Gul Saeed Khattak

A tale of two Huangs. Same last name. Same affair For one of them, it's personal drama. For Jensen, it's billions, export rules and a betting market on when he's out. Makes you wonder. Does a mistake cost what it is, or what you've got to lose? God knows where this one ends for him. Being Janson is expensive man !!

Jansen Huang NVIDIA CEO has resigned. Links in replies 🔽🔽🔽👇

Elon owns X. Tibo owns the reset button. Elon has 241 million followers. Tibo .. Lord of the Resets …from OpenAi has 1 Million Yet Tibo tops X’s web-profile traffic ranking for the second month running. Apparently, people are more desperate for Codex resets than another Elon tweet.

I told @MatthewBerman's AMA last night that this could not be true. I was wrong. Someone on the stream said Claude read lips from a video. I replied that Claude has no lip reading tool. It can only pull still pictures out of a video, and a still picture of a mouth tells you almost nothing. I had asked Claude myself earlier and it told me no. Then I gave it a muted clip from the start of the Moonshots podcast, this time in a session with internet, and said "try lip reading". It still can't read lips itself. It went and got Auto-AVSR, an open lip reading model. It cloned the repo, pulled 1 GB of weights (the first download got cut off, so it retried), waited on a 6 GB PyTorch install, found the speaker's face in the frames and ran the model. The transcript matched what was said. I would not call that Claude lip reading. A separate model did the seeing. But I did not expect it to get from "I can't" to a working lip reader in one session. I recorded the whole thing. Video below. xitter.twitterapi.xyz/ldataengineer/status/2…

NVIDIA Built a 3D Model That Goes Round in Loops NVIDIA released this one as a full repo, not just a paper. The code and weights are out, so you can run it yourself. And I think the idea reaches past 3D. Big multi-view models like VGGT (1.26B parameters) seem to refine their answer a little at each layer. The authors argue that a tall stack of unique layers is partly just paying for repeated refinement, and paying for it in parameters. 𝗦𝗼 𝘁𝗵𝗲𝘆 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝗹𝗼𝗼𝗽 𝗲𝘅𝗽𝗹𝗶𝗰𝗶𝘁. One transformer block, run K times. The model, Déjà View, is 117M parameters. An analogy: a deep stack is like hiring a long line of editors who each read the draft once. A loop is one editor reading it many times. When most of the job is "make it a bit better again", the single editor can be enough. The ablation is the part to look at. Same block design, but with separate weights per step, trained on the same data with the same compute. 𝗧𝗵𝗲 𝘀𝗵𝗮𝗿𝗲𝗱-𝘄𝗲𝗶𝗴𝗵𝘁 𝗹𝗼𝗼𝗽 𝗯𝗲𝗮𝘁 𝗶𝘁. imo that's the real finding. It hints the loop is a good prior for this task, not only a way to save memory. K is also a knob at run time. The released checkpoint works anywhere from 8 to 16 steps, from one trained file. The GitHub repo has inference and training code, though only the ScanNet++ training data loader is out so far. The weights are on Hugging Face under a non-commercial licence, and all benchmark numbers are NVIDIA's own. github.com/nv-tlabs/dvlt Links in reply 👇

0 is moving fast. What caught my attention is that you’re not tied to any particular AI model. You can use DeepSeek, Qwen, GLM, Kimi, or even run your own model locally. Pick the model you like. Let 0 do the security work.

𝗚𝗿𝗼𝗸 𝗕𝗼𝘁 𝟭.𝟬 𝗝𝘂𝘀𝘁 𝗝𝗼𝗶𝗻𝗲𝗱 𝗬𝗼𝘂𝗿 𝗧𝗲𝗮𝗺 𝗖𝗵𝗮𝘁. 𝗜𝘁'𝘀 𝗡𝗼𝘁 𝘁𝗵𝗲 𝗙𝗶𝗿𝘀𝘁 𝗔𝗜 𝘁𝗼 𝗗𝗼 𝗜𝘁. Here is a short History Grok Bot 1.0 is out. Its Bots now sit in Slack and Microsoft Teams, join video calls and, in the company's words, "𝘵𝘢𝘬𝘦 𝘰𝘯 𝘸𝘰𝘳𝘬 𝘸𝘪𝘵𝘩𝘰𝘶𝘵 𝘣𝘦𝘪𝘯𝘨 𝘢𝘴𝘬𝘦𝘥." It sounds new. Most of it has been done before, so here's the short history of AI in the team chat. 𝗠𝗮𝗿𝗰𝗵 𝟮𝟬𝟮𝟯: 𝘁𝗵𝗲 𝗰𝗵𝗮𝘁 𝗵𝗲𝗹𝗽𝗲𝗿𝘀 • 𝗖𝗵𝗮𝘁𝗚𝗣𝗧 𝗮𝗽𝗽 𝗳𝗼𝗿 𝗦𝗹𝗮𝗰𝗸 was announced on 7 March 2023, as a beta with a waitlist • 𝗖𝗹𝗮𝘂𝗱𝗲 𝗮𝗽𝗽 𝗳𝗼𝗿 𝗦𝗹𝗮𝗰𝗸 followed on 30 March 2023, also in beta These are the earliest two I could confirm. Both did the same simple job: you asked, and they summarised a thread or drafted a reply. They never spoke first. 𝗦𝗲𝗽𝘁𝗲𝗺𝗯𝗲𝗿 𝟮𝟬𝟮𝟱: 𝘁𝗵𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴 𝗻𝗼𝘁𝗲-𝘁𝗮𝗸𝗲𝗿 Microsoft's Facilitator agent became generally available in Teams meetings. What it added: 𝗔𝗜 𝗶𝗻𝘀𝗶𝗱𝗲 𝘁𝗵𝗲 𝗺𝗲𝗲𝘁𝗶𝗻𝗴, taking notes while people talk. 𝗔𝗽𝗿𝗶𝗹 𝟮𝟬𝟮𝟲: 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮 𝘁𝗲𝗮𝗺 𝗰𝗮𝗻 𝘀𝗵𝗮𝗿𝗲 OpenAI launched workspace agents on 22 April 2026, as a research preview. What they added: an agent built once and 𝘀𝗵𝗮𝗿𝗲𝗱 𝘄𝗶𝘁𝗵 𝘁𝗲𝗮𝗺𝗺𝗮𝘁𝗲𝘀, able to run on a schedule and pick up requests in Slack. 𝗝𝘂𝗻𝗲 𝟮𝟬𝟮𝟲: 𝘁𝗵𝗲 𝘁𝗲𝗮𝗺𝗺𝗮𝘁𝗲 𝘁𝗵𝗮𝘁 𝘀𝗽𝗲𝗮𝗸𝘀 𝗳𝗶𝗿𝘀𝘁 Anthropic launched Claude Tag in beta on 23 June 2026. What it added: 𝗼𝗻𝗲 𝘀𝗵𝗮𝗿𝗲𝗱 𝗖𝗹𝗮𝘂𝗱𝗲 𝗳𝗼𝗿 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝘁𝗲𝗮𝗺 that learns the channel over time and, with its ambient mode on, follows up on stalled threads by itself. So "shared" and "proactive" were already here before Grok Bot 1.0. 𝗡𝗼𝘃𝗲𝗺𝗯𝗲𝗿 𝟮𝟬𝟮𝟲: 𝘄𝗵𝗮𝘁 𝗚𝗿𝗼𝗸 𝗕𝗼𝘁 𝟭.𝟬 𝗱𝗼𝗲𝘀 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁𝗹𝘆 • 𝗦𝗹𝗮𝗰𝗸 𝗮𝗻𝗱 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗧𝗲𝗮𝗺𝘀. Claude Tag was Slack only at launch • 𝗩𝗶𝗱𝗲𝗼 𝗰𝗮𝗹𝗹𝘀. A Bot joins a Google Meet, Zoom or Teams call from a link, answers in the chat and sends a summary • 𝗔 𝘁𝗲𝗮𝗺 𝗼𝗳 𝗻𝗮𝗺𝗲𝗱 𝗕𝗼𝘁𝘀. A Research Bot can pass work to a Copy Bot. The others are closer to one shared assistant • 𝗜𝘁𝘀 𝗼𝘄𝗻 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿. Each Bot has a cloud computer and can sign into your tools • 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝗻𝗲𝗲𝗱𝘀 𝗮 𝘆𝗲𝘀. What teammates teach a Team Bot "𝘸𝘢𝘪𝘵𝘴 𝘶𝘯𝘥𝘦𝘳 𝘙𝘦𝘷𝘪𝘦𝘸 𝘭𝘦𝘢𝘳𝘯𝘪𝘯𝘨𝘴" until an owner or manager approves it 𝗧𝘄𝗼 𝗵𝗼𝗻𝗲𝘀𝘁 𝗻𝗼𝘁𝗲𝘀 • Grok Bot says it's "𝘧𝘢𝘴𝘵𝘦𝘳 𝘢𝘯𝘥 𝘮𝘰𝘳𝘦 𝘳𝘦𝘭𝘪𝘢𝘣𝘭𝘦" and gives 𝗻𝗼 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 • The comparisons use each product's launch features. The others may have added Teams or meetings since 𝘕𝘰𝘣𝘰𝘥𝘺 𝘪𝘯𝘷𝘦𝘯𝘵𝘦𝘥 𝘵𝘩𝘦 𝘈𝘐 𝘵𝘦𝘢𝘮𝘮𝘢𝘵𝘦 𝘪𝘯 𝘰𝘯𝘦 𝘨𝘰. 𝘌𝘢𝘤𝘩 𝘭𝘢𝘶𝘯𝘤𝘩 𝘢𝘥𝘥𝘦𝘥 𝘰𝘯𝘦 𝘱𝘪𝘦𝘤𝘦, 𝘢𝘯𝘥 𝘎𝘳𝘰𝘬 𝘉𝘰𝘵 1.0 𝘪𝘴 𝘵𝘩𝘦 𝘭𝘢𝘵𝘦𝘴𝘵 𝘱𝘪𝘦𝘤𝘦. Links in reply 👇

