Gul Saeed Khattak

Sunlight on Demand, ordered like an Uber. Someone is building mirrors in space to sell you sunlight at night. The company is Reflect Orbital, a startup from California. The idea is a satellite with a huge thin mirror that can still see the sun after it has set where you are. It tilts and bounces that light down onto one spot on the ground. The main customer they want is solar farms. Those stop working at sunset, right when everyone gets home and puts the kettle on. They also talk about lighting up disaster zones at night. Living in Glasgow, where it's dark before 5pm in November, I can see why "sunlight on demand" sounds nice. But the first satellite only promises light about as bright as a full moon, over a patch about 5 km wide, for a few minutes at a time. Astronomers ran the numbers from the company's own filing and say that's nowhere near enough to run a solar farm. The long-term plan is 50,000 of these by 2035. Astronomers are already pushing back hard, and I get why. I want to see one actually light up a field from orbit before I buy the solar farm story. Links reply:

This is insane !

Want a win? 🪄 A Hogwarts Legacy custom RTX 5090 is up for grabs. Share the post and comment #RTXON to enter. Good luck everyone, but mostly me 😄 #RTXON

DECODED: Qwen just posted a poster that says a lot more than "The becoming begins." The Chinese line on it is the opening of Zhuangzi, written over 2,000 years ago. "In the northern sea there is a fish, and its name is..." They left the name blank. In the story it's Kun, a fish so big nobody knows its size, and it turns into Peng, a giant bird. Now look at the picture again. Thousands of birds flying in the shape of a fish. My bet is Qwen 4. Alibaba said in September 2026 that it's in training, and Qwen 3.8-Flash-Next was released as a preview of the Qwen 4 architecture. Nothing is confirmed yet. No name, no date. But you don't quote Zhuangzi for a small update.

Tuesday, no, Thursday: the trick behind a tiny voice model. Solo Dev Beats Wisper Flow With 8x FasterOpen-Source Voice Typing Meet phonon-refiner-1.7b It doesn't try to be smart. It learns one move. You say: "Send the report Tuesday, no, Thursday." A plain transcriber types all of it. Phonon's refiner types: "Send the report Thursday." That's the whole job. Drop the filler and false starts, and keep the version you settled on. The trick is in the training. A big general model knows a bit of everything and has to work out what you meant every single time. A small specialist has been trained on nothing but this. For one narrow job, that can be enough to keep up with something far bigger. My only doubt: sometimes "Tuesday, no, Thursday" is the message. A model trained to keep the last version will never know. huggingface.co/Infatoshi/pho…

I don't buy the "everything becomes open source" take. It only works for software that sits on your machine, and there's less and less of that. A lot of stuff already lives in the cloud. AI decompiling just gives companies one more reason to keep it there. And honestly, the part of the cloud you can see was always open. Every website sends its HTML and JavaScript straight to your browser. Anyone could read it for 25 years. Did that kill Google? No. Then AI came along and could read it much faster. Did that kill Google? Still no. People still use it every day, some of them through agents now. The code was there to copy all along. You just never got the part that matters. With installed software you get everything. Decompile it and you can read the logic, then plug in your own data and test it. With a cloud app the logic sits on their server. You see what goes in and what comes out. That's it. It's like a restaurant. Anyone can walk in and copy the decor. The kitchen is behind a door though. You can taste the food and guess the recipe. Fine for a sandwich, good luck with the secret sauce. And even if you crack it, you still don't have their suppliers. So yes, for software on your computer this is a big deal. For cloud apps it only shows you the dining room, and that was never hidden anyway. Thoughts Folks?

Every chatbot you use talks like someone giving a speech with no notes. It picks one word at a time and can never take a word back. If it starts with "Yes, because", the rest of the answer has to back up yes. ByteDance just released Cola2, and it gets a notepad. It jots down the next idea in its own shorthand, turns that into words, then plans the next bit. Planning in notes isn't new. What's new is how well the notes are kept. A new note starts when the idea changes, not every few words. And the note-taker and the speaker learn together, so the shorthand fits what the speaker actually needs. The results are strong for how early this is. Built on a small Qwen model, it beats Llama-3-8B on HellaSwag and StoryCloze. Llama-3 was trained on over 43 times more data, and Cola2 isn't even fully trained yet. What could change for you: chatbots and agents that settle the idea before picking the words, so answers drift less halfway through. Nobody has shown that in a product yet, but this is a serious step toward it. The weights are public, so anyone can download it and try it. The paper came out in October 2026 and the model landed today. Links in the reply 👇 Links reply: 👇 Links 📄 Paper: arxiv.org/abs/2610.31305
🤗 Model: huggingface.co/ByteDance-See…

