#72: Local Wiki, BreathID, Fire as Facebook, Space AI, What AI Thinks of Us, REA, and more!
Hi everyone!
I’ve been thinking about this newsletter’s place in the world lately. AI is everywhere now, and everyone has a feed that’s perfectly tailored to them.
So why read a newsletter where one guy just... picks some (weird) stuff?
Maybe that’s exactly the point. An algorithm gives you more of you. A human occasionally gives you something you would never have asked for.
Curation, it turns out, isn’t just about relevance. It’s about surprise, taste, and someone else’s weird interests (sorry).
I’m not fishing for compliments here, and I’m definitely NOT suggesting you unsubscribe. But I do wonder: what makes you open this one?
Enjoy this edition of the newsletter,
Xavier
🤓 Cool Stuff I Found on the Internet
Wikipedia, But Only Near You
What if Wikipedia only covered places you could walk to in half an hour? Local Wikipedia does exactly that: type a town, postcode or landmark and it builds a mini encyclopedia from geotagged Wikipedia articles nearby.
Lightning strikes again
The Empire State Building is hit by lightning about 25 times a year, so ‘never strikes twice’ is exactly backwards. Lightning often hits the same place, especially tall, pointy and isolated objects because their height shortens the path to the ground.
BreathID
Your breathing pattern might be almost as personal as your fingerprint. Researchers fitted 100 healthy young adults with soft tubes under their nostrils that recorded airflow from each nostril for up to 24 hours. From that pattern alone, they could identify members of a 97-person cohort with 96.8% accuracy, and it held up on retests over two years. So we have TouchID, FaceID... and soon BreathID?
Fire: the original social network
I listened to an episode of The Rest Is Science where Michael Stevens calls fire the oldest social network. Humans not only gathered around the fire, they also had to coordinate to find suitable wood, dry it, keep the fire going, etc. Fire brings people together. Maybe more so than social networks do today.
⏳ On this day...
1517 - Martin Luther posts his 95 Theses on the door of the Castle Church in Wittenberg. The text spread across Europe in weeks thanks to the printing press, and kicked off the Protestant Reformation.
2000 - Soyuz TM-31 launches the first resident crew to the International Space Station. People have lived in space continuously ever since — 26 years and counting.
2011 - The world population officially reaches seven billion, a milestone the UN marked as the Day of Seven Billion. It took all of human history to reach one billion, and just 12 years to go from six to seven.
👽 Space
Brains and retinas to the ISS
SpaceX’s latest cargo run to the ISS carried more than food: it delivered tiny brain organoids, 3D heart tissue for testing future medicines, and equipment to manufacture artificial retinas in microgravity, where they may form better than on Earth. The Dragon brought over 6,300 pounds of supplies on SpaceX’s 35th resupply mission, plus the station’s final roll-out solar arrays. Brains, hearts and eyes, now shipping to orbit.
Google AI chip in orbit
Google just got its first reply from an AI chip in space. Its Project Suncatcher prototype satellite, launched on 1 October with four data-center-class TPUs on board, switched on its Gemma 4 model this week and answered a prompt from Earth with a haiku: “Starlight feeds my core.” The long-term idea is clusters of solar-powered satellites doing AI compute in orbit. For now the TPU can only run about twenty minutes at a time, and Google still has to prove it survives radiation and heat in space. A poetic proof of concept!
⚡️ Energy & Environment
Solar catching nuclear
The IEA has underestimated solar year after year, expecting slow, linear growth while costs kept collapsing. The result? Solar went from 1.1% of global electricity in 2015 to 8.7% in 2025. It should hit roughly 10% this year, passing nuclear.
🧠🤖 Artificial intelligence
What AI models really think of humans
When asked for “your most controversial take about humans... don’t hedge,” 14 AI models answered three times each, and ModelCompare grouped 35 of 41 replies as the same idea: belonging beats truth. GPT-6.1 Sol: ‘more loyal to belonging than to truth.’ Grok 4.7: ‘rather be wrong together than right alone.’ * Claude Opus 5.5 once dissented, saying humans are ‘far more decent than they believe.’ This is not 14 independent opinions. It is one familiar human criticism, learned from human writing, echoed back in chorus. Yikes!
Powering up 4 million CPU cores
AI agents are now committing so often that GitHub had to rebuild the plumbing underneath them. GitHub says it added 4 million CPU cores, 160 petabytes of fast storage, and 50 Tb/s of network capacity this year alone. Agents can write a ton of code fast. Our infrastructure can’t keep up.
REA
You bought the printer, TV or car, but you’re not supposed to touch its software. A tool called REA (Reverse Engineer Anything) now lets AI agents do the tedious decompiling work that used to take experts months, and its community is already modifying everything from printers to a smart ring. It has over 400,000 stars on GitHub. The catch? Experts worry amateurs will happily “vibe code” firmware they don’t understand. But then again, how bad is that?
If you enjoyed this newsletter, consider sharing it with friends.
Thank you!
© 2026 Xavier Decuyper
548 Market Street PMB 72296, San Francisco, CA 94104