Together AI just raised $800M at an $8.3B valuation. I got their product team to open up the actual repo they run on: 0:00 - Slop is the new party foul 1:44 - Why individual output backfired 3:26 - Inside the shared product repo 8:57 - Who owns the context files 10:41 - Team skills vs personal skills 11:57 - Picking a harness and a model 15:17 - Feature research running live 16:57 - What stays human in the loop 19:56 - The PRD skill that interviews you 23:09 - Half a day of research in 5 min 24:25 - Their customer insights MCP 27:31 - What a good PRD looks like now 32:02 - Reading the finished PRD 35:07 - Every repo in one place 40:24 - Context is a hierarchy, not a pool 41:20 - Build your own orchestrator 44:19 - Agent evals on shipped features 49:41 - Design validated every few hours 52:26 - Where PM ends, engineer begins 55:10 - What it actually cost them

A video game that cost $436 million to make now looks like security camera footage, because someone leaked an NVIDIA file that was never meant to be turned on. The file is DLSS 5. It fell out of a basketball game, NBA 2K27, before NVIDIA released it. Modders grabbed it and jammed it into Cyberpunk 2077 within days. What the file does is simple. It looks at each frame the game draws and repaints it to look like a photo. Skin gets pores. Walls get grime. Everything drifts toward real life. That is the whole problem. Cyberpunk was designed to look like a neon comic book, not real life. So an AI that only knows "make it look real" turns Night City into an LA underpass. Same buildings, same cars, all the style washed out. It also costs half your performance. A $2,000 graphics card drops from 63 to 42 frames per second. Now look at where NVIDIA chose to launch it. A basketball game where every face is a 3D scan of a real player. When the goal is already "look like a photo," an AI that pushes toward photos is pure upside. Nobody has an art style to protect. NVIDIA also knew games like Cyberpunk would break. After the first preview got called AI slop in March, they built controls for developers. Three different models. Masks so a studio can apply it to a jacket but not a face, or turn it off on characters completely. The leaked file has none of those controls. Every one of these mod videos is DLSS 5 cranked to 100% on every pixel with nobody steering. Official launch is Thursday. Modders spent a week showing the world the one version NVIDIA spent five months making sure no game would ship.

Grok Bot's App Store listing names the developer as Anysphere. That's Cursor. xAI closed the $60B Cursor acquisition this week and shipped Grok Bot alongside it. The "product launch of the year" chatter is really the first public look at what that deal bought. Claude Code's unlock was handing the model an actual computer. Grok Bot runs the same play at fleet scale. Every bot gets a persistent cloud VM with a browser, a filesystem, and a terminal. It logs into your apps, keeps working while your laptop is shut, and only surfaces for approvals. I ran a fleet for a week and wrote up the setup plus 100 jobs to hand your bots → Two features carry the product. Skills self-correct as they run, so a task you teach once keeps improving without you touching it. And bots route work to each other, which means you stop being the middleware between your own agents. That second part explains the 100x usage jumps people are reporting. Chat usage was capped by how fast you could type. A fleet on its own computers is capped by how much finished work you can approve. Different ceiling entirely. Cursor spent years learning to build agent harnesses. xAI spent $60B to skip the line.

When India accidentally wiped out its vultures, half a million people died in five years. The bird doing zoomies in this video runs the most extreme digestive system in the animal kingdom. Vulture stomach acid sits near pH 1, strong enough to destroy anthrax, rabies, and botulism in the carcasses they eat. A flock strips a cow to bone in about 45 minutes. India had 50 million of them. Free sanitation at national scale, in a country where religious custom means most cattle die natural deaths and every dead cow becomes a carcass someone has to deal with. Then the patent on diclofenac expired in 1994. Farmers started using the cheap painkiller on livestock. Trace amounts left in a carcass cause kidney failure in vultures. Within a decade, 50 million birds fell to a few thousand. The fastest collapse of any bird species since the passenger pigeon. The carcasses didn't disappear. Feral dogs took the vultures' place and brought rabies with them. Rotting remains leached into waterways. In districts where vultures once thrived, human death rates rose more than 4 percent. Economists at Chicago and Warwick put the toll at 100,000 extra deaths a year and $69 billion in annual damages. So yes, the goofy bird sprinting down the walkway is adorable. It's also a disease firewall that evolution spent millions of years building and one generic painkiller nearly erased in ten.