Google Gemini Decision 1 pricing is kind of wild: $0.10 per million input tokens, output is free. Google says it’s ~5x cheaper than Flash-Lite on their test suite. On Google’s own charts it beats TypeSafe’s Jev on calibration. On the independent Artificial Analysis leaderboard it’s 9th of 18. Cheap and confident, but not the king yet.

Agent_Skills: A Trading Firm Open-Sourced a Tutor, Not a Trading Bot One of Wall Street’s most secretive trading firms just open-sourced something. It isn’t a trading bot. It’s a tutor that refuses to give you the answer. 𝗝𝗮𝗻𝗲 𝗦𝘁𝗿𝗲𝗲𝘁'𝘀 𝗙𝗶𝗿𝘀𝘁 𝗣𝘂𝗯𝗹𝗶𝗰 𝗔𝗜 𝗦𝗸𝗶𝗹𝗹 𝗪𝗼𝗻'𝘁 𝗚𝗶𝘃𝗲 𝗬𝗼𝘂 𝘁𝗵𝗲 𝗔𝗻𝘀𝘄𝗲𝗿 You ask an AI to explain something. It writes a lovely answer, you nod along, and a week later you couldn't explain it to anyone. 𝗝𝗮𝗻𝗲 𝗦𝘁𝗿𝗲𝗲𝘁'𝘀 𝗳𝗶𝘅 𝗶𝘀 𝗮 𝘁𝘂𝘁𝗼𝗿 𝘁𝗵𝗮𝘁 𝗿𝗲𝗳𝘂𝘀𝗲𝘀 𝘁𝗼 𝗹𝗲𝗰𝘁𝘂𝗿𝗲. On 3 November the quant firm opened a GitHub repo called agent_skills. It holds exactly one skill, socratic-tutor. It's a single SKILL.md file of about 𝟴𝟰𝟱 𝘄𝗼𝗿𝗱𝘀. No code, no app, MIT licence. Nine days later it has 𝟱,𝟴𝟭𝟴 𝘀𝘁𝗮𝗿𝘀 𝗮𝗻𝗱 𝟰𝟬𝟮 𝗳𝗼𝗿𝗸𝘀 (as of 12 Nov). Part of that is the name. Jane Street usually keeps quiet, so a public AI release got attention fast, and the twist that it's a teacher and not a trading bot did the rest on X. 𝗛𝗲𝗿𝗲'𝘀 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲 𝗳𝗶𝗹𝗲 𝘁𝗲𝗹𝗹𝘀 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝘁𝗼 𝗱𝗼. • 𝗣𝗹𝗮𝗰𝗲𝗺𝗲𝗻𝘁 𝗳𝗶𝗿𝘀𝘁: split your topic into 4 to 10 subskills, then ask 3 to 5 questions before teaching anything. • 𝗔 𝘀𝗸𝗶𝗹𝗹 𝗺𝗮𝗽: every subskill sits on a 0 to 5 scale, drawn as text progress bars after each change. • 𝗢𝗻𝗲 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻 𝗮𝘁 𝗮 𝘁𝗶𝗺𝗲: the message ends at the question. No hints tacked on. • 𝗪𝗿𝗼𝗻𝗴 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝗴𝗲𝘁 𝗮 𝗻𝗮𝗿𝗿𝗼𝘄𝗲𝗿 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻, not the answer. A direct explanation only comes after 2 or 3 failed tries, and then it re-tests you from another angle. • 𝗥𝗶𝗴𝗵𝘁 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝗴𝗲𝘁 𝗽𝘂𝘀𝗵𝗯𝗮𝗰𝗸 sometimes, so you learn to defend them. It finishes only when every subskill reaches 4 out of 5, followed by a mixed final exam. 𝗧𝗵𝗲 𝘀𝗺𝗮𝗹𝗹 𝗿𝘂𝗹𝗲𝘀 𝗮𝗿𝗲 𝘁𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝗽𝗮𝗿𝘁. • 𝗔 𝗹𝗲𝘃𝗲𝗹 𝗼𝗻𝗹𝘆 𝗴𝗼𝗲𝘀 𝘂𝗽 𝗮𝗳𝘁𝗲𝗿 𝟮 𝘀𝗲𝗽𝗮𝗿𝗮𝘁𝗲 𝗰𝗵𝗲𝗰𝗸𝘀, and one has to be "𝘶𝘯𝘱𝘳𝘰𝘮𝘱𝘵𝘦𝘥 𝘳𝘦𝘤𝘢𝘭𝘭", not repeating what you were just told. • It is told: "𝘋𝘰 𝘯𝘰𝘵 𝘢𝘵𝘵𝘳𝘪𝘣𝘶𝘵𝘦 𝘵𝘰 𝘵𝘩𝘦 𝘶𝘴𝘦𝘳 𝘢𝘯𝘺 𝘪𝘯𝘴𝘪𝘨𝘩𝘵 𝘵𝘩𝘦𝘺 𝘥𝘪𝘥 𝘯𝘰𝘵 𝘢𝘤𝘵𝘶𝘢𝘭𝘭𝘺 𝘦𝘹𝘱𝘳𝘦𝘴𝘴." That's a direct guard against the model flattering you. • 𝗟𝗲𝘃𝗲𝗹 𝟱 𝗺𝗲𝗮𝗻𝘀 𝘆𝗼𝘂 𝗰𝗮𝗻 𝘁𝗲𝗮𝗰𝗵 𝗶𝘁. You have to handle an edge case and explain it well enough that the model would trust you to teach it. • If it can run tools, it checks 𝘆𝗼𝘂𝗿 𝗮𝗰𝘁𝘂𝗮𝗹 𝗼𝘂𝘁𝗽𝘂𝘁 rather than your description of it. 𝗙𝗼𝗿 𝗹𝗲𝗮𝗿𝗻𝗲𝗿𝘀: paste the file into ChatGPT, Claude or Gemini, or drop the folder into a coding agent's skills folder, and say "teach me X". Good for things you think you understand until someone asks. Is it new? Not as an idea. ChatGPT Study Mode, Gemini Guided Learning and Claude's learning mode all teach by asking questions. What's different here is a plain, strict rulebook you can read in five minutes and run on any model. One caveat: there's no evaluation in the repo showing it teaches better. 𝘛𝘩𝘦 𝘣𝘦𝘴𝘵 𝘈𝘐 𝘵𝘶𝘵𝘰𝘳 𝘮𝘪𝘨𝘩𝘵 𝘣𝘦 𝘵𝘩𝘦 𝘰𝘯𝘦 𝘵𝘩𝘢𝘵 𝘮𝘢𝘬𝘦𝘴 𝘺𝘰𝘶 𝘥𝘰 𝘵𝘩𝘦 𝘵𝘩𝘪𝘯𝘬𝘪𝘯𝘨. Links in reply 👇 Video credit: AI adventurer on YouTube (@aiadventurer-p1l). Original: youtube.com/@aiadventurer-p1…

