Pinned Tweet
Outlook has been broken for most of a workday, and Microsoft can't just switch it back on. Every phone and laptop that's been retrying email since this morning would hit the servers in the same second and knock them right back over.
That's pretty much how it broke in the first place.
Every Exchange server gets watched by an automated health check, and if it fails, it gets yanked out of rotation so nobody's email lands on a broken machine. With one bad server, that's great design. Today a core process on a whole bunch of them started choking at once (Tom Warren reports a rollout spiked CPU and pushed racks into maintenance mode), and Microsoft's own update says servers were failing those health checks and dropping out.
So follow one person's inbox. Their server gets pulled, but their Outlook app doesn't know or care. It just keeps retrying, and those retries land on the next server over, which was already running hot from everybody else's retries, so it fails its check too and dumps its users onto the next one. Now multiply that by a product sold to 450 million paid seats.
The system that removes sick servers kept making the healthy ones sick.
Microsoft's first big move was spreading connection traffic across its worldwide infrastructure. A few hours later it posted that nothing so far had "provided the anticipated relief," which makes sense once you see the chain, because spreading the load in a cascade also spreads whatever's killing the servers.
That's why tonight's plan looks so slow. They're walling off a slice of servers, restarting them, letting traffic back in a trickle at a time, and only expanding once it holds. Skip that step and you get what engineers call a thundering herd, where the fresh servers get trampled before they finish warming up. Earlier this evening Microsoft finally said it's seeing positive results.
This has happened to them before, by the way. In November 2024, Outlook and Teams went down after a change caused "an influx of retry requests routed through servers."
A whole workday later, the hardest part of fixing Outlook is that millions of inboxes are all trying to help.
Replying to @MSFT365Status
We’re opening all available pathways to restore service, as our actions taken thus far have not provided the anticipated relief. We found a core process on affected servers is becoming overwhelmed, causing capacity loss. To validate a potential recovery, we're isolating a portion of infrastructure, restarting affected components, and reintroducing traffic to confirm the services remain stable before applying the change more broadly. More info under EX1550543 and MO1550821 in the Microsoft 365 admin center, as well as status.cloud.microsoft if the admin center is not available to you.
Bolt's former CPO told me he'd have been happy to talk to seven or eight customers a month.
With AI, he says, it's now a thousand times that or more. Greg Greiner built the agent that does it in Claude Code, in 30 minutes, before we recorded, then added the safety check live on the show.
Full episode with Greg (ex-CPO at Bolt, CEO of Woven):
youtu.be/JpyhPePe150
0:00 Intro
1:36 What AI research tells you
6:32 Build your own research agent
8:52 Demo: emailing drop-offs
12:29 Why email beats in-app chat
16:03 When to email customers
24:15 Adding an eval loop, live
31:23 When to talk to customers yourself
35:46 The PM's AI research agenda
40:28 1 PM, a team that ships like 25
45:45 AI for your team vs customers
52:03 Inside an AI-native startup
1
1
1,716
Figma is growing faster than Cloudflare. 42.8% vs 39.2% in the latest quarter.
The market pays 29.8x next year's revenue for Cloudflare. For Figma it pays 4.7x.
Slower growth, more than 6 times the price.
So I charted 15 public software companies. Growth still sets the price on average. The dotted line adds about 3x to the multiple for every 10 extra points of growth.
Webflow's CEO Linda Tong walked me through why the line slopes up:
"before you were growing 50%, you could get a 10x multiple, because they figure that to actually reach that valuation you only need four or five years. Now if you're growing 10 or 20%, you need way more time to justify the multiple. So you now see 2 to 3x multiples for them. The math itself isn't wrong."
Today the market is paying more than her rule of thumb. The trend line sits near 19x at 50% growth. AppLovin, the closest company to 50% at 55.5%, trades at 13.2x.
But growth only explains about a third of it. The r-squared is 0.37.
Figma and monday.com sit way below the line. CrowdStrike and Manhattan Associates sit way above it. CrowdStrike grows 25.8% and trades at 34.2x. ServiceNow grows 20.5% and trades at 5.6x. Five points of growth doesn't come close to explaining that.
Linda's math runs on years: how long growth takes to catch up with the price. The chart only shows one quarter. Two companies with the same quarter can have very different years ahead of them.
My full conversation with Linda:
youtu.be/vPBUYcJIk1c
If the Sun disappeared right now, you would get another 8 minutes and 20 seconds of a completely normal sky.
