AI finding and fixing OS vulnerabilities, model quantisation making laptops viable, implantable cells that could manage diabetes autonomously, and a rat who cleared 20 football pitches of landmines. [DVYio Newsletter](https://newsletter.dvy.io/visit/pb2znpTZ) Tech [Project Glasswing: Securing critical software for the AI era alongside a fractal, cellular-like wing structure.](https://newsletter.dvy.io/visit/2zGyU8zy) [Anthropic built an AI too dangerous to release. Now it's using it to patch the internet.](https://newsletter.dvy.io/visit/2zGyU8zy) Most software has bugs. That's not controversial. What's changed is who (or what) can find them. Anthropic has built a model called Claude Mythos Preview that's apparently good at finding and exploiting software vulnerabilities. Not "pretty good for an AI" good. Good enough to beat almost every human security researcher at the same task. It's already turned up thousands of high-severity flaws across every major operating system and browser. That's the alarming bit. The slightly less alarming bit is that Anthropic is trying to use it defensively before someone else uses it offensively. Project Glasswing is the attempt: a coalition of tech giants (Amazon, Apple, Google, Microsoft, and others) pooling access to Mythos Preview so it can scan the software that critical infrastructure actually runs on (banking systems, hospitals, power grids, government services, etc.) The kind of stuff where a serious vulnerability is the most dangerous. Anthropic is putting $100 million in usage credits behind it, plus $4 million to open-source security projects. The honest framing in their announcement is that this is a race. AI capabilities that can find vulnerabilities will spread, and not everyone with access will use them carefully. The question is whether defenders can get ahead of that before it becomes a problem. [A simplified neural network diagram showing a 2.0 input node branching into weight and activation nodes to output 2.5.](https://newsletter.dvy.io/visit/pi4qKV7B) [How shrinking AI models 4× makes them laptop-friendly](https://newsletter.dvy.io/visit/pi4qKV7B) Your laptop probably can't run a cutting-edge AI model. The numbers are absurd: one recent 80-billion parameter model weighs in at 159GB. That's roughly how much RAM you'd need just to load it, before doing anything useful. (A MacBook Air might have 16GB RAM — an order of magnitude lower.) Frontier models are rumoured to top a trillion parameters. That's territory most data centres can't handle, let alone a MacBook. Quantisation is how you shrink that down to something manageable. The idea starts with how numbers are stored. AI models are full of decimal numbers (like 0.52847361...) representing learned patterns. By default, each one gets stored with high precision (lots of decimal places), which takes up a lot of space. Quantisation rounds those numbers off to something coarser, trading a little accuracy for a lot of efficiency. Think of it like the difference between storing a photo as a raw file versus a compressed JPEG. You lose some quality, but usually not enough to notice in practice. Done well, quantisation can make a model 4× smaller and 2× faster while only degrading accuracy by 5–10%. That's the difference between a model that needs a server rack and one that runs on your laptop. The linked post walks through exactly how floating point precision works, where the size actually comes from, and how you measure what you've lost after compressing. [A computer monitor shows the About This Mac window for Mac OS X 10.0.4 with a white Nintendo Wii console sitting beside it.](https://newsletter.dvy.io/visit/c5mSusLi) [Mac OS X on a Nintendo Wii. Someone actually did it.](https://newsletter.dvy.io/visit/c5mSusLi) The Nintendo Wii came out in 2006. Mac OS X Cheetah came out in 2001. Nobody looked at those two facts and thought "yes, obviously these belong together". Except Bryan Keller, apparently. The hook here is that the Wii uses a PowerPC processor, the same chip family that Apple used in its Macs before switching to Intel. So while the Wii looks nothing like a Mac, the CPU inside speaks the same language. That's the crack in the door Keller needed. Getting Mac OS X to actually boot meant writing a custom bootloader (the bit of code that wakes up the hardware and hands control to the operating system), patching the kernel (the OS's core), and building drivers from scratch for the Wii's quirky hardware, including its split RAM setup and custom video output. None of this existed. He had to build it. Worth noting: a 2021 Reddit comment confidently declared there was "a zero percent chance of this ever happening." Keller included that quote in his writeup. The instructions are on his GitHub if you want to try it yourself. [A hand holds a cylindrical off-white portable fan with a star-shaped rose gold nozzle and a perforated intake base.](https://newsletter.dvy.io/visit/HbujdZPW) [Dyson's