[DVYio Newsletter](https://newsletter.dvy.io/visit/Ub3VM4za) [OpenAI is shutting down Sora](https://newsletter.dvy.io/visit/ER7zB2Te) [OpenAI is shutting down Sora](https://newsletter.dvy.io/visit/ER7zB2Te) Sora is the text-to-video tool it launched with considerable fanfare in 2024. Gone: the consumer app and the API. Not affected: image generation inside ChatGPT, which carries on as normal. The numbers explain the decision pretty bluntly. Sora made an estimated $1.4m in revenue over its lifetime. ChatGPT made $1.9bn in the same period. That's not a gap, it's a different species of product. Add in moderation headaches and the resource maths didn't work. OpenAI says it's shifting that underlying technology toward robotics instead, using what it learned about generating realistic video to train physical robots. The $1 billion Disney partnership is also gone, apparently. Quietly. [Three economy seats, one flat bed. United is selling the whole row.](https://newsletter.dvy.io/visit/4aYdnw8a) [Three economy seats, one flat bed. United is selling the whole row.](https://newsletter.dvy.io/visit/4aYdnw8a) United are introducing a bookable three-seat row in Economy that converts into a lie-flat space. You get the whole row to yourself, adjustable leg rests that raise up to form a flat surface, plus a mattress pad, blanket, and two pillows. It's rolling out next year across 200+ of their 787s and 777s, with up to 12 Relax Rows per aircraft. No pricing yet, but it's pitched as a middle ground between standard Economy and Premium Economy. Extra comfort without the full fare jump. [GPT just solved an unsolved maths problem](https://newsletter.dvy.io/visit/RKwgWcKq) [GPT just solved an unsolved maths problem](https://newsletter.dvy.io/visit/RKwgWcKq) Maths has a category of unsolved problems called "open problems." They're ones where nobody, including experts who've spent years on them, knows the answer yet. One such problem involved a structure called a hypergraph (think of it like a network, but where a single connection can link more than two points at once) and a question about how large these networks could be under certain constraints. Getting that limit as high as possible is called improving the "lower bound." GPT-5.4 Pro just solved one, in conversation, with a solution the problem's author confirmed is correct and plans to publish. Will Brian, the mathematician who originally posed the problem, said the AI's approach eliminated an inefficiency in the existing construction that he'd previously thought would be too tricky to resolve. He's now interested in follow-on work the solution has opened up, and the researchers who elicited it (Kevin Barreto and Liam Price) may end up as co-authors on the resulting paper. Since then, Epoch's team tested other models on the same problem using a more structured setup. Claude Opus 4.6, Gemini 3.1 Pro, and GPT-5.4 (xhigh) all solved it too. So this isn't a one-off fluke. It looks like frontier models are now crossing into territory that was genuinely frontier mathematics not long ago. This Week I Learnt [Every US president since Nixon has either shared their name with their father or passed it to their son](https://newsletter.dvy.io/visit/tnVCmYDm) [Every US president since Nixon has either shared their name with their father or passed it to their son](https://newsletter.dvy.io/visit/tnVCmYDm) [— That stat in the title is the kind of thing that sounds made up until you actually check it. Nixon to Trump: every single president either shared a name with their father or passed their name to a son. That's not a policy trend or an ideological pattern, it's just a weird, persistent cultural tic hiding in plain sight across half a century of American history. The Wikipedia source itself is fairly dry, but poking around it surfaces some genuinely odd facts. William Henry Harrison served just 31 days before dying in office (the shortest presidency on record). FDR went the other way, serving over 12 years across four terms, a record that prompted the 22nd Amendment capping future presidents at two. And Grover Cleveland and Donald Trump are each counted twice in the official numbering, having served non-consecutive terms, which is why 45 men account for 47 presidencies. The naming thing, though. Is it a self-perpetuating class of political families? Pure coincidence? Something about the kind of household that produces a president?](https://newsletter.dvy.io/visit/tnVCmYDm) How does your life compare to pre-COVID? Completely