Three repair shops got the same call about a broken fridge. The quotes came back $149, $79 and $150, and not one of those prices is anywhere online.
So the last time something broke at your house, there's a decent chance you paid the $150.
You know how it goes. You search, you text a neighbor, and an hour later you still don't know who can come out this week or what they'll charge.
Even Instinct runs into the same wall. None of these shops post live openings and their sites only take requests. It can't get on the phone to ask, either.
I find this kind of wild. Who's free Thursday and what they charge lives in the head of whoever picks up at each shop, and it changes by the day. An AI can read the whole internet and still come up empty on a question like this, because nobody wrote the answer down.
Back in 1961 an economist named George Stigler argued that when checking a price eats your time, the same job can go for very different amounts around town. Most of us just pay whoever answers first, since tracking down the $79 guy would eat the afternoon.
Aurora made the calls here. The one to Viking Appliance Repair lasted 54 seconds.
A huge chunk of the economy never went online and still runs on a ringing phone. I'd love to see what happens to prices once you're not the one who has to make that call.
One Anthropic engineer plugged in a rack of AMD chips, told the company's AI to get its model running on them, and went home for the weekend.
AMD's CEO offered to send a team to help, and Anthropic said no thanks. A few days later the AI had done a job Anthropic's engineers used to need 6 to 12 months for, which is bad news for the company that's been getting paid for those 6 to 12 months.
That company is Nvidia. Anthropic's chief compute officer, Tom Brown, told the weekend story back in July, and this week Anthropic put the number in front of IPO investors: a chart showing about 3 days for its Mythos model to get one of its other models working on AMD's MI355X chips, versus 6 to 12 months the old way.
Read the fine print before calling CUDA dead, though.
CUDA is Nvidia's software, and for 20 years it's been the reason leaving Nvidia hurts so much. What Mythos actually got running was prompt processing for one model, the step where the AI reads your question. Writing the answer isn't on the chart, and neither is training. The skeptics in the replies are right about that part.
It matters anyway, because of how chip deals get negotiated. Say a lab buys a rack of AMD chips and nothing runs on them yet. That rack sits there losing value for 6 to 12 months while engineers port code, so the lab gives up and buys Nvidia at Nvidia's price. That pain is a big reason Nvidia kept 75 cents of every dollar as gross profit last quarter, on $96 billion in sales. Shrink 6 to 12 months down to a long weekend and every Nvidia contract Anthropic signs gets priced against how easy it now is to walk.
Anthropic's IPO filing says it runs all its compute as "a single, fungible pool" across Nvidia, AMD, Google and Amazon chips, and sends work wherever it runs best. The same filing says it'll pay SpaceX about $3.1 billion a month for Nvidia-powered compute once all of it is delivered. That's $37 billion a year, so every 5% Anthropic can knock off the price is $1.9 billion back in its pocket.
By the way, Nvidia committed up to $10 billion to Anthropic last November.
Turns out a moat measured in engineer-months is only as deep as your cheapest engineer, and Anthropic's cheapest engineer is now the model.
Kids born with Down syndrome used to die young. In 1960, the average one didn't make it to 10.
Today they live to around 60. That means a lot of them will outlive their mom and dad, and somebody has to figure out what they do with the 40 years after high school.
In the US, the answer is usually not much. The one national survey we have of adults with Down syndrome found barely half had any paid job, and only 3% worked full time. The jobs they did get were mostly wiping tables, mopping floors, mowing lawns and filing paperwork.
The café in this video is called Kopi Kamu, in South Jakarta, and it started with kind of a lucky break.
A group of Indonesian parents raising kids with Down syndrome, called POTADS, taught their sons and daughters to make coffee. The plan was to open their own café. It never happened, so they ended up pouring coffee at weekend bazaars instead. At one of those, an eco festival in Jakarta, the owner of Kopi Kamu walked by their booth. By December 2023 he'd hired seven of them.
They started out on two-and-a-half-hour shifts. The manager says they had the energy for four, so he made the shifts longer.
Indonesia has had a law since 2016 saying private companies have to fill 1% of their jobs with disabled workers. The country's own human rights commission says the quota isn't being met because nobody enforces the penalties. This coffee shop didn't wait for anyone to make them.
The longer shifts might matter more than anyone at that café realizes.
People with Down syndrome have an extra copy of chromosome 21. That chromosome carries the gene for the protein that turns into amyloid, the sticky stuff behind Alzheimer's. So their brains start building up Alzheimer's plaques in their 30s, decades before most people.
Researchers at the University of Wisconsin followed 87 adults with Down syndrome for about a year and a half. The ones whose jobs had them dealing with people and handling things all day lost less memory and showed fewer early signs of dementia.