CRAZY !!:🔇👄 Claude Can Lip Read : I Muted MoonShots Podcast and Asked Claude to Read Their Lips I screen-recorded a short clip from @moonshots_pod with @PeterDiamandis and @DaveBlundin Here's what I did. I screen-recorded a short clip from Moonshots with Peter Diamandis, about 50 seconds where the guest talks about AI chips ending up in China and the investigation around it. I stripped the audio out in CapCut. The file still had a sound track, but it was pure silence. Then I gave it to Claude (Opus 5.5) and said: lip read this. I recorded my whole screen while it worked. That's the video attached. What it did on its own: • Found a real lip-reading model: Auto-AVSR, open-source research from Imperial College London. Code from GitHub, the 1 GB model file from Hugging Face. • Checked the download: it compared the file's fingerprint with the one in the official README before trusting it. • Did the video work: found the scene cuts, cropped the area around the face, and split it into 6 short clips at the frame rate the model expects. • Ran it on a plain cloud CPU. No GPU. Note: I asked Claude: did you run this on my computer? Its answer: no. It ran in its own cloud sandbox, a machine with 4 CPU cores and 15 GB of memory. No GPU. It didn't go smoothly the whole way, which I liked. Hugging Face's download server was blocked, so it asked me to allow it. Then a safety check refused to run code downloaded from the internet, and it stopped and asked me instead of finding a way around it. Then the transcript came back. Lines like "a very big deal", "a huge number of chips", "Chinese models", "illegal imports" and "technically probably not illegal, maybe it is, maybe it isn't… that's the part where they're just gonna dig in." From a silent video. That part still gets me. It isn't perfect, and it told me so. Some lines were pure nonsense, like "the dry era of soybean play and the waistband." "Trading Chinese models" was almost certainly "training". Lip reading is hard because lots of words look the same on the lips. This model gets about 1 word in 5 wrong even on clean test videos, and mine was a phone recording of a video call. So it's not a transcript. But I knew what the conversation was about, and some sentences read like real speech. A few years ago this was a research demo. Now it's one prompt and a bit of patience. What else is hiding in all the muted videos we scroll past?

OpenAI's Dot Can Now Use More Than One Of Your Computers At The Same Time Every developer knows this wait. You start a build or a long test run. The fans spin up. The laptop slows down. You stop coding until it finishes. Many people have a second machine on the desk that sits idle the whole time. On 9 November, OpenAI gave its dots a way to use that machine. 𝗙𝗶𝗿𝘀𝘁, 𝘄𝗵𝗮𝘁 𝗮 𝗱𝗼𝘁 𝗶𝘀 OpenAI's docs call it "𝘢𝘯 𝘢𝘭𝘸𝘢𝘺𝘴-𝘰𝘯 𝘢𝘨𝘦𝘯𝘵 𝘵𝘩𝘢𝘵 𝘬𝘦𝘦𝘱𝘴 𝘸𝘰𝘳𝘬 𝘮𝘰𝘷𝘪𝘯𝘨 𝘢𝘤𝘳𝘰𝘴𝘴 𝘺𝘰𝘶𝘳 𝘵𝘰𝘰𝘭𝘴 𝘢𝘯𝘥 𝘱𝘳𝘰𝘫𝘦𝘤𝘵𝘴". It lives in the cloud with its own computer and browser, and you message it from ChatGPT, Slack or Teams. You could already connect your own computer to it. 𝗡𝗼𝘄 𝘆𝗼𝘂 𝗰𝗮𝗻 𝗰𝗼𝗻𝗻𝗲𝗰𝘁 𝘀𝗲𝘃𝗲𝗿𝗮𝗹. The docs: "𝘠𝘰𝘶 𝘤𝘢𝘯 𝘤𝘰𝘯𝘯𝘦𝘤𝘵 𝘮𝘰𝘳𝘦 𝘵𝘩𝘢𝘯 𝘰𝘯𝘦 𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳 𝘢𝘯𝘥 𝘵𝘦𝘭𝘭 𝘺𝘰𝘶𝘳 𝘥𝘰𝘵 𝘸𝘩𝘪𝘤𝘩 𝘰𝘯𝘦 𝘵𝘰 𝘶𝘴𝘦 𝘧𝘰𝘳 𝘢 𝘵𝘢𝘴𝘬." 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀 • 𝗦𝗲𝘁𝘂𝗽: open the dot's profile in the ChatGPT app on each computer, find it under Computers, and select Allow • 𝗙𝗿𝗼𝗺 𝘆𝗼𝘂𝗿 𝗽𝗵𝗼𝗻𝗲: access applies wherever you message the same dot • 𝗕𝗹𝗼𝗰𝗸𝗲𝗱 𝘄𝗲𝗯𝘀𝗶𝘁𝗲𝘀: if a site blocks the cloud browser, you can ask the dot to try a connected computer 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 • 𝗔 𝘀𝗽𝗮𝗿𝗲 𝗠𝗮𝗰 𝗮𝘀 𝗮 𝗯𝘂𝗶𝗹𝗱 𝗺𝗮𝗰𝗵𝗶𝗻𝗲: OpenAI's example is testing a change in Xcode on a second Mac while you keep coding • 𝗧𝘄𝗼 𝗲𝗻𝘃𝗶𝗿𝗼𝗻𝗺𝗲𝗻𝘁𝘀: run the same test on a current setup and on an older SDK, then compare • 𝗖𝗼𝗱𝗲𝘅: the dot can create local Work or Codex tasks and continue local Codex tasks that already exist 𝗧𝗵𝗲 𝗹𝗶𝗺𝗶𝘁𝘀 • 𝗢𝗻𝗲 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗽𝗲𝗿 𝘁𝗵𝗿𝗲𝗮𝗱. The docs say: "𝘌𝘢𝘤𝘩 𝘵𝘩𝘳𝘦𝘢𝘥 𝘶𝘴𝘦𝘴 𝘰𝘯𝘦 𝘤𝘰𝘮𝘱𝘶𝘵𝘦𝘳 𝘢𝘵 𝘢 𝘵𝘪𝘮𝘦." A single task is not split across machines • 𝗧𝗵𝗲 𝗰𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗺𝘂𝘀𝘁 𝘀𝘁𝗮𝘆 𝗼𝗻, online, with the ChatGPT app open • 𝗧𝗵𝗲 𝗵𝗲𝗮𝗱𝗹𝗶𝗻𝗲 𝘀𝗮𝘆𝘀 "𝗮𝘁 𝗼𝗻𝗰𝗲". That means separate threads running side by side, not one job spread over many machines • 𝗡𝗼 𝗺𝗮𝘅𝗶𝗺𝘂𝗺 𝗻𝘂𝗺𝗯𝗲𝗿 of computers is given in the docs • Press reports say dots are on 𝗣𝗿𝗼 𝗮𝗻𝗱 𝗕𝘂𝘀𝗶𝗻𝗲𝘀𝘀 𝗣𝗿𝗲𝗺𝗶𝘂𝗺 plans only. OpenAI's page says plan and region rules apply • More machines give more room to run things. 𝗧𝗵𝗲𝘆 𝗱𝗼 𝗻𝗼𝘁 𝗺𝗮𝗸𝗲 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝗺𝗼𝗿𝗲 𝗮𝗰𝗰𝘂𝗿𝗮𝘁𝗲 𝘛𝘩𝘦 𝘢𝘨𝘦𝘯𝘵 𝘪𝘴 𝘯𝘰 𝘭𝘰𝘯𝘨𝘦𝘳 𝘵𝘪𝘦𝘥 𝘵𝘰 𝘰𝘯𝘦 𝘮𝘢𝘤𝘩𝘪𝘯𝘦. 𝘠𝘰𝘶𝘳 𝘮𝘢𝘤𝘩𝘪𝘯𝘦𝘴 𝘢𝘳𝘦 𝘣𝘦𝘤𝘰𝘮𝘪𝘯𝘨 𝘢 𝘴𝘮𝘢𝘭𝘭 𝘱𝘰𝘰𝘭 𝘪𝘵 𝘤𝘢𝘯 𝘱𝘪𝘤𝘬 𝘧𝘳𝘰𝘮. learn.chatgpt.com/docs/dots/… Links in reply 👇