back to normal Mostly normal, minor habits linger Still some anxiety in crowds Permanently changed much of my behaviour Last Week's Poll Results for: How much do you think about the data you share publicly on apps? Never crossed my mind 11% Occasionally, but I don't act on it 23% I've audited my settings 44% I've deleted or quit an app over it 22% [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Image](https://newsletter.dvy.io/visit/PBQLd6Tn) [Cederhusen, Stockholm](https://newsletter.dvy.io/visit/PBQLd6Tn) [— Designed by General Architecture](https://newsletter.dvy.io/visit/PBQLd6Tn) [Lasers the size of a human hair that can spot a single molecule](https://newsletter.dvy.io/visit/mpdgwq6i) [Lasers the size of a human hair that can spot a single molecule](https://newsletter.dvy.io/visit/mpdgwq6i) Imagine you had a sensor so precise it could feel a single grain of sand landing on a football pitch. That's roughly the scale of what these researchers have pulled off, except instead of sand and grass, they're detecting individual molecules landing on a tiny glass bead. Here's how it works. The bead (about the width of a human hair) traps laser light inside it, bouncing it endlessly around the inner surface like a whisper travelling around the curved walls of a cathedral dome. This is called a "whispering gallery mode," and it's incredibly sensitive: when even a single molecule lands on the surface, it disturbs the light just enough to be measured. The team added gold nanorods to the surface to amplify that disturbance further, making the signal easier to pick up. Previous sensors like this could detect small clusters of molecules. These new ones can detect individual ones, including single atomic ions (atoms that carry an electrical charge). Why does it matter? A lot of diseases leave tiny molecular traces long before symptoms appear. If you could build a "lab-on-a-chip" device using sensors like these, you might catch cancer, dementia, or a viral infection much earlier than current tests allow, potentially from a small blood sample processed in minutes rather than days. The researchers also think these sensors could help scientists study molecular processes that are currently too small and too fast to observe clearly. [Why do some conversations flow and others die? Doorknobs](https://newsletter.dvy.io/visit/ME9mVd55) [Why do some conversations flow and others die? Doorknobs](https://newsletter.dvy.io/visit/ME9mVd55) Most conversation advice is about listening better. Be present. Ask more questions. Show you're interested. Good advice, probably. But it assumes the problem is that people aren't generous enough. The research here suggests something stranger: conversations often stall because both people are being too polite. Each person is waiting for the other to take the floor, leaving plenty of goodwill but no momentum. Sociologists call this give-and-take. The alternative is take-and-take, borrowed from improv comedy, where performers are trained to actively grab the spotlight rather than wait to be handed it. The "doorknob" idea is a good way to picture it. A conversation with lots of doorknobs has plenty of places to grab hold and push: strong opinions, odd tangents, things that invite a response. A conversation with few doorknobs just... sits there, polite and flat, each person holding the door open for the other. "Taking" sounds rude, but apparently it isn't felt that way. In group conversations especially, a confident talker who holds the floor can actually lower the stakes for everyone else, making it easier to relax and jump in when the moment comes. The problem isn't takers. It's that givers sometimes don't take enough. [Scientists found a colour the human eye has never seen before](https://newsletter.dvy.io/visit/KcUcf9df) [Scientists found a colour the human eye has never seen before](https://newsletter.dvy.io/visit/KcUcf9df) Your eye has three types of colour-sensing cells (called cones), each tuned to roughly red, green, or blue light. Every colour you've ever seen is your brain mixing signals from those three. The catch: they overlap a lot, so you can't stimulate one type completely without nudging the others. That overlap is why a truly saturated, pure green, for instance, is physically impossible to see under normal conditions. Your cones always "cross-talk". A system called Oz gets around this by skipping the optics entirely. Instead of shining light into your eye the normal way, it fires precise laser microdoses directly at