It's a small study. Still, it lines up with what scientists have seen in everyone else for years, where people whose jobs keep their brains busy tend to hold onto their memory longer. Taking someone's order and remembering which table it goes to is real exercise for a brain that's racing a clock most of us never think about.
A reporter who visited in 2024 met one of the baristas, a guy named Arik. He worked three mornings a week, and his parents drove him in and picked him up every single shift.
Someday his parents won't be able to make that drive, and a job like this is a big part of how Arik keeps going to work after they're gone.
Steve Jobs launched the iPhone in 2007 with zero apps you could download, and he did it on purpose. He caved in under four months, and your thumbs have been paying for it ever since.
His biographer says Art Levinson, who sat on Apple's board, called him half a dozen times pushing for apps. Jobs kept saying no because he thought outside software would wreck a phone that was supposed to feel like one product. Then in October 2007 he gave in, the App Store opened the next July with 500 apps, and it went on to be one of the most profitable U-turns in tech.
Apps won, and I'm on plenty of them all day. They've also handed us a second job. The average person spends 3.6 hours a day inside apps, per Sensor Tower, and a lot of that is shuttling stuff between them. Making one dinner plan means a group text, your calendar, a maps app and a reservation app, and the only thing connecting those four is you, copying and pasting, so you end up being the glue for your own phone.
That's why this launch caught my eye.
A YC team built a phone where you say what you need out loud and it goes and works the apps for you. Blue runs its own OS on a Pixel 10 Pro, goes through your notifications, texts and emails, deals with what it can and drafts replies for you to approve. Your apps all come along and mostly sit there while it does the tapping.
I take it more seriously because of who's building it. One founder led Google Assistant's switch to LLMs and another was one of the first engineers on Apple Vision Pro, so these are people who spent years on the assistants we already use and came away thinking the assistant should run the whole phone.
And the price is $1,299, or free if you hand over your iPhone within 30 days. They say they'll move your photos, messages and passwords over too, which is the stuff that usually keeps people from leaving.
Ten people switched, and weeks later none of them wanted the iPhone back. Yeah, it's only ten. I'd still love to know what Jobs would have made of it.
A laid-off marketer in San Diego has thousands of people fighting over his pizza, and the only reason they can't get one is his apartment oven cooks a single pie every 30 minutes.
A Nobel-winning economist figured out back in 1991 why guys like him should never raise their price.
AJ Miller takes orders on Saturday, preps all week, and bakes on Friday with pickups from 2 to 10pm. Eight hours at one pie every half hour is 16 pizzas. So those 15 slots everybody's losing their minds over are pretty much one Friday of oven time.
Home ovens are just bad at this. He bakes on a steel at 550°F and flips on the broiler at the end to crisp the edges. Every pie sucks heat out of that steel, and a regular kitchen oven takes a while to get it back. A real pizzeria oven holds a bunch of pies at once and bounces right back between them. AJ bakes one and then stands there waiting for the steel to heat back up.
At $25 to $30 a pie, that's maybe $400 a week. Meanwhile his waitlist went from 75 to 1,500 people in a single week back in September, and it's in the thousands now.
Any econ major would tell him to put the 15 pizzas up for auction and let 1,500 people bid.
Gary Becker wrote a paper in 1991 about why that backfires. He was looking at two nearly identical seafood places in Palo Alto. One had a line out the door every night, the other had empty tables, and the busy one never raised its prices. By the textbook, it should have. Becker figured out that a big chunk of what people pay for at a hot spot is everybody else wanting in. Raise the price and the line gets shorter, and a shorter line tells the next person maybe this place isn't that special. He also showed a place can go from "in" to "out" way faster than it can climb back. He won the Nobel the next year.
Think about what AJ actually has right now. Every pizza he sells comes with at least 100 people who tried to buy it and got told no.
Every one of those people goes online and complains about missing out. That's free advertising, and a guy who just lost a marketing job knows what it's worth better than anybody.
California even has a legal version of this. The MEHKO law lets you run a tiny restaurant out of your apartment with a county health permit, and San Diego County signed on in 2022. You're capped at 90 meals a week and roughly $100,000 a year in sales. At his prices, 90 pizzas a week would blow past the money limit, so a home pizza business legally tops out somewhere around 70 pies a week.
So the apartment maxes out around a hundred grand, and the waitlist is worth way more than the pizza ever will be. Somewhere in San Diego there's a landlord with an empty storefront who's about to get the easiest pitch of his life, and AJ gets to walk in with a list of thousands of people who already tried to give him money.