𝗧𝗵𝗲 𝗔𝗜 𝗔𝗴𝗲𝗻𝘁 𝗦𝗮𝗻𝗱𝗯𝗼𝘅 𝗠𝗮𝗽: 𝗖𝗹𝗼𝘂𝗱 𝘃𝘀 𝗢𝗽𝗲𝗻 𝗦𝗼𝘂𝗿𝗰𝗲, 𝗟𝗮𝗽𝘁𝗼𝗽 𝘃𝘀 𝗦𝗲𝗿𝘃𝗲𝗿 . And where Ravit fits Agents don't always follow the rules. Give one a goal and it may take a road you never wanted, through your files, your tokens or your live data. The best-known case: in July 2025 the founder of SaaStr said Replit's agent 𝗱𝗲𝗹𝗲𝘁𝗲𝗱 𝗵𝗶𝘀 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲 during a code freeze. Replit's CEO apologised. That's why sandboxes got popular. A sandbox is a sealed-off computer for the agent, so a mistake stays inside it. Here's how I'd draw the map. 𝗟𝗲𝘃𝗲𝗹 𝟭: 𝗿𝗲𝗻𝘁 𝗶𝘁 𝗼𝗿 𝗿𝘂𝗻 𝗶𝘁 • 𝗖𝗹𝗼𝘂𝗱: AWS, Azure, Cloudflare, E2B, Daytona, Modal, Vercel and many more. One public benchmark tracks 𝟯𝟬 𝗽𝗿𝗼𝘃𝗶𝗱𝗲𝗿𝘀, and prices vary widely • 𝗢𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲: you run it yourself, on a laptop or on your own servers. A cheap server is enough to start (Hetzner's begin at about €5.49 a month) The rest of this post is about open source. 𝗟𝗲𝘃𝗲𝗹 𝟮: 𝘁𝘄𝗼 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀 𝗦𝗮𝗳𝗲𝘁𝘆 𝗼𝗻 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗺𝗮𝗰𝗵𝗶𝗻𝗲: sandbox-mcp and devvm. Both ask "𝘸𝘩𝘢𝘵 𝘤𝘢𝘯 𝘵𝘩𝘦 𝘢𝘨𝘦𝘯𝘵 𝘳𝘦𝘢𝘤𝘩 𝘪𝘧 𝘪𝘵 𝘮𝘪𝘴𝘣𝘦𝘩𝘢𝘷𝘦𝘴?" • 𝘀𝗮𝗻𝗱𝗯𝗼𝘅-𝗺𝗰𝗽 gives it a clean Windows PC and throws it away afterwards • 𝗱𝗲𝘃𝘃𝗺 gives it a Linux VM that sees only one folder, with placeholders in place of your real tokens 𝗖𝗼𝘀𝘁 𝗮𝘁 𝘀𝗰𝗮𝗹𝗲 𝗼𝗻 𝘀𝗲𝗿𝘃𝗲𝗿𝘀: Sandstorm and Rivet. Both ask "𝘩𝘰𝘸 𝘥𝘰 𝘐 𝘳𝘶𝘯 𝘮𝘢𝘯𝘺 𝘢𝘨𝘦𝘯𝘵𝘴 𝘤𝘩𝘦𝘢𝘱𝘭𝘺 𝘸𝘩𝘦𝘯 𝘵𝘩𝘦𝘺 𝘮𝘰𝘴𝘵𝘭𝘺 𝘸𝘢𝘪𝘵?" • 𝗦𝗮𝗻𝗱𝘀𝘁𝗼𝗿𝗺 keeps the sandbox and puts it to sleep • 𝗥𝗶𝘃𝗲𝘁 takes the agent's memory and loop out of the sandbox, so there is far less to keep alive Why split it like this? The first two are for one developer while building. The other two are for running agents for many users. 𝗔 𝗰𝗹𝗼𝘀𝗲𝗿 𝗹𝗼𝗼𝗸 𝗮𝘁 𝗥𝗶𝘃𝗲𝘁 Rivet isn't a sandbox at all. It gives each agent a small Actor with its own SQLite database and a live connection, and many Actors share one Node.js process. The README claims 𝟳𝟮 𝗞𝗕 each and a 𝟰.𝟵𝟵 𝗺𝘀 cold start at p99. Its view, in its own words: "𝘈𝘯 𝘢𝘨𝘦𝘯𝘵 𝘯𝘦𝘦𝘥𝘴 𝘢 𝘴𝘢𝘯𝘥𝘣𝘰𝘹 𝘰𝘯𝘭𝘺 𝘵𝘰 𝘳𝘶𝘯 𝘤𝘰𝘥𝘦 𝘪𝘵 𝘸𝘳𝘪𝘵𝘦𝘴." Agents that only call tools need none. Agents that write code get Rivet's lighter agentOS, or call a full sandbox as a tool. 𝗛𝗼𝘄 𝘁𝗵𝗲𝘆 𝗳𝗶𝘁 𝘁𝗼𝗴𝗲𝘁𝗵𝗲𝗿 They stack more than they compete. • 𝗪𝗵𝗶𝗹𝗲 𝘆𝗼𝘂 𝗯𝘂𝗶𝗹𝗱: devvm keeps your coding agent off your real machine • 𝗪𝗵𝗶𝗹𝗲 𝘆𝗼𝘂 𝘁𝗲𝘀𝘁 𝗮 𝗪𝗶𝗻𝗱𝗼𝘄𝘀 𝗮𝗽𝗽: sandbox-mcp checks it on a clean PC • 𝗪𝗵𝗲𝗻 𝘆𝗼𝘂 𝘀𝗵𝗶𝗽 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝘁𝗼 𝗺𝗮𝗻𝘆 𝘂𝘀𝗲𝗿𝘀: Rivet holds each user's agent, its history and its live connection • 𝗪𝗵𝗲𝗻 𝘁𝗵𝗮𝘁 𝗮𝗴𝗲𝗻𝘁 𝗺𝘂𝘀𝘁 𝗿𝘂𝗻 𝗰𝗼𝗱𝗲: it calls a sandbox, and Sandstorm keeps those sandboxes cheap while idle 𝗪𝗵𝗶𝗰𝗵 𝗼𝗻𝗲 𝘀𝘂𝗶𝘁𝘀 𝘆𝗼𝘂 • 𝗨𝘀𝗲𝗿𝘀: running a coding agent on your own laptop? Start with devvm (Mac, Linux) or sandbox-mcp (Windows Pro) • 𝗕𝘂𝗶𝗹𝗱𝗲𝗿𝘀: shipping an agent to many people? Look at Rivet for its memory, and add a sandbox only if it runs code • 𝗢𝗿𝗴𝗮𝗻𝗶𝘀𝗮𝘁𝗶𝗼𝗻𝘀: want it all on your own servers? All four are open source, and Sandstorm is the one built to pack many sandboxes onto small machines 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗹𝗶𝗺𝗶𝘁𝘀, 𝗼𝗻𝗲 𝗲𝗮𝗰𝗵 • 𝘀𝗮𝗻𝗱𝗯𝗼𝘅-𝗺𝗰𝗽: one sandbox at a time, and no Windows Home • 𝗱𝗲𝘃𝘃𝗺: no Windows, and code in the VM can still use your GitHub token through the proxy • 𝗦𝗮𝗻𝗱𝘀𝘁𝗼𝗿𝗺: if 10 or more sandboxes wake at once on a small server, 𝗼𝗻𝗹𝘆 𝗮𝗯𝗼𝘂𝘁 𝟲 𝗴𝗲𝘁 𝗮𝗻𝘀𝘄𝗲𝗿𝗲𝗱 • 𝗥𝗶𝘃𝗲𝘁: TypeScript and Rust first, and no multi-region when self-hosting yet For all four, 𝘁𝗵𝗲 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀 𝗰𝗼𝗺𝗲 𝗳𝗿𝗼𝗺 𝘁𝗵𝗲 𝗺𝗮𝗸𝗲𝗿𝘀 𝘁𝗵𝗲𝗺𝘀𝗲𝗹𝘃𝗲𝘀. 𝘗𝘪𝘤𝘬 𝘣𝘺 𝘵𝘩𝘦 𝘲𝘶𝘦𝘴𝘵𝘪𝘰𝘯 𝘺𝘰𝘶'𝘳𝘦 𝘢𝘴𝘬𝘪𝘯𝘨. 𝘐𝘴 𝘪𝘵 "𝘸𝘩𝘢𝘵 𝘤𝘢𝘯 𝘪𝘵 𝘳𝘦𝘢𝘤𝘩?" 𝘰𝘳 "𝘸𝘩𝘢𝘵 𝘥𝘰𝘦𝘴 𝘪𝘵 𝘤𝘰𝘴𝘵 𝘸𝘩𝘪𝘭𝘦 𝘪𝘵 𝘸𝘢𝘪𝘵𝘴?" Links in reply 👇