Cookie stuffing already sent two men to federal prison. Shawn Hogan made $28 million as eBay's #1# affiliate before the FBI raided his home. Brian Dunning, the #2#, took in $5.2 million. Both pleaded guilty to wire fraud. Now leaked Slack messages show Phoebe Gates' startup ran the same playbook, and the founders knew. Here's what Phia's extension did. During checkout it opened a background tab and inserted its own referral code, claiming commission on sales it never drove, sometimes overriding the affiliate who actually earned it. Nike and Nordstrom were among the retailers hit. When Bloomberg first reported it in July, Phia called it a bug and said the company had just learned of it. The new leaks show Gates and Kianni discussing the practice with engineers back in December. Seven months earlier. And per Bloomberg, the code was a deliberately built feature with a toggle. Bugs don't ship with on/off switches. The wildest thread runs through one man. In 2006, eBay hired a fraud researcher named Ben Edelman to run compliance checks on affiliates suspected of cookie stuffing, the cases that became the Hogan and Dunning prosecutions. In 2026, the researcher who reviewed Phia's code and concluded it was designed to claim unearned commissions? Ben Edelman. Same investigator, same scheme, 20 years apart. The eBay cases are also why the 20-year max is real. Prosecutors argued the cookies themselves, transmitted across state lines, constituted the wire fraud. That theory won, and it turned cookie stuffing into a federal crime. Hogan got 5 months. Dunning got 15. Both were first offenses.

Neuroscientists found a brain region that physically grows when you do things you don't want to do. It shrinks when you stop. It's called the anterior midcingulate cortex. It's smaller in people who struggle to delay gratification. It gets bigger when obese people diet. It's larger in elite athletes. And in superagers, people in their 80s with the memory of 50 year olds, it refuses to atrophy. Here's the detail almost everyone misses. Hard things you enjoy don't grow it. A runner who loves running gets nothing from a 10 mile run. The region only responds to resisted effort, the workout you wanted to skip, the task you kept putting off. The suck is the signal. That's also why disciplined days feel identical. The aMCC doesn't care about novelty or intensity. It responds to repeated effort against your own resistance. Habit is the wrong word for what's happening. You're building tissue. And it runs in reverse. Avoid enough hard things and the region weakens, which makes the next hard thing feel harder, which makes you avoid it more. Researchers studying aging brains found that where you sit on the apathy versus tenacity spectrum predicts how well your brain holds up over decades. Some neuroscientists call the aMCC the seat of the will to live. The days look the same because that's what the training looks like. The years look different because that's what the tissue does.

Sony's producer pulled each new Spider-Man aside before filming and lectured them not to date their co-star. All three ignored her. Amy Pascal ran the warning like a ritual. She sat down Tobey and Kirsten. Then Andrew and Emma. Then Tom and Zendaya, separately, right after casting them in 2016. Her exact words were "Don't go there. Just don't. Try not to." Her record now stands at 0 for 3. And each pair went further than the last. Tobey and Kirsten lasted about a year. Andrew and Emma made it four. Tom and Zendaya blew past engaged entirely. Tom confirmed this summer that they're already married, after AI-generated wedding photos went viral and his grandmother panicked that she hadn't been invited. He told her she had nothing to worry about, since the family was all there. The pattern was never a coincidence. A chemistry read is the final step of casting a romantic lead. Sony auditions hundreds of actors, then tests the finalists in pairs to find the two people whose attraction looks most real on camera. The winners get paid to gaze at each other for three movies across six years. The studio built a machine that selects for chemistry, then asked it to stop producing chemistry. Pascal has made peace with it. Her current position is that MJ and Peter just always fall in love, and there's nothing anyone can do about it.