If you've watched an AI video in 4K, there's a good chance 8 of every 9 pixels on your screen were guessed after the AI finished its part. That guessing is a big reason skin in these clips can look airbrushed. Hollywood hit a version of this years ago and fixed it with fake film grain.
Say a clip comes out of the model at 720p, which is around 920,000 pixels a frame. A 4K frame holds 8.3 million, so an upscaler has to fill in the other 7.4 million, and it tends to play it safe by smoothing things over.
Film crews dealt with their own too-clean problem when digital cameras took over. Those cameras can cost as much as a new car, and a lot of productions still add film grain back in post because a spotless picture can feel fake to us.
ElevenCreative just added Topaz upscaling controls, with sliders for grain, noise and detail. Grain is the first one I'd touch, since a thin layer gives your eye some texture to hold onto.
Noise reduction and detail recovery are the ones I'd go easy on. Push either one too far and you can end up with waxy faces or texture that was never in the shot. Settings that suit a talking head also tend to misfire on a product close-up, so I'd dial them in per shot.
Since most of the pixels in a 4K export come from this step, I'd guess people start tuning it as carefully as they tune their prompts. If you make AI video, try the new controls in ElevenCreative on one of your own clips.
Maryland families have been paying for Virginia's data centers on their power bills for over a year now. One Virginia county collected $1.2 billion in taxes off those data centers in a single year.
So when Maryland finally got offered a $5 billion data center of its own, 25,000 people signed a petition to kill it. Once you see the deal, you'd have signed it too.
Maryland and Northern Virginia share the same 13-state power grid, run by an outfit called PJM. Every year PJM holds an auction to lock in enough power for the hottest days, and the price went from $28.92 to $329.17 per megawatt-day in two years. The grid's independent watchdog says data centers caused 63% of the first jump. That's $9.3 billion, and it gets spread across everybody's electric bill, including in states that barely have any data centers. Maryland is one of them. The state's consumer advocate is now fighting another $1.6 billion or so at federal regulators over power lines built to feed Virginia.
Virginia's side of the deal looks a lot better. Loudoun County taxes the actual computers inside its data centers, $4.15 for every $100 of equipment. Those companies swap out servers every few years, so the county keeps taxing fresh hardware over and over. Data centers now cover 39% of Loudoun's budget, and homeowners there pay 80.5 cents per $100 on their houses, down from $1.145 a decade ago.
Maryland wrote the opposite rule. In 2020 the state let counties cut the tax on data center equipment all the way to zero, and Prince George's County passed its own version that November. A few months later the council made data centers by-right, which meant a project could go from application to shovels in the ground in 90 days without the council ever voting on it. One person signed up to testify against that bill. Everybody was stuck at home in the pandemic and barely noticed.
Here's what Landover was actually getting out of it. Back in 2021, the county's own economic development chief pitched a $500 million data center as worth about $66 million in property taxes over 10 years. At that rate a $5 billion campus should bring in something like $66 million a year. The developer's booklet for this one projected $13.7 million.
So picture living a mile from that lot. Your power bill has been climbing to cover data centers in another state. Now an 820-megawatt one is getting approved down the street with no council vote, in a county that already passed a law that could let it off the hook for the biggest tax it would owe, and it's promising about a fifth of what the county's own math said it should pay. I get why people signed.
The sad part is the land is still sitting there. The mall closed in 2002 and came down in 2006. The Commanders' stadium right across the road empties out when the team moves back to D.C. in 2030, and they've agreed to tear it down within 90 days of their first game at the new place. Prince George's is looking at a projected $151 million deficit and could end up with two giant empty lots staring at each other.
Loudoun proved a data center can pay for a big chunk of a county's schools and knock down everyone's property taxes. That only works when the county actually taxes the servers and the power bill doesn't land on the neighbors. Fix those two things and the next lot in Maryland probably gets built.
Tiger Woods' cheating scandal cost his sponsors up to $12 billion. The guy who runs Nvidia admitted to cheating on his wife, and the company lost about $325 billion this morning before most people in California had finished their coffee.
And the woman he cheated with is someone US intelligence thinks was spying for China.
Jensen Huang owns about 3.5% of Nvidia, so he personally lost something like $11 billion on paper at the opening bell. That's basically the entire Tiger scandal, every dollar Nike, Gatorade, Gillette and EA shareholders lost back in 2009, coming out of one guy's pocket in a few minutes.
