰𝘸𝘴𝘦𝘳 𝘵𝘢𝘣. huggingface.co/spaces/webml-…

𝗞𝗮𝗿𝗽𝗮𝘁𝗵𝘆'𝘀 𝗛𝗮𝗰𝗸𝗲𝗿𝗧𝘂𝗯𝗲 𝘃𝘀 𝗬𝗼𝘂𝗧𝘂𝗯𝗲: 𝗥𝗲𝗺𝗲𝗺𝗯𝗲𝗿𝗲𝗱 𝘃𝘀 𝗖𝗹𝗶𝗰𝗸𝗲𝗱. There are two very different ways to decide if a video is worth your time. 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗰𝗼𝘂𝗻𝘁𝘀 𝗰𝗹𝗶𝗰𝗸𝘀. Views, watch time, what keeps you scrolling. 𝗛𝗮𝗰𝗸𝗲𝗿 𝗡𝗲𝘄𝘀 𝗰𝗼𝘂𝗻𝘁𝘀 𝘀𝗼𝗺𝗲𝘁𝗵𝗶𝗻𝗴 𝘀𝗹𝗼𝘄𝗲𝗿: someone coming back, months or years later, to tell a stranger "𝘺𝘰𝘶 𝘩𝘢𝘷𝘦 𝘵𝘰 𝘸𝘢𝘵𝘤𝘩 𝘵𝘩𝘪𝘴." Most of the time, the two agree. I wanted to find where they don't. 𝗪𝗵𝗮𝘁 𝗜 𝗱𝗶𝗱 Karpathy just launched HackerTube. Claude read 19 years of Hacker News comments and kept the YouTube videos HN keeps recommending: 𝟲,𝟰𝟱𝟰 𝗼𝗳 𝘁𝗵𝗲𝗺, ranked by how many different months people recommended each one. The data is public. It has HN's side of the story, not YouTube's. So I got a free YouTube API key and pulled the view count for all 6,454. 𝗪𝗵𝘆? If HN's favourites are also YouTube's most-watched, great videos find their audience. If not, some of the best talks are getting lost. 𝗖𝗼𝗺𝗽𝗮𝗿𝗶𝗻𝗴 𝗹𝗶𝗸𝗲 𝘄𝗶𝘁𝗵 𝗹𝗶𝗸𝗲 HackerTube sorts every video into 𝟭𝟰 𝘁𝘆𝗽𝗲𝘀: talks, lectures, explainers, demos, comedy, documentaries, music videos and more. Those types live in different worlds. The typical 𝗺𝘂𝘀𝗶𝗰 𝘃𝗶𝗱𝗲𝗼 in the collection has 𝗮𝗯𝗼𝘂𝘁 𝟲.𝟭 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝘃𝗶𝗲𝘄𝘀. The typical 𝗲𝘅𝗽𝗹𝗮𝗶𝗻𝗲𝗿 has about 2.7 million. The typical 𝘁𝗮𝗹𝗸 has about 85,000. So comparing a 50-minute conference talk with Despacito would prove nothing. I kept only 𝘁𝗮𝗹𝗸𝘀 𝗮𝗻𝗱 𝗹𝗲𝗰𝘁𝘂𝗿𝗲𝘀 (𝟭,𝟳𝟲𝟱 𝘃𝗶𝗱𝗲𝗼𝘀), then only the ones 𝟱 𝗼𝗿 𝗺𝗼𝗿𝗲 different HN commenters recommended. That left 391. The typical one has 𝗮𝗯𝗼𝘂𝘁 𝟮𝟳𝟮,𝟬𝟬𝟬 𝘃𝗶𝗲𝘄𝘀. So usually, good talks do get found. 