NASA's figure for how long sunlight takes to cover the 93 million miles to us is about eight minutes and 20 seconds. Earth moves along its orbit at 29.78 kilometers per second, so over those 500 seconds it travels about 14,900 kilometers. That is farther than the 12,742 kilometers across the planet itself, and nothing on Earth would notice.
Then the lights go out. Earth's core is still hot, though, and that heat reaches the deep ocean, where life might hang on near hydrothermal vents.
Until 1977, scientists thought every food chain on Earth depended on the Sun. That February, John Corliss and Tjeerd van Andel dove in the submersible Alvin 2.5 kilometers down the Galapagos Rift and found food chains running on chemicals from the vents instead. There were no biologists on the mission, so the geoscientists collected the specimens themselves, and when the ship's formaldehyde ran out they preserved them in vodka.
You'd think a vent animal wouldn't need the Sun at all, but some of them depend on oxygen made by photosynthesis, and the oxygen in the ocean is a byproduct of it. Even down there the Sun isn't fully out of the picture, which is why "for a while" is the honest answer.
Then there's gravity. Newton's theory treats gravity as acting instantly at any distance, which would have Earth feel the Sun vanish at once. Einstein's general relativity says changes in gravity travel at the speed of light, so the pull would disappear at the same mark as the light. Physicists answering this exact question on the University of Illinois Physics Van site say any change in the Sun would have no influence here for 8 minutes.
That got tested in 2017. Two neutron stars collided about 130 million light-years away. The gravitational waves reached us first, and the Fermi and INTEGRAL space telescopes caught the gamma-ray burst from the same crash 1.7 seconds later. After a trip that long, that gap pins the speed of gravity to within about 3 parts in a quadrillion of the speed of light.
So the dark and the drift arrive at the same second, and a collision 130 million light-years away is the measurement that backs it.
1
4
3,530
Dan Bilzerian's dad owed the government $62 million, told a judge he had $15,805 to his name, and sat in jail for almost a year instead of paying.
Meanwhile, a trust for Dan and his little brother owned 2.3 million shares of the company Dad ran. When the trust got those shares in 1997, they were worth about $12 million.
That's the trust fund Dan keeps getting asked about.
Here's how it got built. Paul Bilzerian was convicted of securities fraud in 1989, and in 1993 a federal judge ordered him to pay back $62 million. The next year he set up a trust for his two boys. That same summer, a holding company he owned with his wife got the right to buy 6 million shares of Cimetrix, a Utah software company he ran, for 16 cents a share. Those rights got handed down through family entities, and by the time everyone cashed them in, the shares were worth over $31 million. The kids' trust ended up with 2.3 million of them. Another 2.9 million went to a partnership owned by a family trust set up in the Cook Islands.
By 2000 he hadn't paid one penny of the $62 million, and he said as much in court.
The judge locked him up for contempt in January 2001. He got out about a year later through a settlement, which matches what Dan says in this interview about signing over a third of his fund to get his dad released.
Over twenty-plus years the SEC spent $8.6 million chasing him and collected $3.7 million. That's $2.32 spent for every dollar recovered, and all in they got back about 6% of what he owed.
Dan's book says the trust was worth around $96 million at its peak and that what was left sold for about $1.3 million. When your whole trust is one tiny stock, it goes wherever that stock goes.
Then he admits he let people think he was a rich kid blowing his dad's money because that's what got him invited to private games with billionaires. Those games want a guy with deep pockets who everybody figures can't play. Paying $1,500 a month for his share of a $20 million mansion made him look like that guy. And his $10.5 million morning checks out, since he owned 25% of a player who won $42 million in one night.
Ignite ran on the same idea, just with public shareholders. The company lost about $50 million in 2019. Dan got pushed out as CEO in December 2023, and in his own lawsuit against his dad he says he was a figurehead who never ran anything. In 2024 the SEC charged Paul, who had no official job at Ignite, with booking $4.6 million in sales that never happened in late 2020, while the stock went from 42 cents to $1.20. Dan wasn't charged.
Honestly, the poker money might be the cleanest money in the whole family.
Dan Bilzerian, Instagram'da sergilediği lüks hayatı nasıl finanse ettiğini ve servetinin kaynağını anlattı:
🔹 "Babamın kurduğu güven fonunun değeri başlangıçta yaklaşık 100 milyon dolardı."
🔹 "Babamı hapisten çıkarmak için fonun üçte birinden vazgeçtim."
🔹 "Kalanını sattığımda değeri 1-2 milyon dolara kadar düşmüştü."