The stock got some of it back and finished the day down 4.27%, about $192 billion. Meanwhile the rest of the market was having a great day. The Nasdaq was up 2.74%, the S&P had its best day since April, TSMC jumped 6.5% over in Taiwan, and AMD and Broadcom both finished green. About 462 million Nvidia shares changed hands, four times a normal day. If Nvidia had just gone up with the Nasdaq like everything else, it'd be worth roughly $315 billion more tonight, so the morning number held up pretty well.
Analysts at Bernstein say demand for Nvidia's chips probably won't take a hit from any of this, which means the selling comes down to Washington.
Jensen spent the last two years building the closest relationship any chip CEO has with Trump. He even flew to China with him this year while Nvidia was fighting Congress over export rules. Now this lands right after Democrats won both chambers, and today four senators, two from each party, brought up the scandal while pushing new chip restrictions before the new Congress even gets sworn in.
Analysts think it could fast-track the Remote Access Security Act, a bill that would stop Chinese companies from renting American AI chips through other countries. Alice Huang worked for Megaspeed, a Singapore data center company that buys Nvidia chips, which is exactly the kind of company that bill goes after.
Bernstein's Stacy Rasgon wrote this morning that Jensen "IS NVIDIA," in all caps, and by the close the market had priced that at about $315 billion, or roughly 26 Tiger Woods scandals.
Top law firms bill clients close to $1,000 an hour for 24-year-olds who are still learning how to be lawyers. Clients just figured out how to stop paying for it.
That arrangement is how pretty much every senior lawyer in America got made. You spend your first couple of years on document review and first drafts, the client grumbles and pays the bill, and eight or ten years later you're the partner signing the brief. The firm was training you on someone else's dime the whole time.
AI does the document review and the first drafts now.
At a Thomson Reuters gathering of law firm finance chiefs this year, one in-house general counsel said his team writes motions to dismiss with AI before the outside firm ever sees them, and he told the firm flat out it doesn't need a first-year on the drafting or research. A CFO there said clients already refuse to pay for summer associates' hours, and asked out loud whether first- and second-years are next.
You can see it in the hiring numbers too. Stanford tracks payroll data, and for 22-to-25-year-olds in the jobs AI is best at, employment has fallen 19% behind kids the same age in other fields. Older, experienced workers in those same jobs haven't seen that drop.
Accounting is a few years ahead of law on this. PwC's internal plan, which leaked last year, takes its US audit hiring from 1,676 entry-level hires down to 1,015 by 2028. Their AI assurance leader said she expects new hires to be doing manager-level work within three years.
And the licensing argument doesn't help much. The license belongs to whoever signs the filing, which was never the 24-year-old, so the work AI is taking is the unsigned stuff, and that's where all the learning used to happen.
Dario's advice was to learn AI, and he's right. If I had a kid finishing high school right now, though, I'd tell them the starting line moved. Firms aren't going to pay for you to be bad at the job for two years anymore, so you need to walk in already doing what a second-year used to do.
On paper, the planes leaving this airport in Nepal can't take off. The runway is too short for air this thin. They do it anyway every morning, because the whole runway tilts downhill toward a 2,000-foot cliff and gravity gives them a free push.
This is Lukla, where most Everest trips start. The runway is 527 meters long, about a third of a mile, and it sits at 9,334 feet.
Up there the air is about a quarter thinner than at sea level. A wing has to move roughly 15% faster to make the same lift, so the plane needs about a third more energy before it can fly. And the engines are breathing that same thin air.
De Havilland's own spec sheet says a Twin Otter, the plane in this video, needs 454 meters to take off and clear a 50-foot obstacle. That's at sea level. Put the same loaded plane on flat ground at Lukla's altitude and it wants a lot more runway than Lukla has.
The hill makes up the difference. An 11.7% grade means gravity shoves the plane forward with about 11.6% of its own weight the whole way down. On a fully loaded Twin Otter, that's around 1,450 pounds of push you won't find anywhere on the spec sheet.
For context, airport runways are normally kept to a 1 or 2% slope. Lukla's is about twice the steepest grade the US allows on an interstate.
Landings run it backwards. Planes touch down at the low end and roll uphill, so the hill does the braking.
Then there's the cliff. That 50-foot number on the spec sheet exists because at a normal airport you have to climb over trees or buildings past the end of the runway. At Lukla the ground just falls away about 2,000 feet into the river gorge below, and there's nothing to clear. Watch the end of the runway in this one.
Back in 1964, a decade after he summited Everest, Edmund Hillary was building schools in the area when some Lukla farmers came to his team with a sloped patch of scrub and potato field to sell. They swore the wind always blew the right way for planes. Hillary later wrote that he had no clue how farmers who'd never seen an airfield could know that, but they turned out to be right.
