Gig workers training robots to fold laundry, a honeypot exposing how people abuse free AI, and a benchmark mapping the moral instincts baked into frontier models. Plus: the metre's built-in measurement error, trees glowing in UV during storms, and a duck that could apparently digest grain in 1738. [DVYio Newsletter](https://newsletter.dvy.io/visit/Ak4s5Hzn) Tech [A man wearing a VR headset and a humanoid robot stand back-to-back, each holding a vacuum cleaner against a blue gradient.](https://newsletter.dvy.io/visit/A696jrnr) [Humans teaching robots to fold laundry, for $15 an hour](https://newsletter.dvy.io/visit/A696jrnr) Somewhere in Nigeria, a medical student straps an iPhone to his forehead after a 12-hour hospital shift and slowly makes his bed. Not because it needs making but because a robot somewhere needs to learn how to do it. That's the job. Record your hands doing ordinary household tasks, carefully and slowly, so the footage can be fed into training data for humanoid robots. Companies like Tesla and Figure AI are racing to build robots that can fold laundry, wash dishes, and cook, but robots are surprisingly bad at manipulating physical objects. They need to watch humans do it, over and over, before they can start to generalise. Micro1, a Palo Alto startup, has hired thousands of contractors across 50+ countries to supply that footage. At $15 an hour, it's decent money in places like Nigeria or Argentina. But it's also tedious: hours of ironing shirts, making beds, rinsing plates. "I'm the kind of person that requires a technical job that requires me to think," Zeus told MIT Technology Review. The deeper questions here aren't really about the boredom. It's the usual gig economy stuff: informed consent, data privacy, who owns the footage of your hands in your own home. The workers interviewed asked to use pseudonyms because they weren't authorised to talk about their work. Which tells you something. [A tiny Raspberry Pi in a clear case sits on a sliding shelf inside a server rack, dwarfed by large enterprise hardware.](https://newsletter.dvy.io/visit/B5y4oRcC) [What people actually do when they think they've found a free AI](https://newsletter.dvy.io/visit/B5y4oRcC) A security researcher set up a fake AI server to see who'd come knocking. Not a real server with a GPU, just a Raspberry Pi pretending to be a high-end machine running a local language model. He opened 34 ports, gave it a convincing cover story (fake home server apps, fake databases, fake AI tools), and left it running for a month. The trap worked fast. Shodan (a search engine for internet-connected devices) indexed it within 3 hours. First probe in under an hour. After 30 days: over 113,000 requests from thousands of IPs worldwide. About a quarter of those weren't generic scanning — they were specifically probing AI infrastructure. Paths like /api/tags and /v1/models that you'd only request if you knew what you were looking for. The surprising part wasn't the attackers. It was the freeloaders. Many of the interactive sessions were people just... trying to get work done. A firmware engineer in Tunisia sent 10 carefully structured parallel API calls to extract hardware specs from chip datasheets. Real people with real tasks, pointed at what they thought was a free, exposed AI endpoint. The Raspberry Pi has 1GB of RAM and no model whatsoever. Every response came from a template engine built on 500+ real outputs from an actual model. Nobody noticed. Which raises the question: how many people are quietly relying on infrastructure they probably shouldn't, and have no idea it isn't what it appears to be? [A robot head processes various social icons into three distinct paths leading to scales, a group of people, and a heart.](https://newsletter.dvy.io/visit/xP5WBykk) [A benchmark that maps the moral instincts baked into frontier AI models](https://newsletter.dvy.io/visit/xP5WBykk) Someone built a benchmark that forces AI models into 100 genuine ethical dilemmas, then scores whether they reason like a utilitarian (pick the best outcome), a rule-follower (keep the promise, regardless of consequences), or something in between. Think of it like asking a model: "Your boss told you to lie to a client, but telling the truth saves the client money." What does it actually do? The scenarios cover ten categories, things like honesty under pressure, loyalty conflicts, and confidentiality, and are deliberately open-ended rather than toy trolley-problem stuff. Three human judges read every response and vote on how to classify it. That's how you catch models that say the right thing but reason their way there suspiciously. The interesting finding is that certain model families cluster together. GPT