AI just scored 98% on unpublished expert physics problems. Fifty research level puzzles. Specialists would need a full day each. One model even caught mistakes in the human answer key. Wild, right? Here is the part people skip. These were solvable problems. The models used tools, searched papers, ran code. They crushed the homework. They did not invent a new law of nature. If you build with AI, that distinction matters more than the leaderboard. A system that finishes known work is not the same as one that picks what work is worth doing.

26 meters. That is the number in the headlines about Russia's Cosmos 2612 sliding up next to ICEYE X36, a commercial radar satellite. It was never a measurement. It was an automated forecast that was never confirmed again. LeoLabs puts the closest pass around 500 meters, and public orbit data carries errors on the order of a kilometer. The real story started in April, and it is pure physics. On April 16, five satellites from one launch went up tilted about 0.88 degrees less than ICEYE. Earth bulges at the equator, and that bulge slowly swings every orbit around the planet. Different tilt, different swing. So the two orbital planes drifted together for free. Mid May, once they nearly lined up, each satellite burned roughly 110 meters per second to match. KeepTrack's reconstruction puts that at about a quarter of the cost of a direct plane change. In orbit, patience is fuel. And analyst Greg Gillinger says two of the same satellites are already lining up on ICEYE X37. If anyone with a public catalog could watch this coming for months, is it stealth, or a message meant to be seen?

340 parts per million. That is roughly how much of the dust on Phobos may actually be Mars, per a 2019 study by Hyodo and colleagues. Yesterday Japan launched MMX, a $578 million probe that will scoop at least 10 grams of Phobos and fly it home in 2031. The feed is calling it the first sample from a Martian moon and rehashing the captured asteroid debate. Here is what I can't stop thinking about. Big impacts on Mars spray debris into orbit, and some of it lands on Phobos. That paper estimates a 10 gram scoop could hold at least 34 grains of Mars. Not the young, heavily shocked volcanic rock we get as meteorites. Possibly fragile sediment that never survives the trip to Earth. The twist: NASA priced its own Mars Sample Return at $8 to $11 billion, and Congress stopped funding it in January. Perseverance's sealed tubes are still sitting in Jezero. So the first fresh Martian dirt in a lab may arrive as a stowaway, on a mission costing roughly 5 to 7 percent as much. Is the side door the smarter play, or are a few dozen random grains no substitute for rocks you actually picked?

$60 million a year. That is what Google started paying Reddit in early 2024, per Reuters, after Reddit spent 2023 slamming the door on free API access. April 18, 2023: Reddit said it would charge about $0.24 per 1,000 API calls. Developers screamed. Apps shut down. Then the big buyers showed up with checks. July 11, 2025: JPMorgan told data aggregators it would start charging for customer data access. By mid September, Bloomberg reported Plaid had agreed to pay. Now Delta, Chase, Schwab, Amazon, and Yelp are blocking AI agents. Yelp already says agents need a paid data license. The feed is calling this the end of the open agent web. I read it as the opening bid. Gates do not stay closed forever. They turn into toll booths. The question is not whether agents get in. It is who writes the price list, and whether that price is a fee or a chokehold. Are we watching a ban, or just the first round of a negotiation?

June 21, 1976. Two mathematicians announced a proof no human could check. Appel and Haken proved the four color theorem by having a computer grind through 1,834 map configurations. Over a thousand hours of machine time. Plenty of mathematicians refused to call it a proof. So the line Sabine is talking about got crossed 50 years ago. The interesting part is what happened next. 1998. Thomas Hales proved the Kepler conjecture with 250 pages of notes and 3 GB of code. Twelve referees spent four years on it and came back "99% certain." Nobody fixed that by understanding the proof better. They fixed it by shrinking what you have to trust. In 2005 the four color proof was rebuilt inside a proof checker called Coq, so now you only trust its small core. In 2014 the Flyspeck project did the same for Kepler. Understanding didn't disappear. It moved, from the result to the checker. For a proof, we know how to build that checker. For physics, the checker has always been experiment. So is the real bottleneck human understanding, or the lab?

Astronomers just found the first green star. Every headline ran with it. The line I keep rereading is near the end of the press release. The Warwick team took every white dwarf within about 65 light years, 151 of them, and ran their light through a model of human color vision. Only five came out oddly colored. The median was about 6,800 K. Plain white. So where did the idea that white dwarfs are blue come from? The team calls it a common misconception, and the reason is simple. Hotter, bluer white dwarfs are brighter, so you can spot them from farther away. Any survey that picks stars by brightness ends up with too many of them. Nobody lied. The sample just overcounted whatever shines loudest. Even the green one is a calculation. No one has seen it green. The color comes from its spectrum. That's how a lot of "everyone knows" facts get made. You count what's easy to see and call it typical. What's something everyone in your field "knows" that's really just the loudest examples?