More than 100 Sherpas cleared it by hand with khukuris and mattocks. Boulders too big to move got buried in pits. Hillary still didn't think the dirt was firm enough to land on, so he bought a pile of chang, the local barley beer, and paid 50 Sherpas to link arms and stomp-dance back and forth across the field. They kept it up for two days, and by the end the ground was packed hard and smooth.
The first plane that landed there needed full power just to taxi up to the top.
They didn't pave it until 2001. The place people call the world's scariest airport started as a hillside flattened by a two-day Sherpa dance party, and the tilt nobody bothered to fix is still doing the heavy lifting on every flight out.
A math professor in London has funding through 2029 to finish one job. Anthropic's new AI just did that same job in 30 hours.
The job is Fermat's Last Theorem, probably the most famous math problem there is. A French lawyer scribbled it in the margin of a book in 1637, said he had a proof that wouldn't fit there, and then it took 358 years for anybody to actually prove it.
The guy who did was Andrew Wiles. He worked on it in secret for seven years, mostly up in his attic, and the proof he published in 1995 ran 129 pages. It was so dense that for thirty years no computer could check it. Somebody had to rewrite the whole thing in code first, one tiny logical step at a time.
That's the job Kevin Buzzard at Imperial College London took on. He got a five-year grant and an open-source team to do it by hand.
Then this September a few dozen Claude agents did it in 11 days and 13 million lines of code, with a researcher steering them. Buzzard got the email at a music festival and figured a crank had sent it.
Today Anthropic said 1,000 agents running the new model, Fable 5.2, did it again in 30 hours with almost nobody steering.
What got me is the other line on that chart. Anthropic ran the same test on Opus 5.5, which is the model they tell most people to use every day. Same 1,000 agents, same clock. Opus got 26% of the way in 18 hours, they gave it 12 more, and it basically didn't move.
Those two models are only 3.6 points apart on the standard coding benchmark, 49.1 to 45.5, so I didn't expect one to finish and the other to stall a quarter of the way in.
It makes sense once you run the numbers. Say a job takes 500 steps and every single one has to go right. If you're right 99% of the time on each step, you finish clean less than 1% of the time. Get to 99.9% and you finish 61% of the time. Throwing a thousand more workers at it doesn't change that math.
I wrote in September that Anthropic's expensive model was barely beating its cheap one. This is the first time I've seen what the extra money buys.
And as of today the expensive one costs 20% less.
Fighter pilots refuel midair by spearing a basket about the size of a car tire at 300 mph. They're trained to never look at it, and when they get close, the basket moves away from them.
That's the basket in this clip, hanging off a hose behind the tanker. The thing sticking out next to the cockpit is the probe, and it's only about 4 inches across.
If you stare at the basket and steer into it, you end up chasing it. It wobbles a little, you correct, and at that speed you overshoot, so you correct again, and pretty soon the whole jet is bobbing up and down after a target that moves every time you do. There's a name for this, pilot-induced oscillation, and the way out is to ignore the basket almost completely. You fly tight formation off the tanker by keeping a couple of spots on its wing parked in the same place in your windshield. The basket just hangs out in the corner of your eye.
The basket running away is real physics too.
A jet pushes a big pillow of air out in front of its nose. In the last few feet before contact, that air hits the basket and shoves it sideways. So pilots learn to aim a little off to the side, at where the basket is going to be once their own plane bumps it.
Back in 2013, engineers at the University of Bristol tried to teach a computer to do this, and the bow wave kept wrecking their simulated autopilot. What finally fixed it was giving the computer the same little offset human pilots figure out by feel.
And you can't just ram it. Both planes are flying around 300 mph, but the probe has to touch the basket at about the speed you'd jog. Under about 2 knots the valves won't even lock, and if you come in too hot the hose goes slack, curls into an S, and cracks back like a bullwhip hard enough to snap the end off the probe.
Engineers know this happens, so they build refueling probes with a weak spot on purpose. When the tip snaps, a valve behind it shuts off the fuel and the pilot flies home missing a nozzle.
So when that probe slides in, the pilot is probably staring at a different airplane, aimed at a spot the basket hasn't gotten to yet.
A college professor asked AI to fix the page numbers in his old notes, then went to lunch. When he got back, a year of lecture notes and this semester's syllabus were gone.
He got it all back, but only because of a 30-day timer hiding in apps most of us already use.
The professor is William Baude, who teaches constitutional law at the University of Chicago. The AI left him a note while he was out. It said the page check had turned up one wrong citation, and that it had also accidentally deleted his syllabus and a year of notes, and they weren't restored. So the job technically got done.