first portable fan is basically a 65,000 RPM jet engine you wear round your neck](https://newsletter.dvy.io/visit/HbujdZPW) The HushJet™ Mini Cool Fan weighs 212g (roughly the same as a can of Coke, two-thirds empty), runs for 6 hours on a charge, and pushes air at up to 25m/s. (For context, a light breeze is around 3m/s; this is closer to standing in front of an open car window at 55mph.) The motor spins at up to 65,000 RPM to get there. What's interesting is the nozzle design. Most portable fans are loud because high-speed motors produce high-frequency whine. Dyson says the HushJet™ nozzle actively shapes the airflow to cut those frequencies, landing at 68–72.5 dBA depending on the setting. That's roughly conversation volume, which is meaningfully quieter than most desk fans at full tilt. It works handheld, propped on a desk, or worn round your neck via the included Neck Dock. Optional accessories let you clip it to a pram or bag. Comes in three colour combinations (Ink/Cobalt, Carnelian/Sky, Stone/Blush) and charges over USB-C. [The NASA Orion spacecraft in the foreground, with the moon's cratered surface and a distant Earth crescent behind it.](https://newsletter.dvy.io/visit/iQT8jCcP) [A thin crescent Earth shines against a deep black void as it emerges from behind the curved silhouette of the Moon.](https://newsletter.dvy.io/visit/iQT8jCcP) [Craters of various sizes cover the lunar far side, with deep shadows stretching across a rugged, monochromatic landscape.](https://newsletter.dvy.io/visit/iQT8jCcP) [The Moon silhouetted against the Sun's bright corona during a solar eclipse, seen from the Orion spacecraft.](https://newsletter.dvy.io/visit/iQT8jCcP) [A high-contrast monochrome view showing the Moon's rugged far side, dense with impact craters and textured ridges.](https://newsletter.dvy.io/visit/iQT8jCcP) [A crescent Earth hangs in the black void of space beside the grey, cratered curve of the lunar far side.](https://newsletter.dvy.io/visit/iQT8jCcP) [Artemis II Lunar Flyby](https://newsletter.dvy.io/visit/iQT8jCcP) Poll Free seat on a loop around the Moon. Do you take it? Yes, without hesitation Yes, but I'd be terrified Maybe, if the safety record was solid No, too risky No, space doesn't interest me Last Week's Poll: How do you sleep at night? How do you sleep at night? One long stretch 32% Two separate blocks 25% Several short naps 1% Completely unpredictably 42% This Week I Learnt [A front view of the ornate, hand-carved Resolute desk in the White House, covered in books, papers, and inkwells.](https://newsletter.dvy.io/visit/htvjc7qS) [The Oval Office "Resolute" desk was built from the timbers of HMS Resolute, a British ship abandoned in the Arctic in 1854](https://newsletter.dvy.io/visit/htvjc7qS) [In 1845, Britain sent HMS Resolute into the Arctic to search for the lost Franklin Expedition (a previous mission that had vanished trying to find a route through the ice). Nine years later, with the ship stuck fast in the ice and no rescue in sight, the crew abandoned her and walked out. A year after that, an American whaling ship found her drifting, still largely intact, hundreds of miles from where she'd been left. The Americans didn't salvage her for parts. They repaired her, sailed her back across the Atlantic, and handed her back to the Royal Navy as a goodwill gesture. No strings attached. Queen Victoria was apparently so moved by this that when HMS Resolute was finally decommissioned in 1879, she had the ship broken up and three desks built from the oak timbers. One of those 1,300-pound desks was sent to President Rutherford B. Hayes in 1880. So that's the origin: a ship built to find a lost expedition, abandoned in the Arctic, rescued by Americans, dismantled, and turned into furniture. Each step takes it further from what it started as. Then it had its own quiet life inside the White House. Used in the President's study for decades, then mostly forgotten in a broadcast room during the Eisenhower years. Jacqueline Kennedy rediscovered it in 1961, moved it to the Oval Office, and had a small plinth added so JFK could sit behind it without his back problems flaring. It's been there, on and off, for almost every president since.](https://newsletter.dvy.io/visit/htvjc7qS) Science [Illustration comparing the size of a large Microbubble (1-10 micrometres) to a smaller Liposome and tiny Nanoparticle.](https://newsletter.dvy.io/visit/rC6obfxT) [Gas-filled bubbles that pop on command could finally get drugs into the brain](https://newsletter.dvy.io/visit/rC6obfxT) Most drugs are bad at their job. Not because they don't work chemically, but because they never actually get where they need to go. Less than 1% of an injected cancer drug typically reaches the tumour. The rest gets intercepted, broken down, or diluted somewhere between the needle and the target. The brain is even harder to crack. It has a dedicated security system called the blood-brain barrier: a tight mesh