Signals from the Fringe of Science & Technology
Google trained a 400M parameter model on four decades of Atlantic spotted dolphin vocalizations, with the goal of enabling two-way dolphin-human communication.
AI systems working through the 340,000+ tablet CDLI database completed first translations of previously unread Babylonian texts, including a 250-line hymn and Herculaneum scrolls thought too charred to read. The newly surfaced content includes sophisticated economic records with credit instruments and commercial law, medical tablets documenting hundreds of conditions, and literary texts from the Iron Age. The AI is rewriting assumptions about the complexity and richness of pre-modern civilization.
China approved NEO by Neuracle Medical Technology for commercial use on March 13, 2026, making it the first BCI cleared outside clinical trials anywhere in the world. Eight electrodes decode imagined hand movements via AI and transmit commands to a robotic glove that restores grip. China has also designated BCI as a national strategic priority, signaling intent to lead globally in neural interface technology.
Published in Nature Neuroscience (2026), researchers trained adversarial neural networks on 680,000+ neuroelectrophysiology samples from 565 patients, validating predictions about basal ganglia and inhibitory cortical wiring across patient scans, RNA sequencing, and rat models. The model independently identified high-frequency subthalamic nucleus stimulation as a promising new therapy for disorders of consciousness. This is the first AI framework to produce independently confirmed mechanistic predictions about the neural basis of consciousness.
IISc researchers synthesized ruthenium-complex molecular devices whose function is tunable by adjusting surrounding ligands and ions without any rewiring. The same physical device can behave as memory, a logic gate, a selector, an analog processor, or an electronic synapse depending on how it is stimulated. The team is now integrating these onto silicon wafers, creating neuromorphic hardware where learning is encoded in the material itself rather than in software weights.
Published in Nature Computational Science (2026), MIT DiffSyn was trained on 23,000+ synthesis recipes from 50 years of scientific literature and uses a diffusion AI approach to propose novel synthesis pathways. Given a desired material, it outputs 1,000 candidate recipes in under a minute, replacing the historic one-to-one structure-to-synthesis mapping with a one-to-many model. The team validated it by synthesizing a novel zeolite using DiffSyn-suggested pathways.
University of Cambridge researchers built a nanoelectronic memristor from modified hafnium oxide that combines memory and processing in one location — just like biological neurons — eliminating the energy-wasting data shuttling of conventional AI chips. The device switches at currents roughly one million times lower than standard oxide-based memristors and supports spike-timing dependent plasticity, the biological learning rule used by real brains. The main obstacle to production is a manufacturing temperature of ~700 degrees C, which exceeds standard semiconductor fabrication thresholds.
Oratomic (launched March 2026, spun out of Caltech) used OpenEvolve — an open-source LLM-powered optimizer — to iteratively evolve quantum algorithms the way natural selection works, discovering that a single logical qubit can be encoded with just three atoms instead of the previous 100 to 1,000. The resulting breakthrough means a utility-scale fault-tolerant quantum computer may require only ~10,000 physical qubits, versus earlier estimates in the millions. The AI combined knowledge across niche quantum sub-disciplines in a way no single human expert would have explored.
Google partnered with Georgia Tech and the Wild Dolphin Project to train DolphinGemma, a ~400M-parameter model that learned the hidden structure of Atlantic spotted dolphin vocalizations from four decades of underwater recordings, distinguishing signature whistles, burst-pulse squawks, and courtship clicks. The model — small enough to run on a Pixel phone in the field — can generate novel dolphin-like sound sequences, and researchers are now deploying it toward a shared vocabulary for two-way interaction with free-ranging dolphins. Google plans to open-source DolphinGemma for adaptation to other cetacean species.
Sony AI Ace became the first known autonomous robot system to compete at an elite human level in a widely played competitive sport, defeating 13 out of 29 amateur and professional human players in real matches and winning rallies against world-ranked professionals. Unlike game-playing AIs, Ace must contend with physical unpredictability — ball spin, bounce variance, real-time opponent reads — making this a qualitatively harder class of problem than board games. The project demonstrates that physical-world AI can now close the gap with top human athletes.
UCL researchers built a hybrid where a 20-qubit quantum computer first identifies invariant statistical patterns in chaotic data, then feeds those patterns as structure into classical AI training. Tested on fluid dynamics and chaotic physical systems, the method delivers ~20% greater accuracy than standard AI while requiring hundreds of times less memory and maintaining stable long-horizon predictions. Published in Science Advances, with potential applications in climate modeling, transportation, and medicine.
Berkeley Lab Accelerator Assistant — an LLM multi-agent system — was the first to autonomously prepare and execute multi-stage physics experiments on a live synchrotron light source, cutting setup time by two orders of magnitude compared to expert manual scripting. Engineers issue natural-language goals; the system resolves variables, writes and runs analysis code, and safely controls accelerator hardware within operator-standard safety constraints. Published in Physical Review Research.