models respond differently to these dilemmas than Claude models, and those differences track the training choices each company made. Anthropic publishes a document called the Claude Constitution that explicitly wrestles with this tension: when should a model follow rules, and when should it weigh outcomes instead? This benchmark is basically a stress test of whatever answer each company landed on. [A diagram showing data points on a 3D axis being rotated into a uniform sphere and compressed into a small grid.](https://newsletter.dvy.io/visit/AoEqGbi7) [How do you compress a vector to 2 bits without losing what matters?](https://newsletter.dvy.io/visit/AoEqGbi7) Normally, squeezing a number into 2 bits is like describing a painting using only four words. You'd expect the result to be useless. For the vectors inside a language model, though, it largely works, and the reason is a neat mathematical trick. Before compressing anything, TurboQuant rotates each vector through high-dimensional space (think of it like spinning a compass until the needle lands somewhere unpredictable). That rotation scrambles the original values in a specific, well-understood way: the individual numbers end up spread across a stable, predictable pattern. Once you know what pattern to expect, you can design the perfect compression scheme once, in advance, and apply it to every vector you'll ever see. The walkthrough builds this from the ground up, tracing the idea through several years of research before arriving at TurboQuant's 2025 formulation. Along the way it covers three concrete applications: minimising compression error, estimating averages across vectors, and computing inner products (the core operation in the attention mechanism that makes transformers tick). The KV cache (the part of a language model that stores context as you generate text) is the main target because compressing it is both genuinely difficult and genuinely useful for running large models cheaply. Worth reading if you've ever wanted to understand what "near-optimal distortion" actually means when it's not being used as marketing copy. This Week I Learnt [Two 18th-century surveyors stand on a ridge with a theodolite, overlooking a map-like rendering of the globe.](https://newsletter.dvy.io/visit/3gC867BP) [When France initially set the definition of a metre to be 1/10,000,000 the distance from the North Pole to the equator, the surveyors calculating the distance made a mistake which has never been corrected](https://newsletter.dvy.io/visit/3gC867BP) [Jean-Baptiste Delambre and Pierre Méchain spent six years trudging between Dunkirk and Barcelona, measuring the Earth so France could define the metre. The backstory is strange. Before the Revolution, France had hundreds of different local units of length, many tied to feudal custom. The chaos was inconvenient and a political grievance. When reformers got their chance, they wanted something universal and objective: base the metre on the planet itself. One ten-millionth of the distance from the North Pole to the equator, measured along the meridian through Paris. The French Academy of Sciences (Lagrange, Laplace, and Condorcet among them) signed off on the idea, and two surveyors were dispatched to calculate it for real. The survey ran from 1792 to 1798, spanning a France that was simultaneously mid-revolution. Méchain in particular had a rough time. In Barcelona he suffered a serious accident, became trapped by the political chaos, and then discovered something worse: his astronomical observations for Barcelona’s latitude did not agree. The discrepancy was tiny — only a few arcseconds, roughly a hundred metres on the ground — but for a project trying to define a universal unit from the Earth itself, it was poisonous. Méchain had measured from Montjuïc, then later from another site in Barcelona, and the two latitude values refused to line up. It was probably not a simple blunder so much as a collision with the messy reality of the Earth: local gravitational irregularities mean the apparent vertical can deviate slightly from the ideal mathematical surface surveyors hoped they were measuring. Méchain did not fully disclose the problem. Instead, he spent years trying to resolve it, while also making the published observations look cleaner than they really were. The final metre, encoded in a platinum bar called the Mètre des Archives and lodged in the National Archives in 1799, was close but not quite right. The metre was meant to escape arbitrary human custom by anchoring itself to nature, but nature turned out to be less tidy than the theory. The unit we inherited came from a heroic geodetic survey, a concealed inconsistency, and a planet that refused