individual cone cells on your retina, one cell at a time, across thousands of cones simultaneously. It can target, say, only your green-sensing (M) cones, while leaving the red and blue ones completely untouched. No normal light source can do that. When researchers activated M cones in isolation, subjects reported seeing a blue-green colour they couldn't compare to anything they'd seen before. Unprecedented saturation. Outside the normal human colour gamut. They then showed that this works for full images and video, not just isolated flashes. We assume our colour experience is a fixed property of being human. But it turns out the "gamut" is just a hardware constraint, and if you go around the hardware, there's apparently more space out there. [How LA stole a river and became a world city](https://newsletter.dvy.io/visit/8Ye9mj2N) [How LA stole a river and became a world city](https://newsletter.dvy.io/visit/8Ye9mj2N) [— Water doesn't flow uphill. That one constraint shaped everything about how Los Angeles gets its water. The LA Aqueduct, finished in 1913, runs roughly 300 miles from the Sierra Nevada down to the city. No pumps for most of that journey. Instead, the whole thing is engineered as one long, precisely graded slope: canals, tunnels, pipes, and the occasional bridge, all tilted at just the right angle so gravity does the work. Get the grade wrong by even a fraction and the water either stops or starts eroding the channel. It's a remarkably delicate balancing act across wildly varied terrain. It starts at the Owens River Diversion Weir, where snowmelt coming off the Sierra is redirected into the system. From there it winds through a landscape most Angelenos never see: high desert, volcanic rock, mountain passes. At the end, in the foothills of the San Gabriel Mountains, it spills down two concrete chutes called The Cascades. When the gates opened for the first time in November 1913, the project's chief engineer William Mulholland turned to the mayor and said: "There it is, Mr. Mayor. Take it." The engineering is genuinely impressive. The politics are murkier. Getting that water meant buying up land and water rights in the Owens Valley, a farming community that had its own plans for the river. Residents there didn't exactly celebrate the aqueduct's opening. That tension never fully went away. But the water did flow, and it kept flowing, and it's roughly a third of what LA drinks today.](https://newsletter.dvy.io/visit/8Ye9mj2N) [McDonald's sells more than Wendy's and Burger King combined](https://newsletter.dvy.io/visit/Ks5HPZjX) [McDonald's sells more than Wendy's and Burger King combined](https://newsletter.dvy.io/visit/Ks5HPZjX) [— The McDonald's CEO nibbling daintily on his own burger while calling it a "product" was always going to end badly online. Burger King responded with a video of their president attacking a Whopper. Wendy's followed the next day. It was a good week for food content. But zoom out to the actual sales numbers and the "rivalry" looks a bit different. The average McDonald's restaurant in the US now generates roughly 2.4× the revenue of a Burger King, and more than a Wendy's and Burger King combined. That gap opened up around 2015 when McDonald's was having its worst sales decade in years and decided to do something about it: simplified menu, all-day breakfast, heavy investment in app loyalty and kiosks, and a big push to refranchise stores. Meanwhile Burger King and Wendy's were dealing with underinvestment and customers quietly drifting elsewhere.](https://newsletter.dvy.io/visit/Ks5HPZjX) [Image](https://newsletter.dvy.io/visit/8L2LmV9Q) [Image](https://newsletter.dvy.io/visit/8L2LmV9Q) [Image](https://newsletter.dvy.io/visit/8L2LmV9Q) [Image](https://newsletter.dvy.io/visit/8L2LmV9Q) [Beinecke Rare Book & Manuscript Library](https://newsletter.dvy.io/visit/8L2LmV9Q) [— Yale, Connecticut](https://newsletter.dvy.io/visit/8L2LmV9Q) [Broadband cables that listen for leaking pipes](https://newsletter.dvy.io/visit/4QAyn9p7) [Broadband cables that listen for leaking pipes](https://newsletter.dvy.io/visit/4QAyn9p7) Britain leaks roughly 3 billion litres of water every day through cracked and ageing pipes. Finding those leaks is slow, expensive, and mostly involves someone walking around with a listening stick hoping to get lucky. Openreach ran a three-month trial near London testing a different approach. Distributed Acoustic Sensing works by shining laser pulses down fibre-optic cables and measuring tiny distortions in