Sergey Brin is the third richest person alive, worth roughly $240 billion, and he just took direct command of Gemini. He retired in 2019. He owns around 6% of a $3.9 trillion company. He could spend the next 40 years on a yacht without ever checking a price tag. Instead he's been at the office three to four days a week, in the weeds, writing code. Demis Hassabis confirmed it himself this year, "Sergey has been in the weeds, programming." The last time Brin was this hands-on, Google was running out of a garage in Menlo Park. And he fits a pattern. Zuck went founder mode on AI and reorganized all of Meta around superintelligence. Elon started xAI from scratch and built it into a frontier lab in two years. Now Brin, at 52, is personally running the model that decides whether Google wins the next 20 years. That tells you everything about the size of the prize. Smartphones, cloud, crypto, none of it pulled a retired founder back onto a product team. AI did it in under two years. A $240 billion man just gave himself a job. The AI war is fully on.

Classical music has never paid for itself. Ticket sales cover roughly a third of an American orchestra's budget, and that's the healthy ones. Since 2010, orchestras in Philadelphia, Louisville, Honolulu, and Syracuse have all filed for bankruptcy. Then gamers started buying tickets. The model was always subsidy. Beethoven lived on an annuity from three Viennese aristocrats. Wagner had a Bavarian king paying his debts. The modern version swapped princes for donors and endowments, but the arithmetic never changed. The music on stage loses money and someone rich covers the gap. Video game concerts broke that arithmetic. The Legend of Zelda's Symphony of the Goddesses sold out most of a world tour. Final Fantasy VII Rebirth sold out the Royal Albert Hall. Sonic filled London's Barbican for three straight nights. Same players, same halls. Different composer names on the program, and suddenly the box office covers the cost. The audience explains why. The average classical subscriber is near retirement age and the field loses old attendees faster than it gains young ones. The Zelda crowd shows up in their teens and twenties, some in costume. A Royal Philharmonic poll found more young people now discover orchestral music through games than through live concerts. So the annoyance is aimed at the wrong people. Koji Kondo and Nobuo Uematsu built the biggest onboarding funnel orchestral music has ever had. The 19-year-old hearing a 90-piece orchestra play Zelda tonight is the only plausible Mahler subscriber of 2046. Haydn needed a prince to fund a symphony. Zelda fills the hall at face value. In 300 years of orchestral music, this is the first time the audience paid the whole bill.

The open weights letter launched Thursday with 25 signatures. Reports put it near 50 a day later, and names keep landing. Bolt's is the one I'd flag. Eric's post below lays out the record: years of funding the OSS projects the web runs on, the OSS Pledge, open sourcing And Bolt sits at the application layer, where open weights are supply chain. What stays downloadable, inspectable, and swappable underneath decides what you can ship. The range of the list is the signal. Chip makers, clouds, labs, VCs, and a growing slice of the application layer are signing the same page, which makes it harder to wave off as labs talking their own book. Every AI product carries assumptions about what stays open underneath it. Bolt just put its assumptions in writing.

The Matrix took its name from a 1984 novel written on a manual typewriter by a man who didn't own a computer. 42 years later, that novel finally gets its own screen. William Gibson typed Neuromancer on a 1927 Hermes portable. He coined "cyberspace" before he'd ever logged onto anything. The book swept the Hugo, the Nebula, and the Philip K. Dick award in the same year, the only novel ever to take all three. Then Hollywood spent four decades strip-mining it while calling it unfilmable. The Wachowskis lifted the matrix, the jacked-in hackers, the AI pulling strings behind a corporate veil. Every razor-girl assassin in sci-fi traces back to Molly. Johnny Mnemonic, an actual Gibson adaptation, flopped so hard in 1995 that studios treated his work as radioactive for a generation. At least five directors attached and detached over the years. Chris Cunningham in 2000. Vincenzo Natali in 2010. Tim Miller at Fox in 2017. Each attempt died the same death. By the time the technology existed to film Neuromancer, audiences had watched its ideas in a dozen movies that borrowed them first. The original became unfilmable because it looked like a copy of its own copies. Now the timing loops back on itself. A story about a rogue AI maneuvering to escape its corporate constraints, written before the web existed, lands in January 2027 as the least speculative thing on television. And one detail makes the whole arc perfect. When Neuromancer royalties came in, Gibson bought his first computer, an Apple IIc. Four decades later, Apple is the company putting his book on screen.