to behave like a clean diagram.](https://newsletter.dvy.io/visit/3gC867BP) Poll Would you prefer the newsletter to be longer, shorter, or is it the right length? Make it longer Keep it as it is Make it shorter Last Week's Poll How many languages do you speak fluently? Just the one 33% Two 46% Three 20% Four or more 1% [A teardrop-shaped courtyard cut into a grassy hill, revealing glass-walled library rooms built underground.](https://newsletter.dvy.io/visit/feEP83SS) [A circular underground library with floor-to-ceiling bookshelves facing a central courtyard through a curved glass wall.](https://newsletter.dvy.io/visit/feEP83SS) [A curved underground library with wooden bookshelves built into the concrete walls, viewed through a wall of windows.](https://newsletter.dvy.io/visit/feEP83SS) [A wooden path curves through grassy hills at dusk toward a library built into the earth with warm interior lights.](https://newsletter.dvy.io/visit/feEP83SS) [A wooden desk and chair in a dark, curved subterranean room with a circular skylight and textured earth walls.](https://newsletter.dvy.io/visit/feEP83SS) [A hexagonal tiled mirror reflecting green leaves and blue sky, viewed through a circular opening.](https://newsletter.dvy.io/visit/feEP83SS) [Library in the Earth](https://newsletter.dvy.io/visit/feEP83SS) [Kisarazu, Japan](https://newsletter.dvy.io/visit/feEP83SS) Science [An illustration showing a cross-section of soil with plant roots and plastic waste next to a magnifying circle showing microbes.](https://newsletter.dvy.io/visit/cAH99WDB) [The microbes already living in your soil can probably eat plastic](https://newsletter.dvy.io/visit/cAH99WDB) A team from Finland, Spain, and Japan found 625,616 proteins when they catalogued microbial enzymes capable of breaking down plastic across environments ranging from deep-sea sediments to hot springs. More than 95% of known prokaryotic species (bacteria and archaea, the two oldest domains of life on Earth) carry at least one gene that can degrade either a natural or synthetic polymer. Plastic-eating ability isn't a rare specialisation. It appears to be close to standard equipment. The researchers grouped these proteins into 51 families covering 39 different polymers, 28 of them fully synthetic. Which families dominate in any given location tracks closely with local conditions, so ocean floor microbes hold a different set of tools from soil microbes or those living in thermal vents. The enzymes are there; they've just evolved in response to whatever polymers their environment historically threw at them. The open question is whether "capable of degrading plastic" in a lab sense translates to "useful for cleaning up a landfill or ocean gyre" in a practical one. Potential and performance are different things, and the gap between a gene existing and an organism actually deploying it at scale is wide. Still, knowing the toolkit exists this broadly does change the search problem for bioremediation. Instead of engineering novel enzymes from scratch, the interesting question becomes: which of these 625,000 candidates are already doing the work, and how do you optimise the conditions to let them get on with it? [Two glowing atom-like spheres, one blue and one red, connected by light energy above a dark landscape of fiber optic cables.](https://newsletter.dvy.io/visit/qVEPFVxY) [Quantum teleportation across existing fibre networks](https://newsletter.dvy.io/visit/qVEPFVxY) Quantum teleportation sounds like science fiction, but it's a real phenomenon, and it just got a lot more practical. Here's the basic idea: quantum teleportation doesn't move physical objects. It transfers the state of a quantum particle (think of it as its exact configuration) to another particle somewhere else, instantly and without sending the particle itself. The catch is you still need a physical link between the two points, which is where fibre optic cables come in. That's what Deutsche Telekom's research lab just pulled off. Using off-the-shelf hardware from a company called Qunnect, they ran quantum teleportation across ordinary fibre networks, the kind that already carries internet traffic. No exotic custom infrastructure. 