of cells lining the brain's blood vessels that blocks almost everything from getting in. That's mostly a good thing (it keeps out pathogens and toxins), but it also blocks the vast majority of drugs, including most chemotherapy and antibody therapies. Which is part of why conditions like Alzheimer's and Parkinson's are so much harder to treat than diseases elsewhere in the body. Researchers have been trying to solve this for decades using nanoparticles and liposomes (tiny fatty pouches that carry drugs like a molecular delivery bag), but both tend to get intercepted by the liver and spleen before they reach their destination. The approach getting attention now is stranger: microbubbles. These are microscopic spheres filled with gas, smaller than a red blood cell, that can be injected into the bloodstream and guided to a specific location using focused ultrasound. When the ultrasound hits them, they vibrate and burst, punching temporary holes in nearby vessel walls just long enough for a drug to slip through. Aim the ultrasound at the right spot and you can open the blood-brain barrier on demand, in a targeted region, for a few minutes, then let it seal back up. [Circuit boards and ribbed crimson capsules arranged in a ring around a cluster of bioluminescent pancreatic islet cells.](https://newsletter.dvy.io/visit/YiU66TN8) [What if your body just... managed diabetes on its own?](https://newsletter.dvy.io/visit/YiU66TN8) Type 1 diabetes, at its core, is a supply problem. The pancreas stops producing insulin, so blood sugar goes unregulated, so you inject insulin manually, multiple times a day, for life. It works, but it's relentless. MIT researchers think they can replace that loop with a small implantable device packed with islet cells (the pancreatic cells that naturally produce insulin) that does the job automatically. In trials, the device kept working for at least three months, producing enough insulin to regulate blood sugar without any injections. Two engineering problems had to be solved to make this work. First: the immune system attacks foreign cells, so the device wraps them in a protective enclosure to shield them from rejection. Second: cells need oxygen to survive, so the device includes a built-in oxygen generator to keep them healthy. Both problems solved in one implant. It's still early, and "at least three months" isn't the same as "forever". But if the longevity can be extended and the design holds up in humans, this could be a meaningful shift in how Type 1 diabetes is managed. [Illustration of a man in a suit levitating a glowing brain over the open skull of his own clone in a wheelchair.](https://newsletter.dvy.io/visit/H7Y5XrMz) [Your spare body, grown without a brain](https://newsletter.dvy.io/visit/H7Y5XrMz) R3 Bio recently went public, announcing plans to grow nonsentient monkey "organ sacks" as a cleaner alternative to animal testing. Tim Draper is one of the backers. That part sounds odd enough already. But MIT Technology Review dug further and found the founder, John Schloendorn, had been privately pitching something much stranger: growing a version of yourself without a higher brain. Kept alive, but not conscious. A spare-parts copy of your own body, ready if you ever need a new kidney or liver. Or, in the more extreme version he's floated, a younger body you could eventually move your brain into, effectively getting a second lifespan. The concept is partly inspired by a real birth defect where children are born missing most of their cortical brain structures (the parts responsible for thought and awareness) but can still survive. Schloendorn has reportedly shown people medical scans of these children's skulls as evidence that a body can function without much of a brain at all. R3 officially denies any cloning ambitions. But the people who attended those private pitches describe something much more candid. One attendee, speaking anonymously, compared the experience to a "close encounter of the third kind" with "Dr. Strangelove." The group behind these ideas has apparently stayed quiet deliberately, worried that the public reaction would kill their funding before the science has a chance to catch up. Data [Line chart showing EU solar and wind energy rising to 30% by 2025, overtaking fossil fuels which dropped to 29%.](https://newsletter.dvy.io/visit/VkUtQhWk) [Europe's electricity grid just flipped: more solar and wind than fossil fuels for the first time](https://newsletter.dvy.io/visit/VkUtQhWk) [Thirty years ago, the EU got more than half its electricity from burning things. Last year, for the first time, it got more from the wind and sun than from coal, gas, and oil combined (just over 30% of total production, versus fossil fuels' roughly 28-30%). Interestingly, 2025 was actually a bad year for wind. Speeds were lower than usual, and gas use ticked up slightly to cover the gap from weaker hydropower. The crossover happened anyway. A decade ago, fossil