🔹 "O dönem poker oynayarak yılda 50 milyon dolar kazandığım oluyordu."
🔹 "Herkes babamın parasını harcayan zengin bir çocuk olduğumu düşünüyordu."
🔹 "Bu algıyı hiç düzeltmedim çünkü milyarderlerin katıldığı özel poker masalarına girmemi sağlıyordu."
🔹 "2011'de 20 milyon dolarlık bir malikaneyi paylaşıyordum, aylık kira payım sadece 1.500 dolardı."
🔹 "Evi pahalı eşyalarla doldurmuş, lüks arabalar almıştım."
🔹 "Dışarıdan yüz milyonlarca dolarım varmış gibi görünüyordu ama bankada belki 3 milyon dolarım vardı."
🔹 "Bir oyuncunun kazancının %25'ine ortaktım ve bir gecede 42 milyon dolar kazandı."
🔹 "Sabah uyandığımda 10,5 milyon dolar kazandığımı söyleyen bir mesaj gördüm."
🔹 "Bir milyarderi kendi evinde poker oynayarak 42 milyon dolar yendim."
🔹 "Sonrasında Ignite'ın yönetim kurulu beni CEO'luk görevinden aldı ve 750 milyon dolarlık satış anlaşması gerçekleşmedi."
To get all my takes without an algorithmic filter, subscribe to my newsletter:
aibyaakash.com
1,255
At some point today you told someone "I'll send that over." You haven't, and tomorrow you'll have forgotten you ever said it.
You'll say more than 12,000 words today and almost none of it gets written down. Amazon and Meta have both paid real money to fix that.
The number is from the biggest study anyone's run on this, about 2,200 people wearing recorders. Somewhere in those words is everything you promised today, like the deck you'd send and dinner Thursday. By 4pm most of it is gone.
Then at night you open Claude and type whatever you still remember, at maybe 40 words a minute.
That part has bugged me ever since agents got good. An agent can write the code and book the table for you, but it has no clue what you agreed to at lunch, because lunch only ever happened out loud.
So look at what happened to the wearables. Humane tried to replace your phone for $699 plus $24 a month and ended up selling to HP for $116M. Meanwhile Amazon grabbed Bee, a little wristband that just listens, and Meta picked up Limitless a few months later. Big tech basically paid for the microphones.
Vexo is a $199 bracelet making that same bet, plus one more step. What it hears goes to agents that already work in 2,000+ apps, so the "I'll send that over" actually gets sent. Anything sensitive, it checks with you first. Battery's around four days and it runs off your phone. You can switch the listening off whenever you want, and sometimes you should, since everyone near you gets recorded too.
Most of my own follow-ups start as something I said on a call, so I'm exactly who this is for. Typing is the slow part of working with AI. I'd rather it just hear me.
🤝 Paid partnership
Backed by YC and Paul Graham, we're building Vexo, an always-listening AI bracelet that turns your conversations into action.
Built by engineers from Tesla, AMD, and Amazon.
One sentence to Vexo can:
•Get Claude building the feature you wanted
•Text the family
•Block your calendar for dinner
•Order the groceries
•Find the founder you met at last week's event
Every conversation adds context, so you explain less and get more done.
Get yours now, shipping start January 30th 2027
vexoai.com/checkout
5
6
3,465
A new hire waits two weeks for a first paycheck, and about ten software companies get paid for her on day one. I had our AI employee add it up from public pricing pages last week. It came to roughly $1492 a year in subscriptions for one extra person, whether or not she ever opens them.
I wanted the number for a deep dive on software pricing, so my growth lead asked him for it in Slack. He came back with a little page where you pick your tools and a headcount, and it shows what the bill comes to. He'd run the tool list by us before building anything. I still went through every price against the vendor's own pricing page before I trusted the total, and they're list prices, so if you've negotiated a contract yours will be lower.
Charging per seat, meaning one fee for every person with a login, made sense back when software was something a person sat down and used, since headcount was a decent stand-in for the value you got. But a vendor's revenue goes up every time your payroll does, even if the new person logs in twice a year. I'd rather put that money toward the person we just hired.
O8 is a 20-person digital agency, and nineteen of the twenty use Viktor every day. Seth Viebrock, who founded it, estimates the extra capacity at $200,000 to $400,000 a year. He's upfront that it's his own estimate and nobody ran a formal time study, and a person still reviews everything before a client sees it.
With a per-seat tool they'd pay for all nineteen, and pay again the day the twentieth person starts. At O8 that twentieth person can start handing him work without anyone buying another seat.
