Signals from the Fringe of Science & Technology
DeepSeek V4 achieves 97% needle-in-haystack accuracy at 1M tokens using Hybrid Attention that cuts KV cache memory 90% and inference FLOPs 73% — purely architectural, no hardware changes required.
Bradford and RIT researchers found AI produces identical 'conscious-like' signals even when cognitively degraded — proving the signals are complexity artifacts rather than genuine consciousness markers.
Tufts University researchers combined neural networks with symbolic reasoning to build a hybrid AI that uses up to 100x less energy than standard deep learning approaches.
Running 200,000+ simulated conversations, Microsoft and Salesforce found LLMs drop 39% in performance during multi-turn dialogues — not from running out of context but from reliability collapse at conversation turn boundaries.
Stanford study had 43 experts each spend 100+ hours executing LLM-generated vs human research ideas; LLM ideas rated as more novel before execution but fell sharply after — effectiveness dropped 1.879 points vs 0.052 for human ideas.
Meta's V-JEPA 2 learned physics purely from 1 million hours of unlabeled video and transferred to real robot arms zero-shot using only 62 hours of footage — no task-specific training, no reward signal needed.
Google DeepMind unveiled Aeneas in July 2025, trained on nearly 200,000 Roman inscriptions capable of predicting missing words in damaged Latin texts, estimating when and where they were created, and revealing historical links between texts across the Roman Empire. Similar AI tools have now translated hundreds of thousands of Hanja articles in months and decoded the first words from the Herculaneum scrolls buried by Vesuvius. The tools surface hypotheses, not ground truth: scholarly verification is still required.
Researchers at Ohio State University created shiitake mycelium-based memristors operating at ~5,850 Hz with ~90% accuracy, while Cornell built biohybrid robots embedding mycelium sensors that respond to light and chemical stimuli. The roadmap projects 2025-2028 as the era of enhanced fungal-computing prototypes for edge cases like remote sensors, space hardware, and eco-bots. Serious hurdles remain: environmental sensitivity, speed limitations, and difficulty bridging living tissue to conventional electronics.
Researchers at Emory University used a physics-tailored neural network to analyze dusty plasma (the fourth state of matter) and discovered non-reciprocal particle forces with over 99% accuracy. The AI overturned a decades-old assumption: particle size DOES affect how quickly inter-particle forces weaken, and a leading particle attracts the trailing one while the trailing particle always repels the leading one. The framework runs on a desktop and may generalize to biological many-body systems.
Sandia National Laboratories scientists published a paper in Nature Machine Intelligence showing neuromorphic chips can efficiently solve partial differential equations (PDEs) — math previously reserved for traditional supercomputers. For 12 years after the cortical network model was introduced, no one noticed its non-obvious link to PDEs; the new algorithm exploits that link and opens the path to the world's first neuromorphic supercomputer. This directly challenges the assumption that neuromorphic hardware is only good at pattern recognition.
Tufts University researchers combined symbolic reasoning with neural networks in a neuro-symbolic VLA for robotics tasks, achieving a 95% task success rate vs. 34% for standard systems, while cutting training energy to 1% and operational energy to 5% of conventional approaches. The approach mimics how humans solve problems by breaking them into steps and abstract categories rather than raw pattern matching. The paper will be presented at ICRA 2026 in Vienna.
MIT Technology Review named mechanistic interpretability — the science of mapping features and circuits inside large language models — as a top 2026 breakthrough technology. Anthropic research traced the entire path a model takes from prompt to response by identifying recognizable concept-features and the weighted pathways between them. Foundational challenges remain: feature still lacks a rigorous definition and many interpretability queries are computationally intractable.
Clear search or tag filters to see all transmissions on this page.