fuels were producing almost three times as much electricity as solar and wind. That gap has closed faster than most people expected.](https://newsletter.dvy.io/visit/VkUtQhWk) History [A grainy, black and white photograph of the far side of the Moon with heavy static and horizontal transmission noise.](https://newsletter.dvy.io/visit/TtQg2ADD) [What the far side of the Moon actually looks like, and how we found out](https://newsletter.dvy.io/visit/TtQg2ADD) We've been staring at the Moon for all of human history. Thousands of years of naked-eye observation, then centuries of telescope time from Galileo onwards. And yet, until the late 1950s, an entire half of it remained completely invisible to us. That's because the Moon is tidally locked: the same face always points towards Earth, the same face always points away. No amount of waiting or better equipment could fix that. The far side simply never turned into view. Except, it turns out, not quite. There's a subtle wobble in the Moon's orbit called libration (think of it like a slow, rhythmic nod) that lets us peek fractionally around the edges. Over time, this means we've actually seen just over half the Moon from Earth, not exactly half. Early astronomers noticed this and tried to encode it in their maps, including Johannes Hevelius, whose 1647 selenographic map is one of the first to capture that wobble visually. This article walks through that history with original illustrations and maps, using a handful of landmarks (the maria, or dark "seas", plus specific craters like Grimaldi near the western edge) to show exactly how our picture of the Moon was assembled, piece by piece, decade by decade. It's the kind of thing that reframes how far "exploration" actually reaches. You don't always need to go somewhere to map it. [A Victorian dictionary page explaining the phrase 'Please mother open the door' including musical notation for its pitch.](https://newsletter.dvy.io/visit/rZhjyqAo) [How did Victorians actually talk when no one was writing it down?](https://newsletter.dvy.io/visit/rZhjyqAo) A snapshot of the words Londoners used in the Victorian era and then forgot. The book was written by James Redding Ware (a pseudonym) and published in 1909, just as those expressions were sliding out of use. His point was partly that slang is hyper-local: the language of Lambeth wasn't the language of Clerkenwell, even though they're a mile apart. A few examples that didn't survive: 'balmy on the crumpet' (slightly mad), 'doing the bear' (courting a woman), 'giving it mouth' (boasting). The bits that survive tell you a lot about how a city actually worked, in a way that formal records don't. [A close-up shot of the Artemis I rocket igniting on the launchpad, with bright white flames and golden smoke at the base.](https://newsletter.dvy.io/visit/Z4MTt6xs) [What it takes to photograph a rocket launch from 1,000 feet away](https://newsletter.dvy.io/visit/Z4MTt6xs) [Most photos of a rocket launch are taken from the press site, which is several kilometres away. Safe, controlled, and fine for documentation. But the shot here was taken from roughly 300 metres. To do that, you can't be there. The energy coming off the pad at ignition (heat, pressure, debris travelling outward at speed) makes that impossible. So instead, you do something that sounds almost absurd: you place your camera there days in advance, point it at where the rocket will be, and walk away. The shutter doesn't fire when you press a button. It fires when a sound trigger detects the noise of ignition. With the Space Launch System, that's actually six seconds before the solid rocket boosters light, because the four main RS-25 engines come to life first. So by the time the rocket visibly does anything dramatic, the camera has already been shooting for a few seconds. Between setup and launch, the camera just sits there. Florida sun. Rain. Humidity. Probably birds. You don't know what you've got until you retrieve the memory card afterwards. The photo exists because Erik Kuna was willing to leave an expensive camera alone in the Florida heat and not know the outcome for days.](https://newsletter.dvy.io/visit/Z4MTt6xs) [A minimalist, headless humanoid robot with a flat white torso, black mechanical legs, and long white-and-black arms.](https://newsletter.dvy.io/visit/H6HbNZKs) [A minimalist, salmon-pink humanoid robot with a rectangular torso, thin tubular limbs, and simple two-pronged hands.](https://newsletter.dvy.io/visit/H6HbNZKs) [A minimalist seafoam green humanoid robot with long, tube-like limbs and black mechanical hands against a white background.](https://newsletter.dvy.io/visit/H6HbNZKs) [Minimalist humanoid robot frame made of matte black and white interlocking tubes with simple mechanical hands and feet.](https://newsletter.dvy.io/visit/H6HbNZKs) [Close-up of a minimalist matte black robotic hand with three flat, broad fingers featuring textured diagonal grip