the returning light. When a pipe nearby is leaking, the escaping water causes vibrations in the ground, which disturb the light signal in a detectable way. Machine learning then filters out the background noise (traffic, roadworks, someone drilling their kitchen) to pinpoint the actual leak. The trial covered 650 km of water pipes across five locations, and found over 100 leaks in 90 days. Estimated saving: 2 million litres a day, enough to supply around 10,000 people for a year. The fibre cables already run under most towns and cities. They're just not doing anything else while carrying your broadband. This reuses them as a passive sensor network at relatively low cost, without digging anything up. [How do you squash an AI model without losing what makes it smart?](https://newsletter.dvy.io/visit/rPQLPh4j) [How do you squash an AI model without losing what makes it smart?](https://newsletter.dvy.io/visit/rPQLPh4j) AI models think in numbers, specifically in long lists of numbers called vectors. A vector for the word "cat" might encode its relationship to "fur", "claws", "pet", and thousands of other concepts, each as a separate value. The more complex the concept, the longer the list. And long lists eat memory fast. To manage this, engineers use a trick called quantisation: replacing precise, expensive numbers with cruder approximations that take less space. Good enough for most purposes, much cheaper to store and search through. The catch is that traditional quantisation methods need to store little correction factors alongside the compressed data, to track how much they rounded things. Those corrections can eat back 1-2 bits per number, which doesn't sound like much until you're running a model with billions of values. Google Research's TurboQuant takes a different approach. Rather than compressing the raw vectors and then patching in corrections, it first rotates the data mathematically (spreading information more evenly across the vector) and then compresses it. That rotation means the corrections become so small and predictable that you barely need to store them at all. The result: smaller models, faster similarity searches, less memory pressure on the "cheat sheet" that large language models use to avoid recomputing things they've already seen. The practical upside is real efficiency gains in both search engines and large-scale AI systems, without meaningful accuracy loss. Whether that translates cleanly from research benchmarks to production workloads is the usual question. But the underlying idea is tidy. [How a bitter feud sent one of Britain's finest typefaces to be lost to the Thames… and then found 100 years later](https://newsletter.dvy.io/visit/r6ZzSsZp) [How a bitter feud sent one of Britain's finest typefaces to be lost to the Thames… and then found 100 years later](https://newsletter.dvy.io/visit/r6ZzSsZp) In 1916, a 75-year-old printer named TJ Cobden-Sanderson made 170 secret midnight trips to Hammersmith Bridge and quietly dropped his life's work into the Thames. Not stolen, not destroyed by fire. Deliberately dumped. The reason: to stop his business partner ever getting his hands on it. The thing he was destroying was Doves Type, a typeface (the set of lettering designs used to print books) he'd spent years developing at the Doves Press. It was considered among the finest ever cut, used to produce some of the most beautiful books in British printing history, including the Doves Bible. Cobden-Sanderson was part of the Arts and Crafts movement, a reaction against industrial mass production that valued handcraft and beauty for its own sake. The type embodied all of that. His partner, Emery Walker, had a legal claim to it. Cobden-Sanderson had other ideas. So into the river it went. Hundreds of trays of individual metal letters, swallowed by the Thames, and assumed lost forever. A century later, mudlarks (people who search riverbanks at low tide, a tradition going back centuries) and divers started pulling small pieces of it out of the silt near Hammersmith Bridge. Enough to reconstruct a digital version of the typeface, which has since been revived and is in use again. The whole story is a strange knot of obsession, principle, and spite. Whether Cobden-Sanderson was protecting something precious or just being destructive probably depends on how you feel about the partner. [Web](https://newsletter.dvy.io/visit/wgFFhAHH) · [Twitter](https://newsletter.dvy.io/visit/rgNGYWdj) · [LinkedIn](https://newsletter.dvy.io/visit/UZptmLyg)