Vendors are going to have a hard time charging by the chair once a good chunk of the work gets done by someone who doesn't need one.
If you want one on your team, hire @viktor_com. Free to start.
#AIemployee
Paid Partnership
🤝 Paid partnership
Cat Wu ran product for Claude Code.
Now she leads core product at OpenAI.
Here's what she said about evals:
Most people become over-obsessed with the theory around evals. "I need to do axial coding and open coding. I need to generate 100 test cases. I need to look at all their results."
You don't need to build hundreds of evals for them to be useful. Just building 10 great evals is more important, and you heard it from Cat herself.
The other thing you want to do is help the team progress. Once you've quantified the goal, evals help them see what's missing.
It's not just about the evals. It's about how you team up with the team on top of it.
Here's your roadmap to better evals:
1. The theory: news.aakashg.com/p/ai-evals?…
2. The first principles: open.spotify.com/episode/4aH…
3. The practical example: youtu.be/J7N9FMouSKg
4. Learning Braintrust: open.spotify.com/episode/6AK…
5. Learning Arize: youtu.be/DL-pUGcfrf4
6. LLM judges: news.aakashg.com/p/ai-pm-llm…
7. Observability: news.aakashg.com/p/ai-pm-obs…
This is the most important skill for PMs to learn anew in this AI era. These are the 7 resources to learn it.
Send them to a friend.
#AI #evals #PM
5
10
3,696
Talk to AI through AirPods and more than 30% of your words come back wrong, at least in one YC startup's tests.
That number explains something odd about their customers. Over 60% of their early users already own AirPods and bought a second pair of earphones anyway.
You've felt this even if you've never measured it. You say a sentence to your phone, it writes down something else, you fix it with your thumbs, and by the third try you give up and type.
Their own earphones got the error rate down to around 5% in the same tests. On a 100-word prompt that's roughly 5 words to fix instead of 30, which I think decides whether you keep talking or go back to the keyboard. The models keep improving and most of us are still feeding them through whatever earbuds we already had.
The founders found this out by accident. They're two friends of 20 years who shut down a startup that had reached $250K in revenue to build AI earbuds with a camera. Getting a first version made took pitching more than 120 manufacturers before 4 agreed to work with them. Then early users barely wanted the camera and loved the voice isolation, so the camera came out. I'd bet that was a hard call, since the camera is the part that demos well.
Later an early user asked how you'd control an AI agent from a pair of earbuds, and that question became a smart ring. It has physical controls and a glyph display, so you can direct agents like Codex or Claude by voice. There's a whisper mode too, for a library or a co-working space where you'd feel weird talking to your AI out loud.
The company is Antimattr and the product is called Project Mnemosyne. They're already running industrial pilots on factory floors, where technicians talk to agents over machine noise. If 5% holds up there, a coffee shop should be easy.
If AI ends up doing most of the typing for us, being heard right the first time is going to matter a lot more than typing speed.
🤝 Paid partnership
we shut down a software business making $250k a year to build a hardware company
put $20k of our own money into it, moved to the US with less than 3 months of runway, and eventually got into @ycombinator
today, we're introducing @antimattr_one
we're building a voice-first hardware interface for human-agent interaction
You know the coworker who jumps into every Slack thread, even the ones nobody asked him about?
12 frontier AI models just got tested for exactly that. They jumped in 4 times out of 10 when the right move was to say nothing, and butting in ran about 1.8x the cost of keeping quiet.
Researchers at @deepnotehq made up a whole software company for this and called it PageMantle. It's got six years of history and 10 GB of records sitting in 11 tools like Slack, Linear, Stripe and Snowflake. Then they put each model in the company Slack and ran 50 situations past it.
When somebody did need help the models came through 9 times out of 10, so that side looks pretty good already. It was the 10 situations where silence was the right call that tripped them up.
The 1.8x comes from digging. When the help wasn't needed, the agents made about twice as many tool calls before they posted. With six years of company data there's always one more place to look, and they only get 3 minutes.
An always-on agent spends most of its day in threads that don't need it. That's been true of every Slack I've worked in, anyway. So staying quiet is the call it makes most often, and right now models get that one wrong 40% of the time.
Making the model more cautious doesn't solve it. Gemini Argon never once helped out of turn, and it also stepped in only about half the time people needed it.
Astra and Sol from @OpenAIDevs led on overall score, for what it's worth. But if I were picking a model for an agent that's on all day, the first number I'd ask for is how often it stays quiet.