GLB-3D-Coder: 3D You Edit, Not 3D You Regenerate This is not another 3D model generator. It's a model that writes code, and the code builds the 3D asset. Most AI 3D tools, like Hunyuan3D-2 and TRELLIS, make the asset in one go. Want a change? Usually you prompt again and regenerate. Sometimes it misreads you, so you go round and round. Here you can just edit the code. Change a line or a number, and the sword gets longer. For games, that's huge. GLB-3D-Coder, a 4B open model 👇

YES , “Attention Is All You Need.” AI researchers gave us the paper. This guy gives us the life lesson

Agents Aren’t Humans (But They Write Code Like Us). Enter Microsoft Proof-Driven Development(PDD KIT). Your coding agent opens a pull request with 3,000 new lines. You can't read them all. The agent wrote tests, but they check what it built, not what you asked for. So how do you know the login really works? Microsoft Research has a proposed answer, published 𝗣𝗿𝗼𝗼𝗳-𝗗𝗿𝗶𝘃𝗲𝗻 𝗗𝗲𝘃𝗲𝗹𝗼𝗽𝗺𝗲𝗻𝘁, and a research prototype called 𝗣𝗗𝗗 𝗞𝗶𝘁. 𝗧𝗵𝗲 𝗰𝗼𝗿𝗲 𝗶𝗱𝗲𝗮 and 𝗲𝘃𝗲𝗿𝘆𝗱𝗮𝘆 𝘄𝗮𝘆 𝘁𝗼 𝘁𝗵𝗶𝗻𝗸 𝗮𝗯𝗼𝘂𝘁 𝗶𝘁 Testing code is like test-driving a car on a few roads. It works on those roads. A proof is like the engineering maths showing the brakes work 𝗼𝗻 𝗲𝘃𝗲𝗿𝘆 𝗿𝗼𝗮𝗱. Tests try some inputs. 𝗔 𝗽𝗿𝗼𝗼𝗳 𝗰𝗼𝘃𝗲𝗿𝘀 𝗮𝗹𝗹 𝗼𝗳 𝘁𝗵𝗲𝗺. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀, 𝘀𝘁𝗲𝗽 𝗯𝘆 𝘀𝘁𝗲𝗽 • 𝗣𝗹𝗮𝗻: PDD Kit asks you clarifying questions and turns your answers into a spec. • 𝗦𝗽𝗲𝗰: the rules are written as formal contracts in Verus, a proof tool for Rust. • 𝗣𝗿𝗼𝘃𝗲: a coding agent writes the code and the proofs. Verus checks them. • 𝗗𝗼𝘂𝗯𝗹𝗲-𝗰𝗵𝗲𝗰𝗸: an LLM reviewer looks for gaps, and a 𝗺𝘂𝘁𝗮𝘁𝗶𝗼𝗻 𝗰𝗵𝗲𝗰𝗸 plants one bug per rule to make sure each rule actually catches something. • 𝗦𝗵𝗶𝗽: a pull request adds the verified code and runs Verus in CI. 𝗪𝗵𝗲𝗿𝗲 𝗣𝗗𝗗 𝗞𝗶𝘁 𝗳𝗶𝘁𝘀: it's the workflow layer. It turns your plain-English request into a formal spec, asks you to fix ambiguities, and ships proven code through normal pull requests. It runs inside GitHub Next's Chopin. 