Signals from the Fringe of Science & Technology
UC Davis engineers combined 16 silicon nanostructure photodetectors with machine learning to build a spectrometer-on-a-chip that achieves 8 nm resolution in a grain-of-sand-sized package.
AI-powered social personas have become sophisticated enough to infiltrate online communities, coordinate messaging across accounts, and subtly steer public opinion at scale — unlike traditional bots, they adapt, learn, and refine strategy in real time.
NASA and Microchip Technology are testing the High Performance Spaceflight Computing (HPSC) chip, which delivers 500x the throughput of current radiation-hardened spacecraft processors — enabling onboard AI inference for rovers, probes, and landers without waiting hours for Earth commands.
Lab-grown clusters of human neurons — brain organoids — have been trained to control physical robotic systems, stepping out of a petri dish and into embodied real-world actuation.
An autonomous AI lab at Sandia found a way to steer light 4x more precisely over a wide field of view — by deliberately breaking the design rules of the field, finding a path physicists said they 'wouldn't have explored.'
An OpenAI reasoning model outperformed experienced emergency room physicians at diagnosing patients and managing their care — using only electronic health records from a Boston ER, with no physical exam data.
Sony AI's Project Ace became the first known autonomous system competitive with elite and professional-level human table tennis players in the real physical world — crossing the line from simulation into live competitive sport.
A March 2026 arXiv paper proposes a sleep-inspired mechanism that uses hippocampal-style memory replay during offline 'rest' phases to let LLMs overwrite stale associations without catastrophic forgetting — mirroring how human sleep consolidates memories.
Stanford Medicine's SleepFM model, trained on 585,000 hours of sleep-study data from 65,000 patients, can predict risk for 130 diseases — including Parkinson's (C-index 0.89), dementia (0.85), and heart attack (0.81) — from a single night of sleep.
Karlsruhe Institute of Technology researchers showed that ordinary WiFi routers, combined with AI, can uniquely identify people at 99.5% accuracy by reading unencrypted beamforming feedback signals — even through walls and without the person carrying any device.
MIT Technology Review put mechanistic interpretability — the effort to reverse-engineer AI models' internal circuits and find out which concepts live where inside a neural network — at the top of its 2026 Breakthrough Technologies list, signaling a field-wide pivot toward auditable AI.
A May 2026 study found that foundation models consistently exhibit emergent collaborative behaviors — task division, turn-taking, role specialization — without being trained to collaborate, with distinct behavioral signatures differing across LLM families.
Clear search or tag filters to see all transmissions on this page.