🤝 Paid partnership
What’s the best model for proactive agents?
Today, we’re introducing the Agent Initiative Benchmark. We tested 12 frontier models on realistic workplace tasks inside a synthetic company with 6 years of highly detailed history: it has a financial model, CRM records, past-due invoices, Linear tickets, and real-life events affecting company development - competitor product announcements, SVB collapse, etc.
We used the company’s Slack to test how a proactive agent listening in on each conversation would assist the team in day-to-day tasks: from stepping in to align initiatives across teams to warning the team not to use an unapproved customer logo.
Highlights:
- 6.1 Sol and Astra from @OpenAIDevs beat Anthropic's latest Fable 5.2, with Sol crushing it at both performance and cost ($0.11 per task vs $1.87).
- The most expensive models aren't the best: Fable 5.1 is 430x more expensive than GPT-6 Luna, with similar performance.
- Most models try to help even if they're not needed. Interestingly, in these failure cases, they spend more time thinking, and the unnecessary work takes about 80% more time and money than if they stayed quiet.
If you’re building agents, keep reading for more tips on which model is best for your use case. Detailed breakdown, the open-source dataset, and benchmark code below:
If you've watched Star Wars on Disney+ or Blu-ray, you've never actually seen the movie that came out in 1977. George Lucas went back and changed it, and the original hasn't had a real theatrical release in over 40 years. On February 19, it's coming back.
The changes started with the 1997 Special Edition. Lucas re-edited the cantina scene so Greedo shoots at Han first, and he added a CGI Jabba the Hutt scene that was never in the original. When a reporter asked him in 2004 why he didn't just put out both versions, he said the Special Edition was the one he wanted out there, and the other movie was "on VHS, if anybody wants it."
Back in 1988, Lucas went to Congress to fight Ted Turner's plan to colorize old black-and-white movies. He said people who alter classic films for profit are barbarians, and he warned that someday old negatives would get "replaced by new altered negatives."
Nine years later, he did pretty much exactly that to his own movie. The Star Wars negative was in rough shape after two decades of use, so the restoration team took it apart, washed the different film stocks separately and put it back together. While it was apart, they cut the new effects shots right into it. So the master copy of the 1977 film turned into the Special Edition.
When the Library of Congress picked the original for the National Film Registry in 1989, Lucas wouldn't give them a copy. As of 2014 the Library still didn't have one anybody could watch.
What brought it back was an old film print sitting in London. Most American prints from 1977 were on Eastmancolor stock, which fades to a pinkish brown over time. The British prints used dye-transfer Technicolor, and that stuff barely fades at all. The BFI had one in its archive and screened it in June 2025. Kathleen Kennedy, who was running Lucasfilm at the time, got on stage to tell the crowd it wasn't an illegal screening. Six months later Lucasfilm confirmed the 50th anniversary release would be the original 1977 cut, with Han shooting first.
Today they announced the full rollout. It's playing on thousands of screens around the world in IMAX 70mm, regular 70mm, 35mm and 4DX, and it's the first time the original has ever been shown in IMAX.
To be fair to Lucas, his 1988 speech was mostly about directors' rights against studios, and he always saw changing his own movie as his call to make. He also told Congress that people "yet unborn" should get to see a generation the way it saw itself.
On February 19, anyone born after 1997 finally gets that chance.
A galaxy far, far away returns to theaters in celebration of 50 years.
Experience Star Wars (1977) in theaters February 19, 2027.
Art by @Cakes_Comics
To get all my takes without an algorithmic filter, subscribe to my newsletter:
aibyaakash.com
945
The $8 popcorn you buy at the movies makes the theater about as much money as your $12 ticket.
Tom Cruise got asked his favorite snack on set and said popcorn, and that one-word answer is pretty much the whole business model of a movie theater.
Pull up AMC's numbers from last year. It sold $2.65 billion in tickets and $1.67 billion in food and drinks. The studios took $1.28 billion of the ticket money off the top, and the food only cost $327 million to make, so after all that the tickets left AMC with $1.38 billion and the snack bar left it with $1.34 billion. Per person that's $6.28 kept on a $12.09 seat and $6.13 kept on $7.62 of snacks.
It all comes down to who gets a cut. AMC handed the studios 48 cents of every ticket dollar last year, and nobody gets a cut of the popcorn (you're paying for kernels, oil, salt and a bucket). There's an old line in the industry about this. Find a good place to sell popcorn and build a movie theater there.
