🌌 One mind-blowing science moment a day, and what it means for us. Space, AI, robotics, energy. Wonder first.
- Tweets14.1K
- Following2.7K
- Followers18.4K
- Likes18.3K
ALT What if an AI's rulebook read less like a manual and more like a moral compass? 🧭 OpenAI just rewrote its Model Spec, the public playbook for how its models should behave, around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. It also retires much of the detailed style guidance and adds new guidance on context, oversight and agency. What it means for us: as SI takes on bigger real-world tasks, it needs judgment for situations nobody scripted, like using what you've told it before while keeping its assumptions visible and easy to correct. Smarter machines, guided by human values. That's the future worth building. #AI #OpenAI
ALT What if your phone signal could make the weather app lie? 📡 Weather satellites read faint microwave signals from water vapor near 23.8 GHz, and 5G leakage can look the same. A new AGU study modeled a 2040 5G network, with up to 7x more towers than 2023 in some counties, and replayed Hurricane Ida's remnants. Forecast errors reached up to 10 mm of rain and 2°C in temperature. The good news: it's a projection, not a disaster yet. The authors say tuning 5G to limit leakage, especially near cities, can protect the forecasts we count on. Faster networks and sharper forecasts can share the sky. Let's build both. #5G #Weather
ALT What if Antarctica is quietly exhaling? 🧊 Researchers measured almost 22 square km of the Taylor and Wright Valleys and found CO2 emissions rose 5.5-fold from 2020 to 2025. The twist: unlike the Arctic, it doesn't seem to be microbes. It's likely carbon-rich brines seeping up from deep underground through soils that used to be frozen shut. What it means for us: there's a carbon source missing from our climate math, and similar leaks may run along the Antarctic coast. Better sensors and smarter models (hello, SI) can help us measure what we've been missing. You can't fix what you can't see, so let's keep looking. #ClimateChange #Antarctica
ALT Dignified tribute portrait of an elderly bald actor as a starship captain, with gold lettering saying "Goodbye, Captain", the years 1940–2026, a starship silhouette and a quill in ink suggesting a solemn farewell.
ALT Imagine discovering a third kind of magnet hiding in plain sight for decades. 🧲 For decades, physics recognized two flavors of magnetism: ferromagnets (your fridge) and antiferromagnets. Now a SLAC-led team under Daniel Jost reports in Science that manganese telluride is an altermagnet, a third phase first proposed in 2022. They fired left- and right-handed X-rays at it and caught a "chiral" spin signal, like a spiral staircase, that ordinary antiferromagnets can't make. Why care? Altermagnets act like antiferromagnets on the outside but can still carry useful spin currents, which could mean faster, denser and cooler memory chips and data centers. Nature still has new chapters. Keep reading. #Physics #Science
ALT What if your next security guard What if your next security guard never sleeps and runs on tokens? 🛡️ Google DeepMind just unveiled Gemini 4 Carbon, only two weeks after Argon. It hit 88.2% on SEC-bench Pro for finding real software vulnerabilities and 71% on CWE-bench for patching them, and sleeps and runs on tokens? 🛡️ Google DeepMind just unveiled Gemini 4 Carbon, only two weeks after Argon. It hit 88.2% on SEC-bench Pro for finding real software vulnerabilities and 71% on CWE-bench for patching them, and it ranks #1 on the Artificial Analysis Intelligence Index at 59. Koray Kavukcuoglu says Google's teams have watched it work for days, pushing through dead ends to root-cause subtle bugs. The catch: cyber defenders get it first, with wider access "as soon as possible." What it means for us: SI that hunts bugs before the bad guys do, so the software we rely on gets a lot harder to break. The best hackers of tomorrow might be guardians. Build the shield. #AI #Cybersecurity
ALT AI can ace olympiad math. Can it design a crankshaft? ⚙️ Xplore just launched RWE-bench, 20 real mechanical, electrical and aerospace tasks: model an engine part in CAD, route DDR and PCIe links on a crowded circuit board, run fluid simulations on turbine vanes and rocket nozzles. No AI judges, no vibes. Physics solvers check every requirement, and missing one means you fail. The scoreboard is humbling. GPT-6 Astra leads at 37%, Claude Opus 5.5 hits 29%, and five of 11 models scored zero. Four tasks beat every model across about 1,100 runs. But give Opus 5.5 five tries and it solves 70% of the tasks, which is already more than most of us could do with the same brief. This is exactly the kind of yardstick that makes progress happen. Today it's 37%. The next rocket engine might have an AI co-designer. #AI #Engineering
ALT What if you could talk again just by thinking the words? 🧠 Neuralink says an ALS participant in its VOICE speech trial hit 75 words per minute with a 5.8% word error rate on a 125,000-word vocabulary, in the very first week. For context, a 2023 Stanford speech implant reached 62 WPM at a 23.8% error rate, and natural conversation runs about 140 WPM. Dr. Nader Pouratian of UT Southwestern, the trial's lead investigator, presented the result at the 2026 NEALS meeting. It's one participant, the numbers are company-reported, and the device is still investigational, so the road is long. But week one looks like a door cracking open for people who've lost their voice. Silence doesn't have to be the last word. #Neuralink #ALS
ALT What if the fastest way to teach an AI is to slip it a cheat sheet? ⚡ Stanford's Herman Brunborg just smashed the NanoGPT Speedrun, training a GPT-2-level model on 8 H100s in 6.702 seconds. The old record was 39.9 s. That one 83% jump beats every record since 2025 combined, and back in May 2024 the same task took 45 minutes. The trick is retrieval-augmented n-gram "hints." While the GPUs train, idle CPUs build a 37B-token hint corpus and feed the model 6-, 10- and 20-gram clues about what comes next. Fun twist: an AI (Opus 5.5) wrote nearly all the code, while the ideas were human. His advisor Tatsu Hashimoto called them "completely absurd improvements." For us, that means cheaper, faster models, built by people and SI as a team. When humans dream and AI builds, even the stopwatch can't keep up. #AI #LLM
ALT Imagine unrolling a ~750-lb solar array like a giant snap bracelet, in orbit. ☀️ This morning NASA's Anil Menon and the CSA's Joshua Kutryk stepped outside the ISS for US Spacewalk 100 to install the station's 7th roll-out solar array (IROSA) on the port truss. Menon rides the Canadarm2 robotic arm holding the array, and Kutryk is making the first Canadian spacewalk in 7 years. No motors needed: once released, the array unrolls on its own in about 10 minutes. With all 8 installed, station power climbs from 160 kW to about 215 kW, roughly 30% more juice for science through the station's final years before its planned 2030 deorbit. The 8th and final array goes up Nov 5. Every watt up there powers a discovery down here. Keep looking up. #Spacewalk #NASA,
ALT What if the solar system has a new sibling we've been staring past for a decade? 🪐 The Vera C. Rubin Observatory just flagged 2026 ED138, a bright dwarf planet candidate about 46 times farther from the Sun than Earth, beyond Pluto's ~40. It's an estimated 600 to 1,500 km across, which would make it one of the biggest candidates ever found, and its orbit is tilted about 25 degrees, likely courtesy of Neptune's gravity. Why did it hide so long? It sits in a part of the sky crowded with Milky Way stars that drowned it out. Harvard–Smithsonian's Matthew Holman and Kevin Napier ran billions of Rubin data points through their search code, then found sightings going back to 2016. "My first reaction was one of excited disbelief," Holman said. Next step: proving it's round enough to join Pluto, Ceres, Eris, Haumea and Makemake. Rubin only just opened its eyes, and it's already redrawing the map of home. Keep looking up. #RubinObservatory #Space