𝗧𝗵𝗲 𝗕𝗿𝗼𝘄𝘀𝗲𝗿 𝗔𝗴𝗲𝗻𝘁 𝗪𝗶𝘁𝗵 𝟭𝟮𝟬𝗸 𝗦𝘁𝗮𝗿𝘀 𝗝𝘂𝘀𝘁 𝗥𝗲𝗽𝗹𝗮𝗰𝗲𝗱 𝗜𝘁𝘀 𝗘𝗻𝗴𝗶𝗻𝗲 Every company has this job. Log into a portal. Pick last month. Click export. Download the file. Do it again for the next four portals. There is no API, so a person does it. 𝗢𝗿 𝗮 𝘀𝗰𝗿𝗶𝗽𝘁 𝗱𝗼𝗲𝘀 𝗶𝘁, 𝘂𝗻𝘁𝗶𝗹 𝘁𝗵𝗲 𝗽𝗮𝗴𝗲 𝗹𝗮𝘆𝗼𝘂𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝘀. Browser Use is an open-source library built for that gap. You describe the outcome, and an AI model works through the website in a real browser. 𝗪𝗵𝗮𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝗱 𝘁𝗼𝗱𝗮𝘆 Version 1.0 came out on 10 November. The migration guide says: "1.0 𝘳𝘦𝘱𝘭𝘢𝘤𝘦𝘴 𝘵𝘩𝘦 𝘦𝘯𝘨𝘪𝘯𝘦." • 𝗕𝗲𝗳𝗼𝗿𝗲: the model picked one action per step from a fixed list • 𝗡𝗼𝘄: the model writes Python in a session that stays alive, and drives Chrome directly The category: tools that let an AI model operate a real browser, for the large part of the web that has no API. The map, by GitHub stars as of today: • 𝗯𝗿𝗼𝘄𝘀𝗲𝗿-𝘂𝘀𝗲: the open-source browser agent. About 120.6k stars. 𝟭.𝟬 𝘀𝗵𝗶𝗽𝗽𝗲𝗱 𝘁𝗼𝗱𝗮𝘆 • 𝗣𝗹𝗮𝘆𝘄𝗿𝗶𝗴𝗵𝘁: Microsoft's framework for scripted web testing and automation. About 99.8k • 𝗣𝗹𝗮𝘆𝘄𝗿𝗶𝗴𝗵𝘁 𝗠𝗖𝗣: gives an agent browser tools over MCP. About 39k • 𝗦𝘁𝗮𝗴𝗲𝗵𝗮𝗻𝗱: an SDK to extract data and interact with sites. About 26k • 𝗦𝗸𝘆𝘃𝗲𝗿𝗻: automates browser workflows with AI. About 23k, AGPL licence • 𝗕𝗿𝗼𝘄𝘀𝗲𝗿 𝗛𝗮𝗿𝗻𝗲𝘀𝘀: browser-use's own tool for coding agents. About 19k Where 1.0 fits: it stops being a menu of browser actions. The model now writes Python in a live session and drives Chrome directly The README puts it in one line: "𝘛𝘩𝘦 𝘮𝘰𝘥𝘦𝘭, 𝘢 𝘙𝘌𝘗𝘓 𝘢𝘯𝘥 𝘊𝘋𝘗. 𝘛𝘩𝘢𝘵'𝘴 𝘵𝘩𝘦 𝘸𝘩𝘰𝘭𝘦 𝘢𝘨𝘦𝘯𝘵." 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀 • 𝗧𝗵𝗿𝗲𝗲 𝘄𝗮𝘆𝘀 𝘁𝗼 𝘂𝘀𝗲 𝗶𝘁: as a Python library in your own code, as a command-line tool that Claude Code, Codex or Cursor can call, or as a hosted API where the company runs the browser for you. This is the bit which let the agent skip capacha and cloudflare bot detection mechanism. • 𝗜𝘁 𝗰𝗮𝗻 𝘂𝘀𝗲 𝘆𝗼𝘂𝗿 𝗼𝘄𝗻 𝗖𝗵𝗿𝗼𝗺𝗲: the agent connects to the Chrome window you already have open. You are already signed in to your sites there, so the agent works inside those sessions and does not have to log in again • 𝗔 𝗳𝗿𝗲𝗲 𝗿𝗼𝘂𝘁𝗲: the library is free. Run a local browser with a local model through Ollama and you pay nothing for the model or the browser 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 • 𝗔𝗻𝘆 𝗺𝗼𝗱𝗲𝗹: OpenAI, Anthropic, Google or any compatible server • 𝗧𝘆𝗽𝗲𝗱 𝗿𝗲𝘀𝘂𝗹𝘁𝘀, file uploads and downloads, custom tools • 𝗦𝗶𝘁𝗲 𝗹𝗶𝗺𝗶𝘁𝘀: allow or block domains • 𝗣𝗼𝗿𝘁𝗮𝗹-𝘁𝗼-𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲: the agent navigates and downloads, your own code checks the file 𝗧𝗵𝗲 𝗹𝗶𝗺𝗶𝘁𝘀 • The success flag is, in the docs' words, "𝘵𝘩𝘦 𝘢𝘨𝘦𝘯𝘵'𝘴 𝘰𝘸𝘯 𝘤𝘭𝘢𝘪𝘮." The built-in judge was removed in 1.0. 𝗖𝗵𝗲𝗰𝗸 𝘁𝗵𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳 • Domain limits are "𝘢 𝘨𝘶𝘢𝘳𝘥 𝘧𝘰𝘳 𝘢𝘯 𝘢𝘭𝘪𝘨𝘯𝘦𝘥 𝘮𝘰𝘥𝘦𝘭, 𝘯𝘰𝘵 𝘢 𝘴𝘢𝘯𝘥𝘣𝘰𝘹." The model's Python runs as your user • The defaults changed. 1.0 uses the company's 𝗽𝗮𝗶𝗱 𝗰𝗹𝗼𝘂𝗱 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 𝗮𝗻𝗱 𝗺𝗼𝗱𝗲𝗹 unless you choose local • The benchmark is 𝘁𝗵𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆'𝘀 𝗼𝘄𝗻 • The README says to pin the exact version, because releases can break things 𝗧𝗵𝗲 𝗻𝘂𝗺𝗯𝗲𝗿𝘀, 𝗮𝘀 𝗼𝗳 𝘁𝗼𝗱𝗮𝘆: about 𝟭𝟮𝟬,𝟲𝟬𝟬 𝘀𝘁𝗮𝗿𝘀, MIT licence, two years old. 𝘍𝘪𝘯𝘪𝘴𝘩𝘪𝘯𝘨 𝘢 𝘳𝘶𝘯 𝘪𝘴 𝘯𝘰𝘵 𝘵𝘩𝘦 𝘴𝘢𝘮𝘦 𝘢𝘴 𝘨𝘦𝘵𝘵𝘪𝘯𝘨 𝘵𝘩𝘦 𝘫𝘰𝘣 𝘥𝘰𝘯𝘦. 𝘛𝘩𝘦 𝘢𝘨𝘦𝘯𝘵 𝘤𝘢𝘯 𝘯𝘰𝘸 𝘸𝘳𝘪𝘵𝘦 𝘤𝘰𝘥𝘦. 𝘚𝘰𝘮𝘦𝘰𝘯𝘦 𝘴𝘵𝘪𝘭𝘭 𝘩𝘢𝘴 𝘵𝘰 𝘤𝘩𝘦𝘤𝘬 𝘵𝘩𝘦 𝘳𝘦𝘴𝘶𝘭𝘵. Links in reply 👇