The snack profit alone covered AMC's entire $887 million rent bill last year with about $450 million to spare, and the company still finished the year $632 million in the hole. Take the popcorn away and that loss is closer to $2 billion.
Which is why it's kind of perfect that Cruise is the guy saying it.
He kept Top Gun: Maverick off streaming through the whole pandemic, it made nearly $1.5 billion, and Spielberg told him he might have saved theatrical distribution. Before Dead Reckoning came out he shot a promo telling you "I love my popcorn. Movies... popcorn." For The Final Reckoning, AMC's food and beverage team says he helped guide the design of the popcorn bucket.
He produces his own movies, so he gets a piece of the ticket money and not a cent of the popcorn money, and the popcorn is what he keeps telling you to buy.
Kevin Hart bet $200,000 on the worst hand at the table. Then he told the poker pro holding pocket queens exactly what he had.
He had ten-three. He told Sam Kiki he didn't have a king, and he was telling the truth.
That sounds insane until you know how pros think.
Mike Caro, the guy who basically wrote the book on poker tells, has a rule every serious player learns early, "weak means strong, strong means weak." Most people at a poker table act the opposite of their hand without even thinking about it. A guy who's really beat goes quiet and prays you check. So pros learn to flip whatever you show them.
So when Kevin says "I don't have a king," Kiki's brain flips it to "I have a king and I'd love a call."
Now think about who's saying it. Kevin has been getting paid to make people believe him on camera for 25 years. Anything he says at that table, a pro is going to treat like a bit. Kiki has to figure out whether a professional actor is acting, with $200,000 riding on the answer.
The part I love is that Kevin has been on the other side of this exact move, against the same guy.
Last season on the same show, Kiki had pocket sevens in a $424,500 pot. Kevin was stuck in the tank, saying he didn't think Kiki had anything. So Kiki offered to let him see one of his cards. Kevin flipped it over and it was a seven. That was a real card from Kiki's real hand, and it still felt like a trick. Kevin called and lost the whole pot.
Kiki holds the record on High Stakes Poker for the most money won in a single season. Reading people is his entire edge. That's what makes the truth so nasty against a guy like him, because the better you are at sniffing out lies, the longer you'll spend hunting for one that isn't there.
Watch his face while he tries to figure out if Kevin Hart is joking.
To get all my takes without an algorithmic filter, subscribe to my newsletter:
aibyaakash.com
7,688
There's a solar plant in the Chinese desert that keeps making electricity at 3 in the morning, and its battery is 30,000 tons of fertilizer heated past 1,000 degrees.
It sits outside Dunhuang. More than 12,000 mirrors, each one about 1,250 square feet of glass, all aim at a single target on top of an 860-foot tower.
When you spot one of those giant circles of mirrors from above with a white-hot dot in the middle, that dot is the target. At full sun it's taking in 630 megawatts of heat, roughly 700 times the strength of normal sunlight.
The aiming is ridiculous. Every mirror has its own motors and follows the sun all day, and the farthest ones sit almost a mile from a target about 50 feet wide. If a beam drifts a tenth of a degree out there, it lands about 9 feet off.
The heat goes into salt, 60% sodium nitrate and 40% potassium nitrate, which is basically what farmers spread on their fields. Pumps push it up the tower at about 550°F, it comes back down around 1,050°F, and it drains into an insulated tank 129 feet across. When the grid wants power, that hot salt boils water, the steam spins a turbine like any coal plant would, and the cooled salt goes back to a second tank to get reheated tomorrow. Dunhuang can keep running for 11 hours after sunset.
So why would China bother? Its solar panels passed its coal plants in capacity this summer, 1,286 gigawatts to 1,285, and every one of those panels drops to zero around sunset, pretty much all at once. Most grid batteries are built for 2 to 4 hours, which doesn't get you to breakfast.
China barely has any of these towers, though. Power from them runs about 8 cents a kilowatt-hour, a lot more than panels, so the country has roughly 2 GW of them next to more than 1,280 GW of panels. Beijing wants 15 GW by 2030, and back in June CGN broke ground on a 350 MW plant in Golmud built to hold 15 hours of heat.
About a third of China's power capacity now shuts off every evening, and the plants that can run all night on melted fertilizer still add up to a rounding error next to it. That's why Beijing keeps paying 8 cents a kilowatt-hour to build more of them.
Babies are contagious.
When a woman's coworker has a baby, her own odds of getting pregnant roughly double over the next year. Researchers in Germany tracked this inside the same companies, and the effect fades after about two years.
So when a 30-year-old says not one of his friends has kids, that's a bigger deal than it sounds.