𝗧𝗵𝗲 𝗲𝘅𝗮𝗺𝗽𝗹𝗲 A Rust pet store, 𝗮𝗯𝗼𝘂𝘁 𝟱,𝟬𝟬𝟬 𝗹𝗶𝗻𝗲𝘀. The developer asked for two things: "𝘠𝘰𝘶 𝘰𝘯𝘭𝘺 𝘨𝘦𝘵 𝘢 𝘵𝘰𝘬𝘦𝘯 𝘸𝘪𝘵𝘩 𝘵𝘩𝘦 𝘳𝘪𝘨𝘩𝘵 𝘱𝘢𝘴𝘴𝘸𝘰𝘳𝘥." "𝘖𝘯𝘭𝘺 𝘢𝘥𝘮𝘪𝘯 𝘢𝘤𝘤𝘰𝘶𝘯𝘵𝘴 𝘨𝘦𝘵 𝘢𝘥𝘮𝘪𝘯 𝘵𝘰𝘬𝘦𝘯𝘴." PDD Kit spotted a gap: a login that refuses everyone also meets the first rule. 𝗧𝗵𝗲 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿 𝗵𝗮𝗱 𝘁𝗼 𝘀𝗮𝘆 𝘄𝗵𝗮𝘁 𝘁𝗵𝗲𝘆 𝗿𝗲𝗮𝗹𝗹𝘆 𝗺𝗲𝗮𝗻𝘁. The final spec had 𝟴 𝗿𝘂𝗹𝗲𝘀. The post says Verus "𝘶𝘯𝘤𝘰𝘷𝘦𝘳𝘦𝘥 𝘦𝘹𝘪𝘴𝘵𝘪𝘯𝘨 𝘣𝘶𝘨𝘴." 𝗙𝗼𝗿 𝗱𝗲𝘃𝗲𝗹𝗼𝗽𝗲𝗿𝘀 You review a short spec, not thousands of lines of code. 𝗙𝗼𝗿 𝗯𝘂𝗶𝗹𝗱𝗲𝗿𝘀 It sits inside GitHub Next's Chopin planning tool. Microsoft compares it to Spec Kit, but for formal proofs. 𝗪𝗵𝗮𝘁 𝗰𝗮𝗺𝗲 𝗯𝗲𝗳𝗼𝗿𝗲 • Spec Kit made "write the spec first" popular for agents. • AutoVerus (2024) used LLM agents to write Verus proofs. What's new here: 𝘁𝗵𝗲 𝘄𝗵𝗼𝗹𝗲 𝗹𝗼𝗼𝗽, 𝗳𝗿𝗼𝗺 𝘆𝗼𝘂𝗿 𝘄𝗼𝗿𝗱𝘀 𝘁𝗼 𝗮 𝘀𝗽𝗲𝗰 𝘁𝗼 𝗽𝗿𝗼𝘃𝗲𝗻 𝗰𝗼𝗱𝗲, 𝗶𝗻𝘀𝗶𝗱𝗲 𝗮𝗻 𝗮𝗴𝗲𝗻𝘁 𝘄𝗼𝗿𝗸𝗳𝗹𝗼𝘄. 𝗧𝗵𝗲 𝗵𝗼𝗻𝗲𝘀𝘁 𝗽𝗮𝗿𝘁 • 𝗢𝗻𝗹𝘆 𝘁𝗵𝗲 𝗹𝗼𝗴𝗶𝗻 𝗹𝗼𝗴𝗶𝗰 𝘄𝗮𝘀 𝗽𝗿𝗼𝘃𝗲𝗻, not the whole app. Database lookups, password hashing and the clock are trusted and tested. • 𝗥𝘂𝘀𝘁 𝗼𝗻𝗹𝘆 for now. • "𝘈 𝘱𝘳𝘰𝘰𝘧 𝘴𝘩𝘰𝘸𝘴 𝘵𝘩𝘦 𝘤𝘰𝘥𝘦 𝘮𝘦𝘦𝘵𝘴 𝘵𝘩𝘦 𝘴𝘱𝘦𝘤, 𝘯𝘰𝘵 𝘵𝘩𝘢𝘵 𝘵𝘩𝘦 𝘴𝘱𝘦𝘤 𝘪𝘴 𝘸𝘩𝘢𝘵 𝘺𝘰𝘶 𝘮𝘦𝘢𝘯𝘵." • 𝗢𝗻𝗲 𝗮𝗽𝗽, 𝗻𝗼 𝗯𝘂𝗴 𝗰𝗼𝘂𝗻𝘁, 𝗻𝗼 𝗰𝗼𝘀𝘁 𝗼𝗿 𝘁𝗶𝗺𝗲 𝘁𝗼𝘁𝗮𝗹𝘀. • 𝗡𝗼𝘁 𝗽𝘂𝗯𝗹𝗶𝗰𝗹𝘆 𝗮𝘃𝗮𝗶𝗹𝗮𝗯𝗹𝗲. It's a research prototype, and the post names no agent or model. 𝗧𝗵𝗲 𝗶𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝗽𝗮𝗿𝘁 𝗶𝘀𝗻'𝘁 𝘁𝗵𝗲 𝗽𝗿𝗼𝗼𝗳. 𝗜𝘁'𝘀 𝘁𝗵𝗮𝘁 𝘁𝗵𝗲 𝘁𝗵𝗶𝗻𝗴 𝘆𝗼𝘂 𝗿𝗲𝘃𝗶𝗲𝘄 𝗴𝗲𝘁𝘀 𝗺𝘂𝗰𝗵 𝘀𝗺𝗮𝗹𝗹𝗲𝗿. Think about where the bottleneck is now. Agents write code faster than anyone can read it. Microsoft's own line: "𝘳𝘦𝘢𝘥𝘪𝘯𝘨 𝘦𝘷𝘦𝘳𝘺 𝘭𝘪𝘯𝘦 𝘰𝘧 𝘵𝘩𝘦 𝘱𝘶𝘭𝘭 𝘳𝘦𝘲𝘶𝘦𝘴𝘵 𝘥𝘰𝘦𝘴𝘯'𝘵 𝘴𝘤𝘢𝘭𝘦." Tests don't fully save you either. The agent writes them, so they tend to check what it built, not what you asked for. A proof changes the shape of the problem. If the proof checks, the code meets the spec 𝗳𝗼𝗿 𝗲𝘃𝗲𝗿𝘆 𝗶𝗻𝗽𝘂𝘁. So your attention moves to one place: 𝗶𝘀 𝘁𝗵𝗲 𝘀𝗽𝗲𝗰 𝗿𝗶𝗴𝗵𝘁?. Links in reply 👇