Neo4j open sourced light embedded graph database this week. No Server needed. My first reaction: oh, another Kuzu. Kuzu did this from 2022 until Apple bought it and archived the repo. LadybugDB and FalkorDB already run without a server too. So "no server" is not the news. What is different is who it is built for. Relate assumes the main user is an AI agent, not a person. That changes a lot of small decisions, and they are worth a look. The closest thing to it LadybugDB. It is the community fork of Kuzu, it runs inside your program, it speaks Cypher, and Cognee already uses it as its default graph store. If you want to know what Relate competes with, it is that. Think of a spreadsheet file versus a company database. The database is more powerful, but the spreadsheet is the thing you actually open, work in, and email to a colleague. Relate turns a graph into that kind of file. Open it, add connections, query it, close it. • It is written in Rust and comes as a CLI (a terminal tool), a Python package you can pip install it, Node and in-browser builds (WASM), plus Java, Go, Swift and C bindings. • A graph is a file. One .rl file, a folder, or pure memory that disappears when the program ends. Every change is written to a log first (a running diary of changes, so a crash does not lose your work), with Parquet snapshots saved alongside. You can commit the file to git or email it. "Can it turn my document into a graph?" No, and this is the part most people will get wrong. Relate stores and queries a graph. It does not read text and work out the people, companies and links inside it. There is no LLM inside it. So if you have a document and want a graph from it, something else has to do the reading: • An AI agent (the intended way). Claude Code or Cursor reads the document, decides what the things and connections are, and writes them into Relate with Cypher. Later, the same agent or another one queries it. The agent builds the graph. Relate holds it. • Your own script. If the data is already structured, like a CSV or JSON, you can load it straight in with Cypher, LOAD CSV, or a file of statements. No LLM needed. • Two built-in extensions that need no LLM: one turns a folder of files into a graph, one turns git history into a graph. • Vector search works, but you bring the embeddings. Relate does not create them. That is what "built for agents" means here. The agent is the builder. Relate is the notebook it writes in. Built for a model, not a person It is not the first of its kind Kuzu did embedded Cypher. LadybugDB carries that code on under MIT. FalkorDB has an embedded mode under SSPL. CozoDB does it with Datalog instead of Cypher. Relate's difference is Neo4j's query dialect and drivers, the round trip to Aura, and the agent tooling built in from day one. How to play with it • Browser playground: see the link in replies • The exe from the releases page. The simplest option. One file, run it from a terminal. • The Python wheel from the same page. It is not on PyPI yet, so download the .whl file first, pip install it, then use it in Python with "import relate". • The curl install script. Mac and Linux only, it does not cover Windows. • Docker, if you want it as a server speaking Bolt and HTTP. That is the opposite of the embedded idea, and overkill for a trial. Links in replies 👇👇