𝗕𝘂𝘁 𝗻𝗼𝘁 𝗮𝗹𝘄𝗮𝘆𝘀 𝗧𝗵𝗲 #𝟳 𝘃𝗶𝗱𝗲𝗼 𝗼𝘂𝘁 𝗼𝗳 𝗮𝗹𝗹 𝟲,𝟰𝟱𝟰 is Armin Ronacher's "𝘏𝘰𝘸 𝘗𝘺𝘵𝘩𝘰𝘯 𝘸𝘢𝘴 𝘚𝘩𝘢𝘱𝘦𝘥 𝘣𝘺 𝘭𝘦𝘢𝘬𝘺 𝘐𝘯𝘵𝘦𝘳𝘯𝘢𝘭𝘴." 𝟭𝟱 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗛𝗡 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝗲𝗿𝘀 recommended it. It has 𝗮𝗯𝗼𝘂𝘁 𝟮𝟬,𝟬𝟬𝟬 𝘃𝗶𝗲𝘄𝘀. That's more than 13 times below a typical talk at the same level. It's not alone: • 𝗝𝗼𝗵𝗻 𝗦𝗼𝘁𝗼𝘀 𝗮𝘁 𝗗𝗘𝗙 𝗖𝗢𝗡 𝟮𝟱: 11 recommenders, 𝗮𝗯𝗼𝘂𝘁 𝟭𝟯,𝟵𝟬𝟬 𝘃𝗶𝗲𝘄𝘀 • 𝗖𝗵𝘂𝗰𝗸 𝗠𝗼𝗼𝗿𝗲 𝗼𝗻 𝗮 𝟭𝟰𝟰-𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗰𝗵𝗶𝗽: 7 recommenders, 𝗮𝗯𝗼𝘂𝘁 𝟭𝟵,𝟭𝟬𝟬 𝘃𝗶𝗲𝘄𝘀 • 𝗠𝗼𝘅𝗶𝗲 𝗠𝗮𝗿𝗹𝗶𝗻𝘀𝗽𝗶𝗸𝗲 𝗮𝘁 𝗗𝗘𝗙 𝗖𝗢𝗡 𝟭𝟴: 6 recommenders, 𝗮𝗯𝗼𝘂𝘁 𝟮𝟭,𝟳𝟬𝟬 𝘃𝗶𝗲𝘄𝘀 𝗙𝗼𝗿 𝘃𝗶𝗲𝘄𝗲𝗿𝘀: the view count tells you what's popular, not what people still remember years later. These four are worth a weekend. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀: HackerTube's data is free to reuse. Joining it with one more source took one API key and about 130 requests. 𝗧𝗼 𝗯𝗲 𝗳𝗮𝗶𝗿 • 𝗧𝗵𝗲 𝗱𝗮𝘁𝗮 𝘀𝗵𝗼𝘄𝘀 𝘁𝗵𝗲 𝘃𝗶𝗲𝘄𝘀 𝗮𝗿𝗲 𝗹𝗼𝘄, 𝗻𝗼𝘁 𝘄𝗵𝘆. Long technical talks on small channels may simply have a small audience. • 𝗛𝗮𝗰𝗸𝗲𝗿𝗧𝘂𝗯𝗲'𝘀 𝗵𝗼𝗺𝗲𝗽𝗮𝗴𝗲 𝗮𝗹𝗿𝗲𝗮𝗱𝘆 𝗺𝗶𝘅𝗲𝘀 𝗶𝗻 "𝗵𝗶𝗱𝗱𝗲𝗻 𝗴𝗲𝗺𝘀." What's new here is the numbers. • I searched YouTube for other copies of each talk. I dropped one (Guy Steele's "Growing a Language") because 𝗮𝗻𝗼𝘁𝗵𝗲𝗿 𝘂𝗽𝗹𝗼𝗮𝗱 𝗵𝗮𝘀 𝗮𝗯𝗼𝘂𝘁 𝟭𝟲𝟱,𝟬𝟬𝟬 𝘃𝗶𝗲𝘄𝘀. • The types and "recommenders" are labels written by Claude from HN comments. Counts are as of 2 Nov 2026. 𝗩𝗶𝗲𝘄𝘀 𝘁𝗲𝗹𝗹 𝘆𝗼𝘂 𝘄𝗵𝗮𝘁'𝘀 𝗽𝗼𝗽𝘂𝗹𝗮𝗿. 𝗛𝗮𝗰𝗸𝗲𝗿 𝗡𝗲𝘄𝘀 𝘁𝗲𝗹𝗹𝘀 𝘆𝗼𝘂 𝘄𝗵𝗮𝘁 𝗽𝗲𝗼𝗽𝗹𝗲 𝘀𝘁𝗶𝗹𝗹 𝗿𝗲𝗺𝗲𝗺𝗯𝗲𝗿. 𝘞𝘩𝘪𝘤𝘩 𝘩𝘪𝘥𝘥𝘦𝘯 𝘵𝘢𝘭𝘬 𝘸𝘰𝘶𝘭𝘥 𝘺𝘰𝘶 𝘢𝘥𝘥 𝘵𝘰 𝘵𝘩𝘪𝘴 𝘭𝘪𝘴𝘵? Links in reply 👇