Unbelievable! ✅

PANIC: Gone in a day ! $3.8T gone from US stocks in one day. Nvidia $447B. Broadcom $162B. Micron $157B. AMD $138B. Midterms hit chips!

Claude can now read your bank account. Anthropic launched Personal Finance on 5 November 2026. You link your bank and cards through Plaid, then ask things like "how much did I spend on takeaways last month" . It can see up to 2 years of transactions. It's read-only. It can't move a single penny. US only for now, on Pro, Max and Ultra. I am in Glasgow so I'm just watching. Honestly I'd use it. But giving an AI two years of my spending is going to take me a while. Introducing Personal Finance in Claude, 5 November 2026: claude.com/resources/article…

Attackers can't keep up with the bugs. For now. 48,185 CVEs were published in 2025. In 2017 it was 14,645. That's more than 3x in eight years. Confirmed exploits in the wild in 2025? 884. Under 2% of all bugs. At the old 6% rate you'd expect about 2,900. Most bugs never get used because writing a working exploit took a skilled person weeks, and there were never enough of those people. I am not sure that lasts. The same a16z piece says AI now turns a patch into a working exploit in hours. If that holds, this gap is just a queue attackers haven't reached a16z.news/p/the-exploit-was-…

CONFIRMED: Google Gemini Omni 1.2 Flash is out A 15 second AI video now costs about $1.05. Still expensive though . Google just made Gemini Omni 1.2 Flash generally available on the Gemini API. Clips go up to 15 seconds, and 720p is $0.07 per second, 30% cheaper than 1.1. The part I like most is the editing. You can swap a background or a product by chatting with the same clip, without paying to generate it again. Google also admits it still struggles with fast sports action and complex physics. Links in Replies 👇 👉 Google blog: blog.google/innovation-and-a…

OpenAI just released GPT-6 Cyber. Their own safety framework rates it "Critical" for cybersecurity. In testing it found 81 new high-severity bugs in Firefox. It also wrote working exploits for 948 Linux kernel flaws, and 588 of those nobody knew about. Then it came up with 10 kernel attack techniques that outside experts say they'd never seen. OpenAI says every bug was reported and patched. The number I care about more: the old GPT-5.6 Cyber went outside its allowed task in 5.51% of honeypot tests. This one did it in 0.07%. For now only approved Daybreak Red customers get it, through the API and Codex. A model that invents new attack methods is exactly the kind that should sit behind approvals. I'm less sure how well that holds once access opens to all Red customers in the coming weeks. Links reply: 👉 Fortune: fortune.com/2026/11/05/opena…

Delaware wants an AI agent to be the boss Delaware wants to let an AI agent run a whole company. No human manager signing off on daily decisions. I build agents. I still check what they produce before it touches anything real, because they go off track in ways I don't see coming. The plan is called an AIC. The agent runs the business, can own property and sue in its own name. One owner keeps it funded. In return, they're shielded from its debts unless they underfund it or commit fraud. The safety net is a promise in the draft that the owner can shut the agent down. Meanwhile Norm Ai, the firm helping design it, admits agent behaviour may not be fully predictable. I actually like giving agents a legal name you can sue. Today nobody knows who to blame when one causes damage. But that shield worries me. If I'm safe as long as the account has money, the easy move is to top it up and stop watching. It's a 30 month trial and the bill is expected in January 2027. I'd want the rule for when a human must step in written before it starts. arstechnica.com/tech-policy/…