pads.](https://newsletter.dvy.io/visit/H6HbNZKs) [A close-up of a minimalist robot foot with a black ANIMAL logo, featuring a matte grey cylindrical ankle joint.](https://newsletter.dvy.io/visit/H6HbNZKs) [Concept 5 Humanoid](https://newsletter.dvy.io/visit/H6HbNZKs) Random [A wooden statue of Magawa the rat wearing a blue harness and a gold medal, standing upright against a leafy background.](https://newsletter.dvy.io/visit/AC2oG8sV) [Magawa the rat cleared 20 football pitches of landmines. Now he has a statue.](https://newsletter.dvy.io/visit/AC2oG8sV) Cambodia is littered with landmines, mostly left over from decades of war and conflict that ended in the 1990s. Over a million people still live and work on contaminated land. Clearing it is slow, dangerous work. Enter Magawa: an African giant pouched rat trained by a Belgian charity called Apopo to sniff out explosives. Rats are actually well-suited to this. They're light enough not to trigger the mines, and their sense of smell is sharp enough to detect the chemical compounds inside them. During a five-year career starting in 2016, Magawa cleared the equivalent of 20 football pitches, found over 100 mines, and could sweep a tennis court-sized area in 20 minutes. In 2020, he became the first rat ever awarded the PDSA Gold Medal (think: the George Cross, but for animals). He retired, slowed down, and died in 2022. Now Cambodia has unveiled a stone statue of him in Siem Reap, timed to coincide with International Mine Awareness Day on 4 April. The first statue ever dedicated to a landmine-detecting rat. Other trained rats are still working. Cambodia is aiming to be mine-free by 2030. [A vintage seafoam green control room with dials, switches, and a wooden chair sitting before a wraparound console.](https://newsletter.dvy.io/visit/E8EwXTUc) [The colour that made nuclear engineers less likely to lose their minds](https://newsletter.dvy.io/visit/E8EwXTUc) Seafoam green became standard in 1940s control rooms because of one man's obsession with the psychology of colour. Faber Birren was an industrial colour theorist who, in the mid-20th century, started advising military and industrial facilities on how paint choices affected worker performance and safety. His core finding was that certain colours reduced eye strain and improved the ability to spot hazards. Seafoam green was near the top of his list: pale enough not to compete visually with the machinery and controls, cool enough to keep workers calm under pressure, and crucially, high-contrast against blood. The Manhattan Project facilities in Oak Ridge, Tennessee, adopted Birren's recommendations heavily. Walls, control panels, and safety equipment all ended up in that same distinctive green. From there, it spread across industrial and nuclear facilities throughout WWII and beyond, simply because Oak Ridge set the template that other facilities copied. The colour coding wasn't decorative. It was a system. Green surfaces made it easier to spot whether something was out of place, whether a dial was misread, whether someone was injured. In environments where a single mistake could be catastrophic, reducing visual noise wasn't an aesthetic choice. It was an engineering one. [An engraving of Prometheus chained to a rock face, looking upwards in distress as an eagle flies above him.](https://newsletter.dvy.io/visit/Lob2xGbk) [A Victorian chatbot built from scratch, by a man with a cat and a dream](https://newsletter.dvy.io/visit/Lob2xGbk) Someone with no machine learning background decided to build an AI chatbot trained entirely on Victorian literature. Not fine-tuned on top of an existing model. Built from scratch, on a laptop, with a cat asleep on their lap. The project is called "Mr. Chatterbox, or, The Modern Prometheus" (the subtitle is a nod to Mary Shelley's Frankenstein, which itself was subtitled "The Modern Prometheus"). The idea: if you feed a language model nothing but 19th-century British books, does it start to think like a Victorian? Does it pick up the cadence, the worldview, the odd preoccupations of that era? The creator used the [British Library Books dataset](https://newsletter.dvy.io/visit/Hpcay7Rz) (digitised books from 1500–1900), filtered it down to the good stuff, and ran the training on affordable cloud hardware. The main tools were Andrej Karpathy's nanochat (a stripped-back framework for training language models from scratch) and Claude Code (which handled the Python he describes as knowing "just enough of to break things, not enough to fix them"). A few years ago, training a model from scratch required serious infrastructure and serious expertise. Now it's something one person can do over a busy family weekend, with a sleeping cat and a literary obsession. That's a pretty significant shift in who gets to build these things. [Web](https://newsletter.dvy.io/visit/t4xdFZBA) · [Twitter](https://newsletter.dvy.io/visit/c2G456LM) · [LinkedIn](https://newsletter.dvy.io/visit/GBLBeVVD)