𝗠𝘆 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗛𝗼𝗺𝗲𝗽𝗮𝗴𝗲 𝘃𝘀 𝗛𝗮𝗰𝗸𝗲𝗿𝗧𝘂𝗯𝗲 I put my YouTube homepage next to Karpathy's HackerTube. Top 6 videos on each. 𝗧𝗲𝘅𝘁 𝗮𝗱𝗱𝗲𝗱 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗵𝘂𝗺𝗯𝗻𝗮𝗶𝗹: • YouTube: 𝟲 𝗼𝗳 𝟲 • HackerTube: 𝟬 𝗼𝗳 𝟲 "?" 𝗼𝗿 "!" 𝗼𝗻 𝘁𝗵𝗲 𝘁𝗵𝘂𝗺𝗯𝗻𝗮𝗶𝗹: • YouTube: 𝟯 𝗼𝗳 𝟲 • HackerTube: 𝟬 𝗼𝗳 𝟲 Part of that is by design. HackerTube doesn't show the creator's thumbnail at all. It shows 𝗮 𝗳𝗿𝗮𝗺𝗲 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗺𝗶𝗱𝗱𝗹𝗲 𝗼𝗳 𝘁𝗵𝗲 𝘃𝗶𝗱𝗲𝗼. Creators design the thumbnail to get the click. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗱𝗲𝘀𝗶𝗴𝗻𝘀 𝘁𝗵𝗲 𝗵𝗮𝗹𝗳𝘄𝗮𝘆 𝗳𝗿𝗮𝗺𝗲. And these are mostly channels I chose to follow. Even the people I like are playing the thumbnail game, because that's what the algorithm rewards. To be fair: • One of my 6 is an ad. • HackerTube still shows YouTube's own titles. One of its 6 is "𝘛𝘩𝘦 𝘴𝘱𝘦𝘦𝘤𝘩 𝘵𝘩𝘢𝘵 𝘣𝘳𝘰𝘬𝘦 𝘵𝘩𝘦 𝘪𝘯𝘵𝘦𝘳𝘯𝘦𝘵 - 𝘔𝘰𝘴𝘵 𝘐𝘯𝘴𝘱𝘪𝘳𝘪𝘯𝘨 𝘦𝘷𝘦𝘳." • My homepage is personal. Yours will look different. 𝘞𝘩𝘢𝘵 𝘥𝘰𝘦𝘴 𝘺𝘰𝘶𝘳 𝘠𝘰𝘶𝘛𝘶𝘣𝘦 𝘩𝘰𝘮𝘦𝘱𝘢𝘨𝘦 𝘭𝘰𝘰𝘬 𝘭𝘪𝘬𝘦 𝘳𝘪𝘨𝘩𝘵 𝘯𝘰𝘸? 𝘗𝘰𝘴𝘵 𝘢 𝘴𝘤𝘳𝘦𝘦𝘯𝘴𝘩𝘰𝘵 👇

𝗡𝗼𝘁 𝗔𝗻𝗼𝘁𝗵𝗲𝗿 𝗞𝗮𝗿𝗽𝗮𝘁𝗵𝘆 𝗛𝗮𝗰𝗸𝗲𝗿𝗧𝘂𝗯𝗲 𝗣𝗼𝘀𝘁: 𝗧𝗵𝗲 𝗥𝗲𝗮𝗹 𝗦𝘁𝗼𝗿𝘆 𝗦𝘁𝗮𝗿𝘁𝗲𝗱 𝗶𝗻 𝟭𝟵𝟰𝟱 For most of 2026 my feed has been one man's weekend projects. In April 2026 it was Karpathy's LLM Wiki. Since Saturday it is HackerTube, and I think I already know how the next few months will look. 𝗟𝗼𝗼𝗸 𝗮𝘁 𝗵𝗶𝘀 𝗹𝗮𝘀𝘁 𝟮𝟬 𝗺𝗼𝗻𝘁𝗵𝘀: • 𝗙𝗲𝗯 𝟮𝟬𝟮𝟱: one tweet coins "𝘷𝘪𝘣𝘦 𝘤𝘰𝘥𝘪𝘯𝘨", meant for "𝘵𝘩𝘳𝘰𝘸𝘢𝘸𝘢𝘺 𝘸𝘦𝘦𝘬𝘦𝘯𝘥 𝘱𝘳𝘰𝘫𝘦𝘤𝘵𝘴." Collins Dictionary made it Word of the Year in Nov 2025. Collins started in Glasgow in 1819 and still has its dictionary team here, my city, so I smiled at that one. • 𝗢𝗰𝘁 𝟮𝟬𝟮𝟱: nanochat, a full ChatGPT pipeline you can train for 𝗮𝗯𝗼𝘂𝘁 $𝟭𝟬𝟬. He wrote it mostly by hand because coding