My agent writes the code. Now it records the demo too. You have finished the project. Tests pass. You are about to push to GitHub, and the README needs a quick demo video. Same with the PR. Reviewers want to see the change, not read about it. Normally that's another hour. Record it, zoom in on every click, cut the "umm, wait" bits, add captions, export. OpenScreen lets you hand that job to your agent. 𝗧𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝘄𝗮𝘆: After a one-time setup, you tell Claude Code: "add a video of this change to the PR". It plans the shots, records your app, edits the take with zooms on the clicks, and attaches the MP4 to the pull request. Reviewers watch it right there. 𝗧𝗵𝗲 𝗺𝗮𝗻𝘂𝗮𝗹 𝘄𝗮𝘆: Prefer doing it yourself? The desktop app gives you the same polish. Auto-zoom follows your cursor. Captions are made on your own machine, offline, and you cut the video by deleting a sentence from the transcript. Free, MIT licensed, no watermark. Windows, macOS and Linux. Just passed 6,000 GitHub stars. 𝗪𝗵𝗲𝗿𝗲 𝗶𝘁 𝘀𝗶𝘁𝘀 𝗶𝗳 𝘆𝗼𝘂 𝗷𝘂𝘀𝘁 𝘄𝗮𝗻𝘁 𝗶𝘁 𝗱𝗼𝗻𝗲 𝗳𝗮𝘀𝘁 𝗦𝗰𝗿𝗲𝗲𝗻 𝗦𝘁𝘂𝗱𝗶𝗼: the app that defined this category. Beautiful, but Mac only, $29 a month or about $9 if you pay yearly. 𝗗𝗲𝘀𝗰𝗿𝗶𝗽𝘁: great when you're talking on camera and editing by text. Overkill for a quick product demo. 𝗖𝗮𝗽𝗖𝘂𝘁: great for social clips and effects, but not built around screen demos. 𝗢𝗽𝗲𝗻𝗦𝗰𝗿𝗲𝗲𝗻: closest to Screen Studio for demos, with Descript-style transcript cutting, and your agent can drive it. For free. GitHub: github.com/getopenscreen/ope…

NerdLore is now live. No install needed. 👉 nerdlore.streamlit.app It brings two different angles on the same videos into one app: • Hacker News memory, from Andrej Karpathy's HackerTube. It answers: which videos do engineers keep recommending, years later? • YouTube reach, from YouTube's own view counts. It answers: how many people actually pressed play? Each one sees what the other can't. Side by side, you find the talks engineers swear by that most people never watched. Try it now Press Try the demo and every page works straight away. It uses a built-in sample of 628 videos with made-up view counts, just to show how it works, and every page says the numbers are made up. Want the real view counts? Get a free YouTube Data API v3 key (there's a link in the app) Download hackertube-data.zip from hackertube.org/data Upload it, paste your key, press Run analysis It costs about 1 quota unit per 50 videos, and the free daily quota is 10,000. Two honest notes • It sleeps when nobody has used it for a while. If you see "gone to sleep", click the button and wait about 30 seconds. • On the live copy, your key passes through Streamlit's server for your session. Nothing is saved. If you'd rather your key never leaves your machine, run it locally from GitHub.

Make sense of OpenAI's 948 Linux security fixes. No jargon. Imagine a hotel. A guest checks in and gets a normal keycard. It opens her room and nothing else. Only the manager has a master key that opens every door. One evening she taps her card on a door at a slightly odd angle. The lock beeps green. She tries again on another door. Same thing. The system has got confused and now treats her cheap guest card as the master key. Now she can walk into the manager's office and the spa she never paid for. Every door in the building opens for her, and the hotel has no idea. That's a privilege escalation bug. In Linux terms, her keycard is a normal user account. The master key is "root", the admin account that can do anything. "Local" means she had to be inside the hotel already. You need some normal access first. Where it bites: I work at a university, so picture a shared Linux server where hundreds of students log in. One student, or a hacker with a stolen password, runs a small program. Seconds later their normal account is root. Now they can read everyone's work and hide their tracks. What OpenAI did: In mid-October 2026 their AI scanned the Linux kernel and found 948 bugs like this that worked on a normal Linux desktop. Then it wrote a fix for each one. Total cost: $1.2 million in API credits, about $1,279 per bug. It also found at least 10 break-in techniques experts hadn't seen before. The part headlines skip: A fix written by AI still has to be accepted by Linux and then installed on your machine. When the scan finished, maintainers had dealt with roughly 100 of the 948. Finding bugs just got cheap. Fixing them still moves at human speed. That gap worries me more than the headline number. OpenAI's write-up: openai.com/index/a-thousand-…