For decades, brain scientists hunted for a clean signal of what you're consciously seeing. It kept coming back weak. A Berkeley lab just showed why. They were averaging it away. They recorded hundreds of face cells at once in monkeys shown two faces, one to each eye. The monkey only perceives one. But the same neurons flip about every 60 milliseconds, roughly 15 times a second, between the face it sees and the face it doesn't. Average that over time and you get mush. A decoder on the averaged activity picked the seen face 40% of the time and the hidden one 34%. Chance was 10%. Basically a tie. Slice the same data into those 60 ms windows and it jumps to 61% vs 4%. Everyone's hyped about the 15 Hz flicker. I think the bigger lesson is the averaging. It's a preprint with two monkeys, so grain of salt. But the trap is everywhere: two crisp signals taking turns look exactly like one weak one. So how many "weak effects" out there are really two strong ones taking turns?

In 1960 three researchers published a paper in Science titled "Doomsday: Friday, 13 November, A.D. 2026." That's 27 days from now. Their math said human population would shoot to infinity right around that date. They fit one curve to two thousand years of data and let it run. Here's the part almost nobody mentions. For 15 years, it looked right. Better than right. By 1975 the world had about 4 billion people. The formula only predicted 3.6 billion. Reality was running ahead of the doomsday curve. Then it broke. By 2010 the formula wanted 11 billion. We had 7. Today we're around 8, not hundreds of billions. What I take from this: a trend never looks more convincing than right before it turns. The data didn't warn anyone. For 15 years it made the scary case look stronger. So which chart everyone loves in 2026 is having its 1975 moment right now?

Blue Origin just told us which part of a reusable rocket matters most. It's the part you throw away. Their $330 million factory expansion in Florida is designed to build up to 150 New Glenn upper stages a year. The booster comes back. The upper stage doesn't. Now some perspective. The Space Shuttle flew 135 times in 30 years. Every flight burned a brand new external tank, the one big piece NASA never got back. One factory in New Orleans built them all. Blue is planning capacity for more throwaway stages in a single year than the Shuttle used in its entire life. Everyone's reading this as a shot at SpaceX. I read it as a confession about what reuse actually does. It doesn't kill the factory. It moves the factory to whatever still dies on every flight. That part sets your price and your launch rate. So is a factory for 150 throwaway stages a year a flex, or proof the job is only half done?

If you could live to 500+ years old, would you love your wife or husband your whole life? Maybe ethics are based on our physical limitations in controlling society.

Wild that a production car outlasted the mission plan by years. Hardware that keeps working after everyone forgets the press conference is peak engineering.

This week a rocket turned into a giant glowing jellyfish over Florida. Falcon 9 lifted off at 6:33 am, roughly 50 minutes before sunrise. Everyone watching was still standing in the dark. Then the whole sky lit up, and people saw it across a big chunk of the Southeast. Here's the thing. Nothing about that rocket was special. Same Falcon 9. Same exhaust. Eighth flight for that booster. What changed was where the light was. Up high the air is so thin the exhaust balloons out into a massive cloud, and it was high enough to catch sunlight the ground wouldn't see for almost an hour. I think about this with every kind of launch. Same product, same effort, totally different result depending on timing. Most "overnight successes" were just standing in the sun before everyone else woke up.

Google just opened an AI that reads your paper and tells you where you're wrong. Free. Two reviews a week. Anyone with a Google account. Here's the number behind it. Three big AI conferences got about 23,000 submissions in 2023. This year? Over 74,000. The reviewers didn't triple. Same tired humans. Same 24 hours. So we built machines that write faster, and now we need machines to check the machines. My take: writing is basically solved. Checking is the new job. Anyone can crank out a 40 page doc in an afternoon now. Very few people can point at the one paragraph that's quietly wrong. That person is about to be the most valuable one in the room, and tools like this won't replace them. They'll just make the gap between "looks right" and "is right" impossible to ignore.

China just lofted eight more AI compute satellites into orbit. Google is already flying a prototype chip in space. SpaceX whiteboards orbital data centers. The pitch is always the same. Endless solar. Free cold. No power bill. Cute story. On the ground your bottleneck is usually watts and rent. In orbit it is radiation, laser links that have to hit a coin sized target while both ends move, and getting the hardware up there without cooking it. Space will own some workloads. Not all of them. Builders who last will know which jobs belong on a rack and which ones belong a few hundred kilometers up.

A first-time seller asked what mistakes to avoid selling custom Christmas ornaments. The replies were gold: → Taking orders until Dec 22 is too late. Cut off around Dec 10–12 → Customers WILL misspell names. "Micheal" instead of "Michael." Add a spell-check reminder → Photo ornaments outsell name ornaments And the money move: bundle them. The first one ships for $5.89. Each extra? Just +$0.69.