Anthropic researchers used Claude to analyze DNA datasets and hypothesize a new protein family. The discovered system exhibits properties reminiscent of CRISPR with minimal human direction.
Over three years, the cost to achieve a specific AI performance level dropped by an average of 47% per quarter, representing a 13-fold annual decline. Model GPT-5.6 Luna matched OpenAI o3's GPQA Diamond score at a cost 725 times lower.
fMRI study reveals distinct neural systems support novelty-seeking and goal-directed exploration in aging, both linked to reduced social network outcomes.
A robust post-processing workflow combining Xenium spatial technology with iterative indirect immunofluorescence imaging was developed for formalin-fixed paraffin-embedded human brain tissue. Applied to calcarine cortex sections with cerebral amyloid angiopathy, amyloid plaques, and tau pathology, the method generated a 15-plex image capturing neuropathological hallmarks.
vLLM added new layers compatible with any GPU architecture, achieving token throughput within 3.4% of native performance on NVIDIA H100.
The llama-cpp-Python library now supports controlling model output via grammar, restricting tokens to ensure exact format compliance. This enables running OpenAI Functions on local devices using Llama 2 models.
Using Bayesian inference to dynamically adjust feedback thresholds, the team achieved ground-state initialization fidelity of 99.44% on a dispersively readout transmon qubit.
Ema secured $77 million in Series B funding led by Creaegis, quadrupling its valuation to $140 million. The startup uses AI agents to automate HR, IT, and finance tasks, capturing spending traditionally directed at enterprise software.
A new pipeline stabilizes AOLCO video to measure erythrocyte velocity in physical units. It is delivered as an interactive 3D Slicer extension and a batch command-line tool.
Phase Ib/IIa data indicate that the experimental antibody 9MW1911 reduces severe COPD flare-ups by up to 100% at the highest dose, signaling potential benefits for aging lungs.
The new KEX-bench benchmark reveals that top coding agents solve 56% of Linux kernel primitive tasks with a reference proof of concept, but only 5% without it.
New chips feature a sensing hub for running local AI agents and models up to 200 million parameters, including a 30 billion parameter mixture-of-experts model.
The system integrates a large language model with neuroimaging tools for quality control, modeling, and visualization. The agent demonstrated capabilities on a custom fNIRS benchmarking suite.
Singapore-based brand Elivity launched a spray that bypasses the stomach, delivering active ingredients under the tongue through NANOBOOST technology co-developed with a biotech spinoff of Hamburg University.
Researchers used LLM GPT-4o and a RoBERTa model to analyze 2,275 tweets containing specific keywords, identifying 80% of users as depressed.
NASA's Orion spacecraft completed the Artemis I mission with a safe splashdown in the Pacific Ocean on December 11, 2022, at 12:40 p.m. EST.
Researchers successfully amplified and sequenced human immunoglobulin variable regions from wastewater, observing diverse sequence polymorphisms suggesting distinct clonal lineages.
The study identified an asymmetry in subproblem difficulty: revealing upstream tokens triggers large confidence changes in nearby undecoded positions, while downstream tokens cause only minor local shifts. The proposed Informed Masking method leverages this structure to create better-posed subproblems for reinforcement learning in diffusion large language models.
The study proves the absence of long-range order in fragments and establishes that a fragment with spin-word period k supports a k-particle channel.
A new electrokinetic waste separation process removes cellular debris and contaminants, allowing for greater recycling of culture media in biotechnology.