Science

A preprint from QuEra, Harvard and Anthropic reports lower gate and qubit requirements across seven of eight quantum circuit building blocks.

A 27B terminal agent improved on reported benchmarks through five human-directed training stages; the study does not show unsupervised self-improvement.

A Lean-checked proof of the Aaronson-Ambainis conjecture limits quantum advantages in input queries on most random inputs, not total computing time.

ToyBlocks lets researchers train and test robot arms in a controlled block-sorting simulation, but physical-world results have not been reported.

MTSU and university partners held a three-day beginner-focused quantum hackathon, with 250 registrants and four weeks of introductory training.

Berkeley startup SuperMatics says its AI models are designed to find superconductors that can be manufactured as wire for energy and computing uses.

EigenOrbit’s study evaluates its open-source eigenkit library across 12 quantum execution providers, finding performance depends on workload and hardware fit.

Researchers report that GPT-6 Astra changed two Qwen3.5-0.8B weight tensors, cutting benchmark loops and raising its MATH-500 score.

A source headline links Anthropic to progress on provable LLM inference, but provides no method, results or technical documentation.

An FAU study reports up to 94.31% accuracy for a quantum model, but synthetic data and limited testing leave clinical value unproven.