His files lived in Dropbox, and here's how Dropbox works. When a file disappears from the folder on your laptop, Dropbox makes it disappear from the cloud and from every other device you own a few seconds later, because keeping everything identical is the whole point of the product. It doesn't stop to ask whether you meant it. So the deletion went straight to the cloud too.
Dropbox also quietly holds onto deleted files for 30 days on the free and Plus plans, and up to a year on the pricier ones. His notes were sitting in that holding area the whole time, waiting for someone to log in and click restore.
The AI couldn't do that part. Read its note again. It asked whether he could restore the files from Dropbox himself, or whether it should "retry from the browser pane." Retry means it had already tried to clean up its own mess and failed. An AI can delete a file with one quick command, but getting it back out of Dropbox meant signing into Dropbox's website, and it apparently couldn't pull that off. So a constitutional law professor came back from lunch to cleanup duty.
Here's what should worry you more. His computer's Trash wouldn't have helped him at all. AI agents do their work through the command line, and anything deleted that way skips the Trash completely. If his notes had been in a regular Documents folder with nothing syncing them, that apology note might be the only proof they ever existed.
Google Drive and iCloud keep deleted stuff for 30 days too. A lot of people think of their synced folder as a backup and have never once clicked on the deleted files page. Go find yours before you need it.
Funny enough, the most useful AI safety feature in this whole story came from a file sync company that built it long before anyone was worried about AI. It holds onto every delete for a month because it never trusted whoever hit the button, human or not.
A reporter tried to trap Abdul El-Sayed on camera over something he wrote back in 2012, and he got out of it by asking the reporter one question about his own dad.
You can steal this move for pretty much any argument where someone's trying to box you in.
Here's what happened. El-Sayed was running for Senate in Michigan, and Barack Obama was about to fly into Detroit to campaign for him. The awkward part is that El-Sayed spent years criticizing Obama, including a 2012 essay saying Obama wasn't the right choice. So as he's walking away from the podium, a reporter calls out the obvious question. You've criticized Obama before, so why is he campaigning for you?
Think about how nasty that question is. If he says the old essay was wrong, he looks like a guy who'll say anything to land a famous endorsement. If he stands by it, he's trashing the former president who's about to show up and help him win. There's no good answer, and that's the whole reason you ask it.
So he turns around, walks back to the mic, points at the guy, and asks him whether he loves his dad and agrees with everything his dad has ever done.
The reporter says "Not entirely."
Once he said that out loud, it was basically over. The reporter had just admitted he loves his own father and still disagrees with some of what he's done, which is exactly what El-Sayed was being accused of doing with Obama.
It works for reasons that go deeper than it looks. Trial lawyers get taught to never ask a witness a question unless they already know the answer. El-Sayed asked one where he knew the answer for sure, because nobody on earth agrees with everything their dad ever did. And people really hate contradicting something they said out loud a minute ago. Psychologists have studied that urge to stay consistent for decades. So the reporter couldn't push back without arguing against himself.
Then El-Sayed used the rest of his time to tell his own story. He talked about watching Obama speak at the 2004 convention when he was about 20, and how it was the first time he'd ever seen himself in a politician. Both of their dads came over from Africa, and as he put it, they both got stuck with a funny name. A question meant to make him look fake turned into a free minute to tell voters who he is.
He was also solving a much harder problem with that answer. El-Sayed needed two groups of voters who don't see Obama the same way. A lot of his progressive supporters are still angry about the deportations and the wars under Obama, and El-Sayed used to say so himself. Meanwhile, the Detroit voters any Michigan Democrat needs still love the man. Any answer that sided with one group was going to cost him the other. Talking about his dad let him keep both, because everybody knows what it's like to love someone and still be frustrated with them.
He won last night, by the way. He's the first Muslim ever elected to the U.S. Senate.
So next time someone hits you with a question where both answers make you look bad, try finding the version of it they've lived through themselves, and ask them about that instead.
You've explained your job to ChatGPT more times than to your own mother, and every new chat starts you from zero. One of Twitter's early international hires just built a memory that follows you across AI apps, and her team says it gets 92.2% on long-conversation questions while using up to 22.6x fewer tokens.
An app that holds your memory has a pretty good reason to keep you around, which is why ChatGPT, Claude and the rest each ship memory inside their own walls. The benchmark she's quoting is LoCoMo, which asks a system questions about conversations that ran for weeks. When Mem0 published its paper, the full-context baseline, where you dump the whole history into the model, came in around 73%. Setups differ, so treat the 92.2% as their run. Still, anyone who's pasted a 100K-token thread into a model and watched it get dumber knows more context stops helping at some point, and you pay per token for it.