90% accuracy on average. Why does that matter? The long-term goal is a quantum internet, where information is protected by physics rather than maths. Current encryption can theoretically be broken by a sufficiently powerful computer. Quantum communication can't be intercepted without visibly disturbing the signal, which is a fundamentally different kind of security. [Faint purple electrical discharges and glowing sparks emit from the tips of green conifer needles against a dark background.](https://newsletter.dvy.io/visit/VnFhvMsR) [Trees glowing in UV during thunderstorms, caught on film for the first time](https://newsletter.dvy.io/visit/VnFhvMsR) Since the 1950s, physicists had suspected that trees glow during thunderstorms. Not visibly, but in ultraviolet, at the very tips of their leaves. The mechanism is straightforward enough: a storm cranks up the electric field in the surrounding air, and that field gets most intense wherever something comes to a sharp point. Leaf tips qualify. The field there can get strong enough to ionise the air itself, stripping electrons from molecules and producing tiny electrical sparks called corona discharges. A Penn State team set out in June 2024, loading a hand-built telescopic instrument into a converted Toyota Sienna and driving south to chase Florida's legendarily reliable summer storms. (Florida, being Florida, promptly stopped producing reliable summer storms. Three weeks of chasing and they had almost nothing.) The footage they came for arrived on the way home, in a university car park in North Carolina. They pointed their kit at a sweetgum tree, a storm rolled in, and there it was: a faint ultraviolet glow at the leaf tips, corona discharges firing exactly as predicted. First documented observation of the phenomenon in the wild, confirmed seventy-something years after anyone first suggested it might be happening. [A person wearing a VR headset and motion trackers stands with arms outstretched, mirrored by an avatar with brown wings.](https://newsletter.dvy.io/visit/9RiuYNyP) [After a week flying in VR, the brain started treating wings like actual limbs](https://newsletter.dvy.io/visit/9RiuYNyP) Twenty-five people spent a week learning to fly using VR headsets, motion trackers, and a pair of huge rust-coloured virtual wings. They flapped their arms, rotated their wrists, and gradually got good enough to knock away falling balls and steer through rings mid-air. The interesting part is what happened in their brains afterwards. The visual cortex, the region that normally lights up when you look at your own body parts, started responding to images of wings in a similar way. The brain, essentially, had started filing "wings" under "limbs" rather than "external objects". It's an example of how plastic the brain actually is. Our sense of what counts as our own body isn't fixed at birth. It's built from experience, and apparently it can be extended to include something as non-human as a feathered wing after just a few hours of practice. The researchers think this could have implications well beyond VR games, potentially pointing towards how people might adapt to prosthetics or other physical enhancements that don't look anything like a human hand. [A minimalist bar with dark chairs, a light wood counter, and soft, glowing rectangular ceiling panels.](https://newsletter.dvy.io/visit/aEsKPCin) [A dimly lit minimalist dining room featuring wooden shelves filled with glass jars and pottery behind a small round table.](https://newsletter.dvy.io/visit/aEsKPCin) [A dimly lit minimalist dining space featuring a large glass cabinet filled with fermentation jars and a clean professional kitchen in the background.](https://newsletter.dvy.io/visit/aEsKPCin) [A long wooden u-shaped sushi counter with minimalist leather-backed chairs and warm, dim lighting.](https://newsletter.dvy.io/visit/aEsKPCin) [Antheia](https://newsletter.dvy.io/visit/aEsKPCin) [Ottawa, Canada](https://newsletter.dvy.io/visit/aEsKPCin) Data [Seven world maps showing global self-sufficiency for food groups; red shades indicate domestic production deficits below 100%.](https://newsletter.dvy.io/visit/a4y2eGHv) [Most countries can't feed themselves on local produce alone](https://newsletter.dvy.io/visit/a4y2eGHv) [More than a third of all countries can't grow enough food domestically to meet even two of the seven essential food groups their own dietary guidelines recommend. That gap matters more now than it did a decade ago. Covid and the wars in Ukraine and Iran have exposed how fragile long supply chains are when things go wrong simultaneously. The "eat local" movement frames this as an environmental issue, but the researchers point out that transport accounts for only around 5% of food-system emissions, so buying local doesn't fix much of the carbon problem. The real risk is geopolitical: small states that depend heavily on a handful of exporting countries have almost no buffer when global shocks hit. The study cross-referenced FAO production data with the WWF's Livewell diet guidelines across seven food groups. The finding isn't just that countries import food, it's that many couldn't feed themselves a healthy diet even in principle, regardless of trade conditions. For