World of Science & Technology

ALT Google DeepMind just used a new system called AlphaProof Code to formally verify key security properties of the Signal protocol, the cryptography behind billions of conversations in apps like Signal and Google Messages. Led by Pushmeet Kohli and Swarat Chaudhuri, Gemini Carbon agents generated over 6 million lines of Lean proof in just four weeks, a job that used to take expert teams years. This is the first time anyone, with or without AI, has formally verified the actual implementation of a widely deployed protocol like this, down to the compiled code running on devices. To check the system wasn't just rubber-stamping, they reintroduced two old, already-fixed bugs. AlphaProof Code proved both buggy versions were insecure. Follow @worldofscitech

ALT The Carbon in Earth's Heart Was a Long-Distance Delivery ☄️ What if the carbon at the center of our planet came from the far edge of the solar system? A team led by Columbia's Yves Moussallam, with Yale's Damanveer Grewal, studied tiny droplets of magma trapped in crystals from Hawaiian eruptions, including Kīlauea Iki in 1959. Each one is a message in a bottle from deep inside Earth. They found an unusually heavy carbon-13 signature that they read as "carbon leaking from the core." The best match: carbon-rich meteorites from the outer solar system, which likely arrived in the final stage of Earth's assembly, just before the core closed for business. It's the first observational estimate of the core's carbon fingerprint, published in PNAS

ALT Claude Learned to Dream, and Its Memory Got Sharper 🌙 What if your AI got smarter while it slept? Anthropic just shared how Claude uses "dreaming": idle time between sessions to reread its old transcripts, catch its own mistakes, and rebuild its notes into a linked wiki. Basically what your brain does every night, except Claude doesn't snore. The numbers are wild. On a company archive of about 350,000 files, scaling dreaming from 2 to 128 sessions lifted Claude's wiki-only answers from 13% to 48% of graded criteria. In Pokémon Red, old Sonnet 3.7 hoarded 413 memory files and earned one badge. Fable 5 kept nine files plus a single dashboard file and earned four. For us: SI assistants that remember what matters, fix what they got wrong, and stop making us repeat ourselves. Turns out the secret to a great memory is knowing what to forget. Sweet dreams, Claude! #Claude #AI

ALT NASA Is Hiring for a Job With a Moon View 🌕 Imagine your next commute ending at a Moon Base! NASA just opened applications for its next astronaut class, the people who could walk on the Moon, live and work at the Moon Base, and pave the road to Mars. Administrator Jared Isaacman calls it "an extraordinary time to join NASA's astronaut corps." The deadline is Dec. 3. You need U.S. citizenship, a STEM master's (or an equivalent path like a medical degree or test pilot school), and three years of experience. Pay: $164,395 a year. Last round, more than 8,000 people applied and 10 were picked. Final selections come in late 2027. For us: the crew who will build humanity's first off-world neighborhood is being chosen right now. It might be someone you know. It might be you! Shoot for the Moon. This time, literally. #NASA #Artemis

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