Discoveries in the Fields, Space, and the Archives

 〰️ Teaching Robots Which Tomatoes Are Easy to Grab 🍅

Farm robots fail constantly because they can't tell which fruit is actually reachable. Now, some researchers have fixed this with image processing that predicts grab success before the robot even tries. The system spots obstacles like stems, leaves, and other fruit and calculates the odds. Now robots make smarter decisions, they say, instead of blindly reaching and failing. → More

 〰️ A New Pipeline Is Finding Planets We Missed

Astronomers built RAVEN, a system that automatically finds and validates exoplanets in satellite data. Machine learning trained on simulations separates real planet signals from stellar noise. It's already confirmed 118 planets, including dozens that previous searches missed. This tool allows scientists to better understand the distribution of planets across the galaxy. 🔭 → More

 〰️ AI That Imagines Earthquake Damage from the Sky

After a quake, getting ground-level damage views is slow and dangerous. Researchers built an AI called LEGG that takes drone photos and imagines what the destruction looks like from street level. It catches details that satellite images miss like tilted buildings, cracked facades. Tested on the 2023 Turkey earthquakes, it helps responders assess damage safely. 🏗️ → More

〰️ The First Global Rules for AI Testing

South Korean researchers created the first international standard for checking if AI is safe and reliable. ISO/IEC TS 42119-2 lays out how to test data quality, model performance, and bias. It introduces risk-based testing to catch problems like performance degradation or vulnerability to hacker attacks before deployment. Maybe this could be a foundation for future AI certification worldwide? 📜 → More

〰️ Tiny Drones That See with Sound, Just Like Bats 🦇

Inspired by how bats navigate dark and dusty caves, engineers built palm-sized drones that navigate using ultrasound instead of cameras. They work through smoke, fog, and darkness where normal sensors fail. Two small sensors plus deep learning let them process weak echoes and fly safely in rescue zones. This advance allows small drones to operate longer in cluttered environments while using very little power. → More

 〰️ Prehistoric Fish Heard Through Their Lungs

Researchers discovered that coelacanths, fish living 240 million years ago, used their lungs to detect sounds underwater. Using powerful synchrotron X-ray imaging, scientists found that an ossified lung transmitted sound vibrations directly to the fish's inner ear. This rewrites what we know about how ancient life sensed its environment. Those prehistoric organs had jobs nobody expected. 🐟→ More

〰️ Scientists Built the Hardest AI Test Ever

Nearly 1,000 experts created Humanity's Last Exam, a 2,500-question benchmark designed to be impossible for current AI models to solve easily. To ensure the test was tough enough, any question that a current AI model could answer correctly was removed during testing. Early results show top models like GPT-4o scoring as low as 2.7 percent on these highly specialized academic tasks. This exam serves as a durable tool to measure true expert-level knowledge versus simple pattern recognition. 🎓 → More and here.

〰️ Brain-Inspired Chips That Solve Math Problems

Researchers showed that neuromorphic chips (hardware that mimics the brain) can solve complex engineering math more efficiently than regular CPUs. Using Intel's Loihi 2, they tackled problems common in simulations. These chips use timing-based communication instead of binary code. Could mean big scientific computations with way less power. 💻 → More

〰️ AI Found Hidden Links Between Ancient Scripts

A deep learning model found structural similarities between Ethiopic, Armenian, and Georgian writing systems. The AI spotted symbolic patterns suggesting these scripts may share unknown historical connections. Armenian, for example, looks as similar to Ethiopic as it does to its own ancestors. This computational evidence supports theories of sustained intercultural contact spanning from Africa to the Caucasus Mountains. 🖋️📜 → More

〰️ AI Learns How Ancient Pottery Styles Evolved 🏺

Researchers use a deep learning model called PointTransformer to study 3D scans of ancient Japanese Sue ware pottery. The AI looks at point clouds (detailed 3D maps of a vessel’s surface). This helps it notice small shape details that older 2D methods missed. It gets about 93% accuracy. It also uses saliency maps to show which parts matter most, like the rim or inner base, similar to how experts look at pottery. This helps archaeologists clearly show how pottery shapes changed when people in ancient Japan moved from eating with their hands to using spoons and chopsticks. → More

✹ Over half of researchers now use AI for peer review despite bans, and ICML caught 497 researchers using AI to write peer reviews through watermarks.

✹ AI recruitment tools are biased even without gender data. They just find workarounds.

✹ Dow is cutting 4,500 jobs to pivot to AI and automation.

✹ Microsoft Research says AI will now generate hypotheses and run experiments alongside scientists.

✹ IBM says 2026 is when AI agents finally get a control panel.

✹ The UN is hosting a major summit (AI for Good Summit) in Geneva to push AI for healthcare, food security, and climate.

That´s it for today…Thank you for reading!

If you know anyone who would like to read these insights, please share this newsletter with them.

I would also love to get your feedback. Leave a comment to let me know how you think I can improve my newsletter issues or what you would like to see more of.

Stay curious and in the loop.