The Logic Stays Nearby

Researchers used seeded living crystallization-driven self-assembly to make two-dimensional poly(ε-caprolactone) platelets whose thickness sets their melting temperature. Multilayer platelets with programmed thermal-stability gradients selectively disassembled at 37 °C into uniform hollow structures within minutes; cooling drove spatially confined recrystallization, reversible reconstruction and self-healing of the platelet framework. This is nanoscale solution-phase self-assembly, not a load-bearing adaptive component, and the early-access record does not establish cycling count, long-term stability, payload behavior or bulk manufacturability.

TDK demonstrated a monochrome direct-retinal-projection display built around a flat meta-optic mirror embedded in an eyeglass lens. Nanoscale reflectors distribute position-dependent angles across a layer about 150 nanometers thick, steering incident laser light toward the retina while leaving the lens roughly 80 percent transmissive in visible light; TDK also says the arrangement prevents outward image leakage. The claims are company-reported. Field of view, eyebox, brightness, efficiency, resolution, eye-safety margin, yield and cost are undisclosed, while full color and mass production remain future goals.

KSHPV uses ordinary side-view video, lightweight pose estimation, take-off/peak/landing alignment and confidence-aware quality control to describe preparation, propulsion, coordination, flight and landing in a standing long jump. Across 204 participants and 219 analyzed trials, its random-forest model reported 17.25 ± 2.67 centimeters mean absolute error and R² 0.665 ± 0.068, compared with 28.92 ± 2.43 centimeters and R² 0.160 ± 0.091 for an anthropometrics-only baseline. The study covers one task and predicts distance; it does not establish better coaching outcomes or robustness beyond the reported participant strata and simulated recording noise.

Ai2 released AstaBrief 8B, an Apache-2.0 model derived from Qwen3-8B and trained with supervised fine-tuning plus offline preference optimization. Ai2 says quality filtering yielded 47,000 usable supervised examples and about 6,000 preference examples. A supplied workflow generates cited reports from local PDFs, and Ai2 reports a 51.1-second average full-pipeline time for its Fast mode versus 178.5 seconds for its Claude-powered Thinking mode. The comparison is Ai2's own, most evaluation was completed in 2025, and the human study covered 14 questions from three researchers. The training data is CC BY-NC 4.0, not Apache-2.0.

NVIDIA announced a 64-gigabyte DGX Spark configuration built around its GB10 system for availability from OEM partners on October 23, starting at $4,999. NVIDIA says one unit supports models up to 100 billion parameters, while two directly connected systems can reach 128 gigabytes of memory, models up to 200 billion parameters and as much as 1.7 times the performance. The hardware had not shipped when announced, and the capacity and performance claims are vendor-reported. Existing Cluster Assistant software configures networking and SSH rather than transparently turning multiple boxes into one system; a model launcher is still forthcoming.