Apple just sued OpenAI, and the wildest part is how they got caught: one candidate screenshotted confidential Apple files on his Apple work laptop hours before his OpenAI interview. Apple reads its own server logs. The recruiting pipeline generated its own evidence trail. The complaint says OpenAI's hardware chief Tang Tan, a 24-year Apple veteran, directed candidates still employed at Apple to bring "actual parts" (batteries, logic boards) to interviews for show and tell sessions. One candidate was surprised, saying he didn't even know you could take those out of the office. Apple also alleges Tan circulated an internal Apple offboarding document to coach new hires on dodging exit security checks, and that a departing engineer kept his Apple laptop, found a bug that still gave him access to Apple's cloud storage, and downloaded dozens of confidential hardware files after joining OpenAI. Then the supplier: OpenAI allegedly got one of Apple's manufacturing partners to demonstrate a proprietary metal finishing technique by letting the partner believe Apple had approved it. Over 400 former Apple employees now work at OpenAI. Apple says it flagged all of this to OpenAI in February and never got a response. Five months later, it filed. The ask reveals the strategy. Apple wants an injunction barring OpenAI from using the secrets, the return of every file, and full discovery into io, right as OpenAI preps its first device launch and an IPO. If a judge grants it, OpenAI may have to prove the device was built clean, component by component, before it ships. The device was supposed to run on the world's best hardware talent. Now its bill of materials is evidence.

Bridgewater just published numbers that should make every frontier lab nervous. The world's largest hedge fund tested Gemini, Claude, and GPT on six document filtering tasks its investors do every day. Naive prompts scored around 50%. A coin flip. Expert-written prompts pushed accuracy to 78%. Investors needed 80% before they'd trust the system in their workflow, and no frontier model cleared it. GPT 5.4 cost 43% more than 5.2 and was barely more accurate. So they fine-tuned Qwen3-235B on Tinker instead. 84.7% accuracy. 29.8% fewer mistakes than the best frontier model. At 1/14th the inference cost. The smartest part is buried in the middle of the paper. Their vendor-labeled training data was riddled with wrong labels, and expert labeling costs too much to run on everything. Their fix: train a model on the noisy dataset, then run it back over its own training data. Any example the model disagreed with got routed to senior investors, because either the example was genuinely hard or the label was wrong. The model's own confusion became a detector for bad labels. Prompting hit a ceiling for a structural reason. A prompt captures only the judgment an expert can put into words. Twenty years of taste about which central bank memo actually signals a rate move doesn't compress into instructions. It transfers through labeled examples. Every institution sitting on decades of expert decisions just learned that those archives can train a model that beats the frontier at their specific job. The alpha was in the filing cabinet the whole time.

Every time Google Maps gives you directions, your phone does this. 14,671 streets searched to find a single 2.3 km route across Naples. It's called Dijkstra, the undefeated king of shortest path since 1959. Until last month. For 66 years, every GPS, every flight booking, every internet packet route ran on the same algorithm. In 2024, Robert Tarjan and four co-authors won Best Paper at FOCS proving Dijkstra was optimal. The world's most-used algorithm, certified untouchable. Eight months later, a team at Tsinghua led by Ran Duan published a paper proving them wrong. The catch is in what "optimal" means. Tarjan's proof showed Dijkstra is the fastest possible algorithm IF you have to output every point sorted by distance. The Tsinghua group noticed something the field had quietly assumed for 41 years: finding the shortest path does not actually require that sorting. The problem just asks for the distances. They combined Bellman-Ford's batch updates with a recursive partial ordering trick from Duan's own 2023 paper. Instead of sorting the frontier, they cluster the boundary nodes and only explore the representatives. The new bound is O(m log^(2/3) n), beating the 1984 ceiling. Best Paper at STOC 2025. The reframe came before the algorithm. Tarjan did not prove Dijkstra was the best shortest path algorithm. He proved Dijkstra was the best sorted-output shortest path algorithm. The field treated those as the same problem for four decades. They are not. Every speed limit you have memorized has a definition wrapped around it. Crack the definition and the limit breaks. The world's most settled algorithm just got beat by someone asking what problem it was actually solving.