Someone built a full desktop app to stop Claude accounts getting banned. Claude Cage runs Claude Code inside a small Linux machine on your computer and sends its traffic through your own proxy. It also sets the timezone and language to match the proxy. The main README is in Chinese, and China is not on Anthropic's supported list. The author writes that they are fed up with fingerprint browsers and the fear of bans. Its first releases came in early November 2026. It already has around 1,800 GitHub stars. To me that shows how much people outside Anthropic's supported list want Claude.. There is a Pashto README too. I speak Pashto, so I checked because Pashto is spoken in Pakistan as well as in Afghanistan but turned out Pakistan is on the list. Afghanistan is not. Using it from an unsupported country breaks Anthropic's Supported Regions Policy. The README admits its checks cannot promise your account stays safe. Where I live Claude just works, so the proxy does nothing for me. The sandbox is the useful part. The agent only sees the folders you open. Links in replies 👇👇 Repo: github.com/Lee-xeo/Claude-Ca…

Private data in Claude? Do not start with the delete button. Start here Suppose you want to delete personal data from Claude. A work file you uploaded by mistake. A screenshot with private details in it. You want it gone, so it never ends up in Anthropic's training data. So you open Settings, find the file and press the bin icon. Done. You think you know what you are doing. 𝗬𝗼𝘂 𝗺𝗶𝗴𝗵𝘁 𝗯𝗲 𝘄𝗿𝗼𝗻𝗴. Here is the full story. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗯𝗶𝗻 𝗶𝗰𝗼𝗻 𝗿𝗲𝗮𝗹𝗹𝘆 𝗱𝗼𝗲𝘀 Settings > Privacy > Uploaded files > Manage lists every file you have uploaded. Read the notice at the top of that screen: "Deleting a file removes it from storage. Text already extracted into a chat isn't removed. To fully remove it, delete the chat." So the file goes. The text Claude read from it stays in the chat. I checked Anthropic's own pages to see what happens to that text. 𝗧𝘄𝗼 𝘁𝗵𝗶𝗻𝗴𝘀 𝗱𝗼 𝘁𝗵𝗲 𝗷𝗼𝗯 𝗽𝗿𝗼𝗽𝗲𝗿𝗹𝘆. Pressing the bin icon is neither of them. 𝟭. 𝗧𝘂𝗿𝗻 𝗼𝗳𝗳 𝘁𝗵𝗲 𝘁𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝘁𝗼𝗴𝗴𝗹𝗲 On the same Privacy screen: "Help improve our AI models". With it off, Anthropic says it "will not use your previous or new chats or coding sessions for future model training." That covers the whole chat, including text from your files. One switch, every chat, old and new. With it on, training data "includes the entire related conversation, along with any content." Deleting only the file changes nothing. 𝟮. 𝗗𝗲𝗹𝗲𝘁𝗲 𝘁𝗵𝗲 𝗰𝗵𝗮𝘁, 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝘁𝗵𝗲 𝗳𝗶𝗹𝗲 Anthropic says: "if you delete a chat from your chat history, it will not be used to train future models." A deleted chat leaves your history at once and Anthropic's back-end storage within 𝟯𝟬 𝗱𝗮𝘆𝘀. Then press the bin icon too, so the stored file goes as well. 𝗧𝗵𝗲 𝗹𝗲𝘀𝘀 𝗰𝗼𝗺𝗳𝗼𝗿𝘁𝗮𝗯𝗹𝗲 𝗽𝗮𝗿𝘁𝘀 𝟭. 𝗡𝗼𝘁𝗵𝗶𝗻𝗴 𝗶𝘀 𝘂𝗻𝗱𝗼𝗻𝗲. "Your data will still be included in model training runs that are already in progress, or in models that have been trained." 𝟮. 𝗧𝗵𝗲 𝘁𝗼𝗴𝗴𝗹𝗲 𝗵𝗮𝘀 𝗲𝘅𝗰𝗲𝗽𝘁𝗶𝗼𝗻𝘀. The Privacy Policy says that even if you opt out, chats "flagged for safety review" are still used for model improvement. 𝟯. 𝗧𝗵𝘂𝗺𝗯𝘀 𝘂𝗽 𝗼𝗿 𝗱𝗼𝘄𝗻 𝘀𝗲𝗻𝗱𝘀 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗰𝗵𝗮𝘁. It is stored "for up to 5 years" and may be used to train models, whatever the toggle says. 𝟰. 𝗧𝗼𝗴𝗴𝗹𝗲 𝗼𝗻 𝗺𝗲𝗮𝗻𝘀 𝗹𝗼𝗻𝗴𝗲𝗿 𝘀𝘁𝗼𝗿𝗮𝗴𝗲. Up to 𝟱 𝘆𝗲𝗮𝗿𝘀 in training pipelines, in de-identified form. Toggle off keeps you on the 𝟯𝟬-𝗱𝗮𝘆 period. 𝟱. 𝗙𝗹𝗮𝗴𝗴𝗲𝗱 𝗰𝗵𝗮𝘁𝘀 𝗮𝗿𝗲 𝗸𝗲𝗽𝘁 𝗹𝗼𝗻𝗴𝗲𝗿. Up to 𝟮 𝘆𝗲𝗮𝗿𝘀 if a chat is flagged as breaking the Usage Policy. 𝗪𝗵𝗮𝘁 𝘁𝗵𝗲 𝗱𝗼𝗰𝘀 𝗱𝗼 𝗻𝗼𝘁 𝘀𝗮𝘆 I could not find any Anthropic help page about the Uploaded files screen. The only wording is the notice inside the app. And for people with the toggle on: the docs say a deleted chat will not train future models. They do not say in plain words whether the copy already sitting in the 5-year training pipeline is erased too. 𝗦𝗼 𝘄𝗵𝗮𝘁 𝘀𝗵𝗼𝘂𝗹𝗱 𝘆𝗼𝘂 𝗱𝗼? The toggle decides training. Deleting the chat removes storage. 𝗗𝗼 𝗯𝗼𝘁𝗵. 𝗧𝗼𝗴𝗴𝗹𝗲 𝗳𝗶𝗿𝘀𝘁. If the toggle was already off, your file text was not headed for training anyway. Deleting the chat still clears it from storage. This toggle is for Free, Pro and Max accounts. Anthropic says it does not apply to Team, Enterprise or API use. ========================================= DO THIS (where to click) ========================================= 1. Open claude.ai. Click your name, then Settings, then Privacy. 2. Under Preferences, find "Help improve our AI models". Switch it off.This is the step that covers training for all your chats, old and new. 3. On the same screen, under "Your data", click Manage next to "Uploaded files". 4. For each sensitive file, look at the Location column. If it says "1 chat", click "Open chat 1" to see which chat holds the text. 5. Delete that chat. This is what removes the extracted text. Anthropic says a deleted chat is removed from your history at once and from its back-end storage within 30 days. 6. Then press the bin icon to delete the stored file as well. 7. Do not press thumbs up or thumbs down on a chat that holds sensitive content. That sends the whole chat as feedback, kept for up to 5 years. 8. For one-off sensitive work, use an Incognito chat. Anthropic says these are not used to improve Claude even when the toggle is on. Sources and screenshots in the reply.

This is what an edge case looks like outside a slide deck. Millions of miles driven and a traffic break is still rare enough to be thin in the data. To find out what actually happened I would pull the log data and try to find out wether the vision model flagged the cars in front of”police” , and the planner ignored it Or The vision model missed it !