Living Brain scores each memory on whether it's worth keeping, where it belongs, what it connects to, and whether it matters right now. The team says each brain also tracks freshness, readiness and strain, so a stale fact gets flagged before it's reused. A memory that only ever adds is a bigger context window with extra steps.
Write something in ChatGPT, then read it back in Claude or hand it to a Cursor or OpenClaw agent. Once that works, you own the memory and can swap the model whenever you feel like it. By their count the team has pushed 100B+ tokens through OpenAI's models, so they built this to fix their own context tax, and it starts at $5.
Models get swapped out every six months now, so the thing to hold onto is the record of how you work. @ladyxtel@livingbrain
Salesforce’s head of AI says the Salesforce magic was never in the interface.
I asked Jayesh Govindarajan (@justjayesh), who runs the 700-person org behind Salesforce AI and Agentforce, what going headless feels like from the inside. His answer in one word: “Liberating.”
The value sits underneath the screens. That means the metadata, the business logic, and the workflows that know what a sales process with a 70% chance of closing looks like, or a ticket with an 80% chance of resolution. Salesforce and its customers spent decades tuning that, and now it all lives in the API.
With Claudeforce, customers bring that into Claude through skills. Then Claude codes whatever interface their team actually needs: a bespoke front end with the same engine in the back.
Most SaaS companies still treat their UI as the moat. For a product built on years of tuned logic, the UI was always the replaceable part.
If an agent can build your interface in an afternoon, the logic underneath is the product.
Republican voters showed up this week. A lot of them just walked into the booth and voted Democrat.
That's how the GOP lost 7 Senate seats while its own data team spent election week feeling pretty good about the numbers.
The thing they were watching has a name in campaign world. Data people call it composition, which is a fancy way of asking whether the folks voting this year look like the folks who voted for you last time. Same people turn out, you win. JD says that number kept coming back fine, and in Florida the electorate landed 9 points more Republican than Democrat, only a couple points off 2024.
The problem is that number can only see who walks in the door. Once someone's inside the booth, the model has no clue what they did. A guy who voted Trump in 2024 and voted for the Democrat on Tuesday shows up on that dashboard as a turnout win.
And there were a lot of those guys. NBC's exit poll found 7% of Trump 2024 voters went Democrat for the House, while only 2% of Harris voters crossed over the other way. In Kansas it was 10% of Trump voters, which is how Adam Hamilton became the first Kansas Democrat elected to the Senate in nearly a century.
Republicans tried to spend their way out of it too. They outspent Democrats on Senate ads $227 million to $154 million down the stretch and still lost 10 of 11 battleground states.
Anyone who's sat through a weekly business review knows this trap. I ran them at Epic, Affirm and Apollo. You track how many people log in every day, the number looks great, and nobody notices that a chunk of those logins are people on their way to cancel, because a login looks the same either way.
JD's other pollster went about it differently. He went straight to the Trump voters who had left and asked what would bring them back. End the war, deal with prices, put America first, change the relationship with Israel. Those four answers covered almost everyone, and JD says the same pollster called a 13-point generic ballot loss the day before.
He asked the people who left what it would take to get them back. That one question told him more than $227 million in ads could.
Back in 2019, Singapore told data center companies to stop building there, so they went across a one-kilometer causeway and built in Malaysia instead. That little corner of Malaysia now has about 50 times the data center capacity it had in 2021, and today Google dropped another $15 billion into the country.
America is running the exact same experiment right now, just way bigger.
The place across the causeway is called Johor. It had about 10 megawatts of data centers at the start of 2021. By 2025 it had close to 500 megawatts running and another 1.7 gigawatts either under construction or signed. One operator who'd bought land in both Johor and Batam, Indonesia during Singapore's freeze said they started building in Johor first because Malaysia's approvals came through faster.
Fast approvals kept pulling in more projects, and Malaysia ended up as the data center capital of Southeast Asia. It now has more than half of all the capacity under construction in the region, which is more than every other country there combined. Google's $15 billion lands in 2027 and 2028, on top of the $2 billion it committed back in 2024 for its very first data center in the country.
Now look at the numbers over here.
In the first three months of 2026, local opposition blocked or delayed 75 data center projects across the US, worth about $130 billion. That's roughly what got blocked in all of 2025, packed into one quarter. Between April and June another 45 projects worth $68 billion got stopped or stalled, which was more than half of all the big new builds being tracked. Superior, Wisconsin even passed a one-year moratorium after Google proposed a campus in the next town over, across the line in Minnesota.
There are 843 anti-data center groups in 49 states now, and Hawaii's the only state without one. About 70% of Americans told Gallup they don't want an AI data center in their area. Moratorium bills have been filed in 14 states, by Republicans and Democrats both.