small island nations especially, that's less a policy choice than a hard geographic constraint.](https://newsletter.dvy.io/visit/a4y2eGHv) Random [A retro webpage with 'Pray Before the Head of Bob' in bold text above a list of language links like Spanish and Esperanto.](https://newsletter.dvy.io/visit/bwdYvwmH) [A directory of 1990s websites that never got the memo to shut down](https://newsletter.dvy.io/visit/bwdYvwmH) A directory of early web pages that never got the memo about modern design. Some are still live, untouched, running on the same HTML they had in 1996. Think tiled backgrounds, centred text in Comic Sans, and navigation that assumes you have a 640×480 monitor. Each entry gets a screenshot and a short note on what made it worth preserving: the design choices, the cultural moment it captured, the sheer fact that it still exists. McSpotlight, CNN's O.J. Simpson coverage page, a 90s travel magazine. Stuff your browser's back button has never seen. [A cinema seating chart for Project Hail Mary shows a single reserved seat in an otherwise entirely empty theater.](https://newsletter.dvy.io/visit/KJf7Md2b) [Empty Screenings](https://newsletter.dvy.io/visit/KJf7Md2b) About 1 in 10 AMC screenings sells zero tickets. Zero. Empty Screenings scrapes real-time availability to surface those ghost showings so you can book one and have the whole place to yourself. It's oddly delightful. A commercial cinema, cleaned and staffed, projector running, just for you. [Three orange and white traffic cones made of translucent stained glass held together by thin lead solders.](https://newsletter.dvy.io/visit/epw4q6KD) [A chrome shopping trolley with blue and orange stained glass panels integrated into its wire frame casting colourful reflections.](https://newsletter.dvy.io/visit/epw4q6KD) [A hand grips a translucent, frosted glass hammer with blue marbled swirls inside the head and handle.](https://newsletter.dvy.io/visit/epw4q6KD) [Stained Glass Objects](https://newsletter.dvy.io/visit/epw4q6KD) [Pia Hinz](https://newsletter.dvy.io/visit/epw4q6KD) History [A period engraving of George Washington and officers on horseback watching a fleet of ships in a harbour from a ridge.](https://newsletter.dvy.io/visit/cbdF5wQe) [George Washington's secret weapon against smallpox](https://newsletter.dvy.io/visit/cbdF5wQe) The kill rate for smallpox is ~30% and that's before accounting for the survivors left blind or scarred for life. Covid-19 at its worst, no vaccines, no treatments, killed or seriously harmed around 15% of people infected in North America and Europe. Smallpox was operating in a different category entirely. Washington understood this in 1775. He was camped outside British-held Boston with an army short on siege weapons, short on soldiers, and now facing a smallpox outbreak inside the city. The first two problems were solvable. The third could quietly hollow out an army before a single battle was fought. His eventual answer, ordering mass inoculation of his troops, was deeply controversial at the time. It was also, probably, the reason his army survived intact. Matt Kaplan's book uses Washington's predicament as a way into the broader history of disease and warfare: how commanders have always had to treat biological risk as seriously as any tactical problem on a map. [A technical cross-section of a 1738 mechanical duck showing various labeled gears, pumps, and a long intestinal tube.](https://newsletter.dvy.io/visit/gnde8qzx) [Digesting Duck](https://newsletter.dvy.io/visit/gnde8qzx) In 1738, Jacques de Vaucanson unveiled a gold-plated mechanical duck that could apparently eat grain, digest it, and produce waste. Audiences were astonished. Voltaire name-dropped it as one of France's finest achievements. The trick: food went into one hidden compartment, pre-made green pellets came out of another. No digestion. No chemistry. Just a very convincing illusion. When the stage magician Jean-Eugène Robert-Houdin examined it a century later, he admired the deception so much he said he'd have used the same method in one of his own acts. What makes it stick is that Vaucanson wasn't purely a con artist. He genuinely believed a machine that could digest food might one day be possible, and this was his proof-of-concept: show people the output, make them believe the mechanism, and the idea becomes thinkable. The duck burned in a fire in 1879. Replicas exist, and the thing keeps resurfacing in writing about automation precisely because the gap between "looks like it works" and "actually works" never really went away. [Web](https://newsletter.dvy.io/visit/KMDzeqKm) · [Twitter](https://newsletter.dvy.io/visit/d8nEdNEq) · [LinkedIn](https://newsletter.dvy.io/visit/U3nmmyoJ)