agents were not good enough for it yet. • 𝗠𝗮𝗿 𝟮𝟬𝟮𝟲: autoresearch, agents running training experiments overnight. • 𝗔𝗽𝗿 𝟮𝟬𝟮𝟲: the LLM Wiki gist. No program, just one markdown "idea file", and it has 𝗼𝘃𝗲𝗿 𝟭𝟬,𝟬𝟬𝟬 𝗳𝗼𝗿𝗸𝘀. • 𝗢𝗰𝘁 𝟮𝟬𝟮𝟲: HackerTube. On Friday night YouTube showed him a wall of clickbait. On Saturday 31 Oct 2026 he asked Claude Code with Opus 5.5 to build his own YouTube from videos people recommended on Hacker News. Claude pulled HN comments from 2007 to 2026, Sonnet filtered out the spam, and about 𝟰𝟬,𝟬𝟬𝟬 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝘀 that recommend a video were left. That gave 𝟲,𝟰𝟱𝟰 𝘃𝗶𝗱𝗲𝗼𝘀 𝗶𝗻 𝟭𝟱𝟬 𝘁𝗼𝗽𝗶𝗰𝘀. About 𝟮 𝗵𝗼𝘂𝗿𝘀. He says he doesn't do "𝘢𝘯𝘺𝘵𝘩𝘪𝘯𝘨 𝘳𝘦𝘴𝘦𝘮𝘣𝘭𝘪𝘯𝘨 𝘵𝘳𝘢𝘥𝘪𝘵𝘪𝘰𝘯𝘢𝘭 𝘦𝘯𝘨𝘪𝘯𝘦𝘦𝘳𝘪𝘯𝘨 𝘢𝘯𝘺𝘮𝘰𝘳𝘦." He only did the bits that needed his own accounts: a YouTube API key, a server and the domain. So in one year he went from writing nanochat by hand to a live site with 𝗼𝘃𝗲𝗿 𝟮,𝟬𝟬𝟬 𝗛𝗡 𝗽𝗼𝗶𝗻𝘁𝘀, built by talking to Claude. 𝗘𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗶𝘀 𝗰𝗮𝗹𝗹𝗶𝗻𝗴 𝗶𝘁 𝗮 𝗬𝗼𝘂𝗧𝘂𝗯𝗲 𝗰𝗹𝗼𝗻𝗲. 𝗜 𝘁𝗵𝗶𝗻𝗸 𝗶𝘁 𝗶𝘀 𝗼𝗹𝗱𝗲𝗿 𝘁𝗵𝗮𝗻 𝗬𝗼𝘂𝗧𝘂𝗯𝗲. 𝗜𝘁 𝗶𝘀 𝗣𝗮𝗴𝗲𝗥𝗮𝗻𝗸'𝘀 𝗶𝗱𝗲𝗮. In 1998 Brin and Page ranked web pages by treating every link as a human vote, an idea borrowed from how scientists count citations. Google bought YouTube in 2006. After years of tuning for watch time, Karpathy uses Google's first idea to escape Google's own product. The votes are HN comments now, not links. 𝗠𝘆 𝗳𝗮𝘃𝗼𝘂𝗿𝗶𝘁𝗲 𝗱𝗲𝘁𝗮𝗶𝗹: HackerTube ignores the thumbnail the creator uploaded and shows the frame from the middle of the video. Creators design the thumbnail to get clicks. 