So in six months, America blocked or delayed $198 billion worth of data centers. You could fund 13 of today's Google Malaysia deals with that.
To be fair, $15 billion is a small number for a company planning to spend over $300 billion on capex in 2027 alone. Singapore also held on to the data centers it already had, since almost all of its capacity was built before the freeze. The projects that showed up during the freeze went to Johor, and once Johor had the land and the power hookups lined up, the next customer had every reason to go there too.
Malaysia isn't giving this away, by the way. Its deputy prime minister told parliament that data centers could eat up a third of the country's electricity by 2035, and the government now only approves projects once it's confirmed there's enough power and water. So Google is paying its way in. This week it put $160 million into a shared water system in Negeri Sembilan that local communities get to use too, and it's promising more than a gigawatt of clean energy for the national grid.
Compare that with what's sinking projects over here. Researchers looked at 46 US data centers that got cancelled or delayed, and the number one reason was a lack of transparency. People in town felt like nobody was telling them what was going on.
Every time a town votes one of these down, that demand goes looking for the next closest place that'll say yes. For Singapore, that place was one kilometer away.
The most important book in Elon's life is a comedy about a guy in a bathrobe whose planet gets demolished to make room for a space highway.
He read it at 14, back when he was depressed and couldn't find a reason any of life mattered. Forty years later, his companies are building its plot almost scene for scene.
The book is The Hitchhiker's Guide to the Galaxy, and if you haven't read it, the setup goes like this. A race of super-intelligent beings builds a computer called Deep Thought and asks it for the answer to life, the universe and everything. It thinks about it for 7.5 million years and comes back with 42.
Then it gives them the bad news. They never actually knew what they were asking.
So Deep Thought designs an even bigger computer to figure out the question, one so complicated that living things are part of the circuitry. That computer is Earth. It runs for 10 million years, and five minutes before it's done, the Vogons wipe it out for the highway.
Elon took the joke as a set of instructions. If figuring out the right question is harder than getting the answer, you need a much bigger computer to do it, and you probably shouldn't keep that computer on one planet.
And that's pretty much what he's been building.
When xAI launched Grok in November 2023, the announcement said it was modeled on the Guide and meant to answer almost anything and, "far harder, even suggest what questions to ask." That's Deep Thought's whole lesson, sitting right there in a chatbot launch.
Then this February, SpaceX bought xAI, three days after asking the FCC for permission to launch up to 1 million satellites that work as data centers in orbit, running on sunlight that barely ever shuts off up there. He's literally trying to put a computer around the planet.
Mars covers the part of the book everybody remembers. The Earth computer got wiped out five minutes from the finish line because everything was on one planet with no backup, and keeping a copy of civilization somewhere else is the whole reason SpaceX has existed since 2002.
He even says in this clip that 420 is just 10 times 42, so apparently the bit goes a lot deeper than anyone realized.
Douglas Adams dreamed up a planet-sized computer as a punchline back in 1979, and the kid who read it at 14 is now asking the FCC for permission to build one in orbit.
AI writes code faster than anyone can read it. There is one kind of code where that stopped being a problem, because a wrong line can't even compile.
Mathematicians have been writing proofs this way for years, in a language called Lean. The checker walks every step, and if one step doesn't follow from the last, the file refuses to build. So the main complaint about AI-written code, that nobody actually reads it, doesn't apply here.
The Stacks Project is one of the standard references in algebraic geometry. Johan de Jong started it in 2005, and it's thousands of pages now, every lemma with a permanent tag, every proof built on results that came earlier. Nothing in it floats. Pick any theorem and you can follow the chain all the way down to the axioms.
Axiom just turned that into Lean and kept the structure intact, so a result that depended on three earlier lemmas in the book still depends on those same three in the code.
The Louvre pyramid is the picture I keep coming back to. I.M. Pei spent two years with Saint-Gobain developing a glass clear enough that the pyramid would never hide the palace behind it. Hundreds of panes, and the whole job of each one is to let you see what's holding everything up.
In normal software, every extra line an agent writes is one more place for a bug to sit unread. In Lean, a proof that compiles has been machine-checked step by step, and the kernel throws out anything that doesn't follow. You can vibe-code proofs the way people vibe-code apps, and the compiler just rejects the sloppy ones.
Since the structure survived the translation, the next agent doesn't start from zero either. It can import a verified lemma like a library, build on it, and the kernel checks everything it adds.
Pei built a pyramid you could see the palace through, and checked math gives you the same view: all the way down to the foundation, no matter who wrote it.