BART spent $90 million on new fare gates. They're recovering about $10 million a year in fares. That's a 9-year payback on paper. The actual return hit in six months. Embarcadero station went from 112 hours of corrective maintenance in the six months before installation to 2 hours after. Daly City saved 109. Balboa Park saved 75. Across the system, 961 hours of cleanup work disappeared. Corrective maintenance is the term BART uses for graffiti, heavy soiling, vandalism, the damage that needs a crew not a janitor. At several stations it dropped to zero. Crime fell 41% year over year. Riders who reported seeing fare evasion on their trip dropped from 22% to 10%. Citations issued by BART police went from 2,200 in January to under 1,000 in July, because there was nothing to cite. The gates were a filtering project disguised as a revenue project. Old BART gates were waist-high orange fins designed in the 1970s. You could hop them in under a second. That made the station effectively a public space, and the rider mix reflected that. The new gates are 72 inches of polycarbonate with 3D sensors that detect tailgating. You either pay or you don't enter. Once you don't enter, you also don't smoke on the platform, sleep in the elevator, or harass other riders. BART tried hiring more police for years. Blitz operations at high-traffic stations. Increased patrols. Dedicated transit cops. None of it moved the numbers the way six feet of polycarbonate did. The $10 million in recovered fares is the smallest line in the return. Fare revenue used to cover 70% of BART operations. After the pandemic it collapsed to 22%. The gates won't fix that gap directly. They fix the precondition for fixing it: a system that office workers, families, and tourists are willing to use again. Ridership growth at stations with new gates outpaced ungated ones before the rollout finished. A $400 million annual deficit is heading to voters in November as a sales tax measure. Voters don't approve sales taxes for transit agencies they don't feel safe in. The $90 million on gates is buying BART the right to ask the public for more money. That's the real return on six feet of polycarbonate.

The street on the right is 12°C cooler than the street on the left. Same sun, same asphalt, same afternoon. The canopy above it is doing the work of ten air conditioners per tree, and the electricity bill is zero. A mature street tree transpires 100 to 400 liters of water on a hot day. Evaporating one liter of water absorbs 2,260 kilojoules of heat energy straight out of the surrounding air. Run that through a single large tree and you get 70 kWh of cooling capacity per day. Wageningen University measured the live output at 20 to 30 kilowatts while the sun is up. A window AC unit runs at 2.5 kW. One street tree equals ten of them. Shade accounts for about 70% of the cooling. The other 30% is pure thermodynamics. Water molecules can't turn into vapor unless something hands them energy, and that energy comes out of the surrounding air. Every leaf on a canopy tree is a heat pump with no electricity bill. Then look at the ground. 46°C asphalt causes second-degree burns on bare skin in under 35 seconds. It also radiates stored heat into the apartments above it all night, which is why streets without canopy stay warm after sunset. The Camden, New Jersey sensor study found intersecting tree canopies were the only vegetation type that produced statistically meaningful air cooling above 30°C. Grass didn't. Shrubs didn't. The canopy has to actually close over the street. A Nature Communications meta-analysis of 182 studies across 110 cities put the ceiling at 12°C of pedestrian-level air cooling under dense canopy. The right side of the photo is hitting the upper bound of what physics allows. Heat is the deadliest weather event in the United States. 20% of natural hazard mortality. Cities spent 70 years removing the cheapest cooling system ever invented and replacing it with electric units that exhaust hot air into the street, making the outside hotter to keep the inside colder. The right side of that photo built its HVAC 40 years ago. It runs on rain.