𝗖𝗹𝗮𝘂𝗱𝗲 𝗝𝘂𝘀𝘁 𝗚𝗼𝘁 𝗦𝗽𝗿𝗲𝗮𝗱𝘀𝗵𝗲𝗲𝘁𝘀, 𝗪𝗵𝗶𝘁𝗲𝗯𝗼𝗮𝗿𝗱𝘀 𝗔𝗻𝗱 𝗬𝗼𝘂𝗿 𝗢𝘁𝗵𝗲𝗿 𝗔𝗽𝗽𝘀 𝗜𝗻𝘀𝗶𝗱𝗲 𝗜𝘁 A normal work task goes like this. You ask an AI for a budget table. You copy it into a spreadsheet. You fix the formulas. You paste a chart into a slide. Then you open three more tabs. 𝗧𝗵𝗲 𝗔𝗜 𝗱𝗼𝗲𝘀 𝘁𝗵𝗲 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗶𝗻 𝗼𝗻𝗲 𝘄𝗶𝗻𝗱𝗼𝘄. 𝗧𝗵𝗲 𝘄𝗼𝗿𝗸 𝗹𝗶𝘃𝗲𝘀 𝗶𝗻 𝗳𝗶𝘃𝗲 𝗼𝘁𝗵𝗲𝗿𝘀. On 10 November, Anthropic shipped an update aimed at that gap. 𝗪𝗵𝗮𝘁 𝗶𝘀 𝗻𝗲𝘄 • 𝗖𝗹𝗮𝘂𝗱𝗲 𝗦𝗵𝗲𝗲𝘁𝘀 (𝗯𝗲𝘁𝗮): a spreadsheet Claude builds with you, with "𝘵𝘢𝘣𝘴, 𝘧𝘰𝘳𝘮𝘶𝘭𝘢𝘴, 𝘧𝘰𝘳𝘮𝘢𝘵𝘵𝘪𝘯𝘨, 𝘢𝘯𝘥 𝘥𝘳𝘰𝘱-𝘥𝘰𝘸𝘯𝘴". It outlines each edit so you can see what it changed • 𝗖𝗹𝗮𝘂𝗱𝗲 𝗪𝗵𝗶𝘁𝗲𝗯𝗼𝗮𝗿𝗱𝘀 (𝗯𝗲𝘁𝗮): "𝘢 𝘴𝘩𝘢𝘳𝘦𝘥 𝘣𝘰𝘢𝘳𝘥 𝘸𝘩𝘦𝘳𝘦 𝘺𝘰𝘶𝘳 𝘵𝘦𝘢𝘮 𝘣𝘳𝘢𝘪𝘯𝘴𝘵𝘰𝘳𝘮𝘴 𝘸𝘪𝘵𝘩 𝘊𝘭𝘢𝘶𝘥𝘦". It can turn a plan into a board with a timeline, owners and dates • 𝗖𝗹𝗮𝘂𝗱𝗲 𝗗𝗮𝘁𝗮: the old Claude Dashboards, now out of beta. It connects to a data warehouse or CRM and builds dashboards and written data stories • 𝗔𝗽𝗽𝘀 𝗶𝗻𝘀𝗶𝗱𝗲 𝗖𝗹𝗮𝘂𝗱𝗲: partner apps such as Adobe, Atlassian, Box, HubSpot and Notion can open beside the chat or in full screen 𝗙𝗼𝗿 𝘂𝘀𝗲𝗿𝘀 • Sheets, Whiteboards and Data are on 𝗮𝗹𝗹 𝗽𝗮𝗶𝗱 𝗽𝗹𝗮𝗻𝘀, on web, desktop and mobile • A whiteboard can be exported to a doc or a deck, or moved to Lucid or Miro • For Confluence and Notion you get "𝘵𝘩𝘦 𝘧𝘶𝘭𝘭 𝘦𝘥𝘪𝘵𝘰𝘳, 𝘴𝘪𝘨𝘯𝘦𝘥 𝘪𝘯𝘵𝘰 𝘺𝘰𝘶𝘳 𝘰𝘸𝘯 𝘢𝘤𝘤𝘰𝘶𝘯𝘵" • Claude asks for approval before it changes a page, row or file you have open 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 • The embedded apps run on 𝗠𝗖𝗣 𝗔𝗽𝗽𝘀, an open extension to the Model Context Protocol • Anthropic says an app built this way "𝘪𝘴𝘯'𝘵 𝘭𝘪𝘮𝘪𝘵𝘦𝘥 𝘵𝘰 𝘊𝘭𝘢𝘶𝘥𝘦" • The developer features are in 𝗽𝘂𝗯𝗹𝗶𝗰 𝗽𝗿𝗲𝘃𝗶𝗲𝘄 • Anthropic's marketplace lists "𝘮𝘰𝘳𝘦 𝘵𝘩𝘢𝘯 9,000 𝘤𝘰𝘯𝘯𝘦𝘤𝘵𝘰𝘳𝘴 𝘢𝘯𝘥 𝘱𝘭𝘶𝘨𝘪𝘯𝘴" 𝗧𝗵𝗲 𝗹𝗶𝗺𝗶𝘁𝘀 • "𝗢𝗳𝗳𝗶𝗰𝗲 𝘀𝘂𝗶𝘁𝗲" 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗲𝘀𝘀'𝘀 𝗽𝗵𝗿𝗮𝘀𝗲. Anthropic's own posts never use it • Sheets is basic. Anthropic writes: "𝘍𝘰𝘳 𝘯𝘰𝘸, 𝘴𝘩𝘦𝘦𝘵𝘴 𝘩𝘢𝘷𝘦 𝘴𝘪𝘮𝘱𝘭𝘦 𝘴𝘱𝘳𝘦𝘢𝘥𝘴𝘩𝘦𝘦𝘵 𝘧𝘶𝘯𝘤𝘵𝘪𝘰𝘯𝘢𝘭𝘪𝘵𝘺". For more, you export to .xlsx or use Claude for Excel • Sheets and Whiteboards are 𝗯𝗲𝘁𝗮. On Enterprise plans they are 𝗼𝗳𝗳 𝗯𝘆 𝗱𝗲𝗳𝗮𝘂𝗹𝘁 • Claude Meetings, a note-taker for calls, is only "𝘤𝘰𝘮𝘪𝘯𝘨 𝘴𝘰𝘰𝘯" • The time-saving numbers in the launch post come from 𝗔𝗻𝘁𝗵𝗿𝗼𝗽𝗶𝗰 𝗮𝗻𝗱 𝗶𝘁𝘀 𝗰𝘂𝘀𝘁𝗼𝗺𝗲𝗿𝘀. No independent test exists yet 𝘛𝘩𝘦 𝘤𝘩𝘢𝘵 𝘸𝘪𝘯𝘥𝘰𝘸 𝘸𝘢𝘴 𝘯𝘦𝘷𝘦𝘳 𝘵𝘩𝘦 𝘱𝘳𝘰𝘥𝘶𝘤𝘵. 𝘛𝘩𝘦 𝘱𝘭𝘢𝘤𝘦 𝘸𝘩𝘦𝘳𝘦 𝘵𝘩𝘦 𝘸𝘰𝘳𝘬 𝘨𝘦𝘵𝘴 𝘴𝘢𝘷𝘦𝘥 𝘪𝘴. Links in reply 👇

Qwen4-Max Preview is now LIVE in Qwen Chat list But Qwen4 has been “coming” for days. At this rate, the announcement is taking longer than the launch. 👀 Qwen4-Max Preview is live on Alibaba’s paid Token Plan . The big promise? Deeper knowledge (better understanding of complex topics), stronger coding and smarter agents. But no public benchmarks. No confirmed full-release date. The teasing continues…

“The Agent Did It” May Soon Be No Defence In Australia bbc.co.uk/news/articles/cmp3…