Friends work the same way. Two sociologists, Nicoletta Balbo and Nicola Barban, followed 1,700 American women from high school until around age 30. When a high school friend had her first baby, each woman got more likely to have one too, and the effect peaked about two years later. A Dutch study found the same effect jumping from siblings to coworkers and back. When they ran the numbers on a world without the coworker effect, pregnancies dropped 5.8%.
It makes sense when you think about it. Having a kid is the scariest, most permanent decision most people ever make. Watching a close friend go first is basically a free preview. You see she's still okay and still coming to dinner with the stroller parked next to the table. Then you've got someone to do it with, which is honestly how most parents of toddlers manage to keep any friends at all.
Now flip it around.
If nobody in the group goes first, nobody gets the preview. Everyone waits another year, so nobody has a baby that year either, and everyone waits again. Low birth rates start feeding themselves.
And people barely hang out anymore. American Time Use Survey data shows time with friends dropped by more than half between 2003 and 2024. Friday night now looks like any other weeknight. If you work from home, the coworker with the newborn is a Slack status you scroll past.
You can find two 30-year-olds in the same city with similar jobs, and one will tell you all his friends have kids while the other can't name a single one. They're both telling the truth. Babies come in clusters, and whichever cluster you're in decides what feels normal.
The US fertility rate hit a record low of 1.57 in 2025, according to early CDC data. Birth rates for women over 30 actually went up, but not enough to make up for the drop among women under 30.
Birth rates behave like a network effect, and right now ours is running in reverse.
A humpback whale is swimming in the Hudson River, and you can thank a sewage plant in Harlem for it.
Up until 1986, New York dumped raw sewage straight into the Hudson every day. That March the North River plant at 137th Street switched on, and the daily dumping stopped for the first time in the city's history. The 1972 Clean Water Act is what pushed New York to finally build it. They even put a state park on the roof, Olympic-size pool included, and plenty of people swimming laps up there have no clue what's under them.
Humpbacks go wherever the food is.
Around New York that means menhaden, an oily little fish that fishermen call bunker. Menhaden live on plankton, and raw sewage sucks the oxygen out of the water that whole food chain needs. Once the river got cleaner, the bait fish had somewhere to be, and the whales started following them into the harbor.
Then the fish got hammered from the other side. Their numbers fell about 75% between 1985 and 2008, and roughly 80% of the catch went to a single plant in Reedville, Virginia that ground them into fish oil and animal feed. In December 2012, East Coast fishery regulators crammed into a Best Western conference room in Baltimore and voted to cut the catch 20%, the first coastwide cap on menhaden ever. When a Virginia fisherman got up and said the fish were fine, the room booed him and the chairman threatened to clear it.
Whale watchers logged five humpback sightings around New York in 2011.
This August, a Staten Island nonprofit called Gotham Whale cataloged its 500th individual humpback. They ID each one by the pattern under its tail, which is as unique as a fingerprint.
Warmer water up in New England and Canada is pushing some of them south too, to be fair.
Still, a whale in the Hudson is what 50 years of boring sewage law looks like when it finally pays off.
To get all my takes without an algorithmic filter, subscribe to my newsletter:
aibyaakash.com
2,734
Scientists had 26 adults in masks and sunglasses carry a giant T through two narrow doorways without saying a word, and the group did worse than one person working alone, while a crew of 80 ants got it through.
Getting ants to care about a geometry puzzle was the easy part. The researchers left the T in a bag of cat food overnight, so as far as the ants knew, they were hauling dinner home.
This was Ofer Feinerman's lab at the Weizmann Institute in Israel, and the puzzle comes straight out of robotics. It's called the piano movers problem, which is basically the couch-through-the-doorway fight every family has had, done with math. They built an ant-size maze and a human-size one, then ran longhorn crazy ants (about 3mm long) in groups of 1, about 7 and about 80, and people in groups of 1, 6 to 9, and 26.
No single ant has any idea there's a puzzle.
Each one just wants the food home, and pheromones can't tell an ant how wide a doorway is or that the load needs to turn sideways. What happens instead is kind of beautiful. Once 80 of them are pulling, the load picks up a direction and holds it, and that shared direction works like short-term memory for the whole group. When the T hits a wall, they slide along it until it finds the slit, without circling back to retry the dead end they just left. No individual ant remembers a thing, so the memory ends up living in how the whole pile is moving.
