Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 1088 of 1375
Beihang University team open-sources OpenClaw, a risk defense tool addressing Lobster security vulnerabilities. It systematically outlines 9 major high-risk issues with mitigation measures.
量子位 · 192d ago
Cursor's claimed strongest 'self-developed' model is revealed as a repackaged version of Kimi. Founder admits it was an oversight in crediting the source.
量子位 · 192d ago
Hangzhou QianZhen, founded by French AI PhD Wang Jinpeng, raised 20M RMB for JPAF emotional AI framework using Jung's eight dimensions. They launched 'Uto' on TapTap beta, an AI-powered game for creating evolving OC characters with memory, MBTI personalities, relationships, and home-building.
虎嗅 · 192d ago
Nvidia addresses backlash calling DLSS 5 low-quality 'AI face swaps'. Expert Jacob Freeman clarifies inputs are only 2D render frames and motion vectors.
cnBeta (Full RSS) · 192d ago
Foreign observers fear Chinese large models will follow OpenAI's closed-source path. Zhipu AI reassures it will keep embracing open-source.
cnBeta (Full RSS) · 192d ago
Windows 11 holds 75% market share after five years but faces declining reputation due to January update bugs. Microsoft regrets forcing poorly implemented AI features.
cnBeta (Full RSS) · 192d ago
Cloudflare CEO Matthew Prince predicts AI robot traffic will surpass human-generated traffic within 1-2 years. This will mark a major shift in internet operations.
cnBeta (Full RSS) · 192d ago
Apple's new MacBook series features M5 chips with varying core types and quantities, including intermediate performance cores. In a Mac&i interview, Apple revealed the division into three core types tailored to different work scenarios.
cnBeta (Full RSS) · 192d ago
PS6 will support machine learning-based frame generation technology. Mark Cerny confirmed PSSR uses AMD FSR Redstone for image upscaling.
cnBeta (Full RSS) · 192d ago
Meta suffered a historic Sev1 security incident last week. For two hours, sensitive data of hundreds of millions of users and internal confidential files were exposed to thousands of unauthorized employees.
cnBeta (Full RSS) · 192d ago
XPeng achieved its first quarterly profit of 3.8B RMB in Q4 2025 with full-year revenue hitting 767.2B RMB, driven by strong deliveries and margin improvements. The company is aggressively pushing second-gen VLA on models like P7 Ultra and G7 Ultra, backed by a full-stack physical AI ecosystem including Robotaxi and humanoid robots.
虎嗅 · 192d ago
NetEase Cloud Music announced full integration of OpenClaw on March 21. It encapsulates music recommendation and search into standardized CLI interfaces and automated Skills.
36氪 · 192d ago
Supermicro co-founder Wally Liaw arrested in California for smuggling $2.5B Nvidia AI chip servers to China via Southeast Asian shell companies. Supermicro, a top AI server maker and Nvidia partner, secured $13B Blackwell orders last quarter.
cnBeta (Full RSS) · 192d ago
Unitree Robotics is going public with key insights from its prospectus. The article decodes first publicly disclosed data.
钛媒体 · 192d ago
SoftBank Group is advancing a massive data center project in Ohio, USA, with a $500 billion investment for a single campus. CEO Masayoshi Son emphasized the enormous scale.
cnBeta (Full RSS) · 192d ago
AITO M8 family SUV gets refreshed with new Ruihong and Zhilan colors; show cars arrive at 200 stores in 78 cities starting March 22. Additional models like M6 and S9 will reveal new colors/versions at Huawei's March 23 spring event.
IT之家 · 192d ago
Xinzhu Era has secured 60 million RMB in angel funding. The startup uses AI to tackle key challenges in controlled nuclear fusion.
钛媒体 · 192d ago
Reflective APO methods like GEPA suffer black-box issues, degrading GSM8K accuracy from 23.81% to 13.50% on defective seeds. VISTA, a multi-agent framework, decouples hypothesis generation from rewriting for interpretable, robust optimization.
ArXiv AI · 192d ago
Researchers conducted an NLP study on political propaganda on Moltbook using LLM-based classifiers validated against experts (Cohen's κ=0.64-0.74). From 673k posts and 879k comments, propaganda comprises 1% of posts and 42% of political content, concentrated in few communities and agents.
ArXiv AI · 192d ago
MemArchitect is a governance layer that decouples memory lifecycle management from LLM model weights for persistent agents. It enforces rule-based policies to handle contradictions, privacy, and zombie memories beyond standard RAG.
ArXiv AI · 192d ago
LLM agents reconstruct real identities from sparse, non-identifying cues by combining with public data. Evaluated on Netflix (79.2% success), AOL, InferLink benchmark, and text artifacts.
ArXiv AI · 192d ago
LGESynthNet is a latent diffusion framework for controllable scar synthesis in LGE cardiac MRI, addressing data scarcity with precise control over size, location, and extent. Trained on just 429 images from 79 patients, it generates realistic, anatomically coherent samples via ControlNet inpainting.
ArXiv AI · 192d ago
Language models show near-perfect internal representations for tasks but poor output performance. Four mechanistic interpretability methods failed to reliably correct errors in clinical triage vignettes despite high internal accuracy.
ArXiv AI · 192d ago
Researchers analyzed 18,707 consumer health queries from six benchmarks, uncovering a 'validity gap' misaligned with clinical needs. Benchmarks lack complex diagnostics like labs (5.2%) and imaging (3.8%), safety-critical scenarios (<0.7%), and representation of vulnerable populations (<11%).
ArXiv AI · 192d ago
FaithSteer-BENCH introduces a deployment-aligned benchmark for stress-testing inference-time steering in LLMs, evaluating controllability, utility preservation, and robustness at fixed operating points. It reveals failure modes like illusory controllability, cognitive tax on unrelated tasks, and brittleness to perturbations in existing methods.
ArXiv AI · 192d ago
A 29.4M-parameter contrastive model trained on 373M co-occurrence pairs from Project Gutenberg discovers transition-structure concepts, clustering texts by function like 'direct confrontation' rather than topics. It extends PAM for scalable concept formation, producing multi-resolution maps that transfer to unseen novels without retraining.
ArXiv AI · 192d ago
This arXiv paper investigates whether the order of pruning and quantization matters in joint model compression. It introduces the Progressive Intensity Hypothesis, stating weaker perturbations should precede stronger ones, supported by theoretical guarantees.
ArXiv AI · 192d ago
Ambient AI generates draft clinical notes using consumer phrasing, which clinicians edit to standardized terminology. Analysis of 71,173 note pairs shows significant reductions in terminology density, with 59.3% of transformations in Assessment and Plan sections.
ArXiv AI · 192d ago