Signals from the Fringe of Science & Technology
A research team from the Institute for Basic Science, Université de Montréal, and NYU published a new analysis in Neuron warning that current scientific methods may not yet be capable of reliably determining whether AI systems are conscious — calling for more careful methodological standards.
Researchers from the Karlsruhe Institute of Technology, in collaboration with scientific partners, combined large language models with machine learning to build concept graphs from scientific papers — connecting terms that are mentioned together — and used the resulting structure to predict which combinations of scientific concepts could become more significant in the next two or three years, supporting researchers' creative thought processes by surfacing new avenues of research.
Inspired by neuroscientist Erik Hoel's 'overfitted brain hypothesis' — that dreams prevent neural overfitting by injecting noise — researchers argue that AI hallucinations are not bugs but underdeveloped features, and that deliberately engineered 'dreaming cycles' using synthetic scenario rehearsal could make LLMs more robust than current suppression-based approaches.
Worcester Polytechnic Institute built Saranga, a bat-ear-inspired acoustic shielding system + neural network that enables centimeter-scale drones to navigate 3D darkness using ultrasound — consuming 1,000x less power, weighing 10x less, and costing 100x less than standard drone sensors, while detecting obstacles as thin as a human hair.
MIT researchers drew on the fully mapped connectome of C. elegans — a millimeter-long worm with just 302 neurons that communicate via graded analog signals rather than digital spikes — to build 'liquid neural networks' that adapt in real-time without retraining and run on a single edge device.
NOAA deployed three operational AI-driven global forecast models in February 2026, including the world's first hybrid 62-member physical+AI grand ensemble (HGEFS) that outperforms pure physics models while a 16-day global forecast now uses only 0.3% of the computing resources of the traditional GFS system.
University of Pennsylvania physicists demonstrated all-optical switching in AI hardware using exciton-polaritons (hybrid light-matter particles) consuming only 4 femtojoules — eliminating the energy-burning electron-to-light conversion bottleneck in photonic AI chips.
Scale AI, DoorDash, and Chinese firms are hiring thousands of gig workers across 50+ countries to film household chores and wear exoskeletons to “puppet” humanoid robots — generating over 100,000 hours of training footage for the next generation of home robots.
Researchers demonstrated that combining traditional neural networks with symbolic reasoning — mimicking how humans decompose problems into steps and categories — can cut AI energy consumption by up to 100x while actually improving model accuracy.
Sony AI’s Project Ace became the first autonomous robot to defeat professional human table tennis players in real matches, achieving 20ms end-to-end latency vs 230ms for elite humans, with results published on the cover of Nature.
MIT Technology Review named mechanistic interpretability — the field that reverse-engineers exactly which computations inside a neural network produce a given output — its #1 breakthrough technology of 2026.
Harvard SEAS researchers published in PNAS that robot swarms reach destinations faster when given a small degree of random movement — too little causes deadlock, too much causes chaos, but the sweet spot produces fluid collective flow.
Clear search or tag filters to see all transmissions on this page.