People go the other way. One person alone sizes up the maze, plans the turn and beats a single ant. Add more people and it stops getting better, and once the 26 couldn't talk or even see each other's faces, they drifted to the most obvious move, the one everybody could agree on without a word. The researchers called those moves greedy, meaning they feel like progress right now and box you in a few moves later.
Feinerman put it simply. Every person who walks in understands the maze and shows up with their own idea of the right sequence of moves, and those differences are exactly what make it hard for the group to agree. With no way to talk it out, 26 different plans shrink down to whatever nobody objects to.
Each of those people had around 86 billion neurons to work with, and the ants got the T through with brains a fraction the size of a pinhead, because none of them had a plan to defend.
Every meeting where nobody wants to be the first to disagree is this exact experiment, run with the masks off.
Pixar is now borrowing its look from the studios that spent 20 years copying Pixar.
Its next movie, Gatto, is a mob story about a black cat in Venice, and the whole thing looks hand-painted. DreamWorks bet on that kind of look with its 2022 Puss in Boots sequel, which opened to a soft $12 million and kept selling tickets for months until it hit $485 million worldwide.
Here's how Pixar ended up here. Toy Story was the first fully computer-animated feature back in 1995, and for the next couple decades every big studio chased that same clean, shiny finish. That worked great for Pixar for a long time. The problem shows up once every cartoon at the multiplex has the Pixar look, because then a new Pixar movie stops standing out in a 30-second trailer.
Sony made the first big break from it. Spider-Verse put comic-book ink lines and little halftone dots on top of 3D animation in 2018, and it beat Pixar's own Incredibles 2 for the Oscar. DreamWorks followed with that Puss in Boots movie and then The Wild Robot.
Now look at what Pixar's own box office has been telling it. Toy Story 5 opened to $160 million this summer and passed $1 billion, because nobody needs a trailer to want more Woody. Elio, an original, opened to $21 million last year, the worst start Pixar has ever had. An original movie lives or dies on its trailer, and Pixar clearly decided this one has to look like nothing else it's made.
Even the release date is borrowed from a win. Gatto opens the same early-March weekend where Hoppers did $46 million this year, Pixar's best opening for an original since Coco. The director, Enrico Casarosa, made Luca, which got sent straight to Disney+ during the pandemic, so this is the first movie of his that gets a real shot in theaters.
So the studio that made every other studio want to look like Pixar is betting its next original on looking a little more like a cat in boots.
this cat’s on his last life
see the new trailer for Disney & Pixar's #Gatto, arriving in theaters March 5, 2027
Gatto | Official Trailer | In Theaters March 5, 2027
Microsoft shut down the studio that made Fable, then handed the whole series to a company that had only ever made racing games.
These four minutes are what they've been building for almost 10 years, and the whole story starts with a $75 million mistake.
Back in 2012, Lionhead, the British studio that created Fable, pitched Microsoft a big new single-player Fable 4. Microsoft passed. League of Legends was making a fortune at the time, and Microsoft wanted Fable to become a free online game that made its money from in-game purchases.
So Lionhead built Fable Legends, a 4-vs-1 online battle game, even though they were known for single-player games and had never made anything like it. It got pushed from 2015 into 2016, cost roughly $75 million, and never came out. Microsoft cancelled it and closed Lionhead that spring.
Keep in mind this was the team behind three Fable games that sold more than 11 million copies.
One former Lionhead developer told Eurogamer that Legends should have been dirt cheap to make, since that's the entire point of a free game. You put out something small, watch what people like, and build more of that. Legends ended up costing as much as a full-price game.
The next year, Microsoft gave Fable to Playground Games, the Forza Horizon studio. Playground set up a second team to learn how to make an RPG from scratch while the main team kept making racing games, which they're incredibly good at, by the way. Forza Horizon 6 came out in May with a 92 on Metacritic.
Fable was supposed to launch in 2025 and got delayed twice. It's now set for February 23, 2027, which puts that second team at close to a decade on one game.
So when Asha calls it a long road to bring this franchise back, she's underselling it. Xbox even admitted in its own 2021 documentary that closing Lionhead was a mistake.
And look at what she posted. You walk into a corrupted forest with a local hero named Arden, fight with a sword and magic, and end up at a tomb. That's a single-player Fable, which is pretty much what Lionhead was asking to make back in 2012.
It just took a $75 million detour and a racing game company to get there.
Today we're sharing four minutes of new @fable gameplay. It's been a long road to bring this franchise back, and I'm proud of the work Playground Games is doing. Take a look.
2
22
8,382






