𝗡𝗼𝗯𝗼𝗱𝘆 𝗱𝗲𝘀𝗶𝗴𝗻𝘀 𝘁𝗵𝗲 𝗵𝗮𝗹𝗳𝘄𝗮𝘆 𝗳𝗿𝗮𝗺𝗲. The LLM Wiki and HackerTube look like different things. To me they are one idea. People already wrote down good judgment in notes and in comments. Nobody could pay the cost of reading and organising all of it. In the gist he points to Vannevar Bush's Memex from 1945: "𝘛𝘩𝘦 𝘱𝘢𝘳𝘵 𝘩𝘦 𝘤𝘰𝘶𝘭𝘥𝘯'𝘵 𝘴𝘰𝘭𝘷𝘦 𝘸𝘢𝘴 𝘸𝘩𝘰 𝘥𝘰𝘦𝘴 𝘵𝘩𝘦 𝘮𝘢𝘪𝘯𝘵𝘦𝘯𝘢𝘯𝘤𝘦. 𝘛𝘩𝘦 𝘓𝘓𝘔 𝘩𝘢𝘯𝘥𝘭𝘦𝘴 𝘵𝘩𝘢𝘵." Bush's essay was also an idea file. No machine, just a description. Doug Engelbart, a 20-year-old Navy radar technician, read it in a Red Cross library in a hut on stilts on Leyte in the Philippines in autumn 1945. He showed the mouse and hypertext in 1968. The web came in 1991. 𝗧𝗵𝗮𝘁 𝗶𝗱𝗲𝗮 𝗻𝗲𝗲𝗱𝗲𝗱 𝟰𝟲 𝘆𝗲𝗮𝗿𝘀 𝗮𝗻𝗱 𝗮 𝗳𝗲𝘄 𝗿𝗮𝗿𝗲 𝗽𝗲𝗼𝗽𝗹𝗲 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱. 𝗞𝗮𝗿𝗽𝗮𝘁𝗵𝘆'𝘀 𝗶𝗱𝗲𝗮𝘀 𝗻𝗲𝗲𝗱 𝗮 𝘄𝗲𝗲𝗸𝗲𝗻𝗱 𝗮𝗻𝗱 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁. He even put a data export on HackerTube so you can hand it to your agent and make your own version. I don't think he ships products anymore. He ships 𝘁𝗵𝗶𝗻𝗴𝘀 𝘀𝗶𝘇𝗲𝗱 𝘁𝗼 𝗯𝗲 𝗰𝗼𝗽𝗶𝗲𝗱, and the months of posts after each drop are all of us building our copy. 𝗠𝘆 𝗱𝗼𝘂𝗯𝘁: HN comments were a clean signal because almost nobody ranked anything by them. Goodhart's law says once a number becomes a target, it stops measuring what it used to. If HackerTube gets big, people will start dropping their own videos in HN threads, and 𝘁𝗵𝗲 𝗰𝗹𝗶𝗰𝗸𝗯𝗮𝗶𝘁 𝗺𝗼𝘃𝗲𝘀 𝗳𝗿𝗼𝗺 𝘁𝗵𝘂𝗺𝗯𝗻𝗮𝗶𝗹𝘀 𝗶𝗻𝘁𝗼 𝗰𝗼𝗺𝗺𝗲𝗻𝘁𝘀. 𝗠𝘆 𝗴𝘂𝗲𝘀𝘀 𝗳𝗼𝗿 𝘁𝗵𝗲 𝗰𝗼𝗺𝗶𝗻𝗴 𝗺𝗼𝗻𝘁𝗵𝘀: HackerTube for Reddit, for arXiv, for podcasts, one per niche forum. It has already started: "𝘏𝘢𝘤𝘬𝘦𝘳𝘉𝘰𝘰𝘬𝘴 – 𝘏𝘢𝘤𝘬𝘦𝘳𝘛𝘶𝘣𝘦, 𝘣𝘶𝘵 𝘧𝘰𝘳 𝘣𝘰𝘰𝘬𝘴" hit HN the next day. Then some Saturday he drops the next thing and we start again. Links in reply 👇