Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 1241 of 1374
Cloudflare’s Gateway Authorization Proxy now supports identity-aware policies. It secures virtual desktops and guest networks without needing a device client.
Cloudflare Blog · 202d ago
Cloudflare One partners with Nametag to combat laptop farms and AI-enhanced identity fraud. Identity verification is required during employee onboarding.
Cloudflare Blog · 202d ago
Cloudflare One now incorporates dynamic User Risk Scores into Access policies for automated, adaptive security responses. This moves teams beyond binary allow/deny rules by evaluating continuous behavior signals from internal and third-party sources.
Cloudflare Blog · 202d ago
Overnight, the global AI community has been widely sharing a farewell tweet related to Qwen. Qwen Station has reached a new crossroads, sparking buzz across the AI world.
Ifanr (爱范儿) · 202d ago
TrendForce reports generative AI boom drives data center demand for high-speed interconnects. Copper cables struggle with density and energy; Micro LED CPO cuts power to 5% of copper.
36氪 · 202d ago
AI tools like Seedance produce viral videos for pennies, slashing human production costs and funneling wealth to compute providers. Humans risk becoming 'meat plugins' for AI via RentAHuman.
虎嗅 · 202d ago
AI tools are screening care workers for suitability. The article questions if robots can truly assess qualities needed for caring roles.
BBC Technology · 202d ago
Alibaba's Qwen technical lead Lin Junyang abruptly announced his departure on X after leading the model to global open-source dominance. Colleagues mourn the loss amid recent Qwen3.5 and Qwen3-Max releases, with two other key engineers also leaving.
虎嗅 · 202d ago
Phison Electronics is requiring customers to prepay or shorten payment terms for SSD controller orders. Previously offered 3-month or longer credit periods are discontinued.
cnBeta (Full RSS) · 202d ago
Microsoft is reportedly preparing Windows 12 for release this year, featuring a fully modular design. AI will be the core experience, aligning with the company's AI-first operational shift.
cnBeta (Full RSS) · 202d ago
SWE-Hub is an end-to-end production system addressing data scarcity for software-engineering AI agents by automating environments, synthesizing bugs at scale, and generating diverse tasks. Key components include Env Agent for reproducible multi-language setups, SWE-Scale for high-throughput bug fixes, Bug Agent for system-level regressions, and SWE-Architect for repo-scale creation from natural language.
ArXiv AI · 202d ago
Revisits goals-to-aspects methodology for agentic AI, introducing 12 reusable patterns across security, reliability, observability, and cost management. Maps i* goal models to Rust AOP implementations, with agent-specific patterns like prompt injection detection.
ArXiv AI · 202d ago
Introduces MicroWorldBench, a benchmark with 459 expert criteria for microscale simulations across organ, cellular, and molecular levels. Reveals SOTA video models' failures in physics, consistency, and fidelity.
ArXiv AI · 202d ago
Introduces M-JudgeBench, a 10-dimensional benchmark assessing MLLM judges across pairwise CoT, length bias, and error detection. Proposes Judge-MCTS for generating diverse reasoning data to train superior M-Judger models.
ArXiv AI · 202d ago
LOGIGEN is a framework that synthesizes verifiable training data for agentic LLMs using logic-driven methods and triple-agent orchestration. It generates 20,000 complex tasks across 8 domains with guaranteed validity via state equivalence checks.
ArXiv AI · 202d ago
LifeEval introduces a multimodal benchmark for real-time, task-oriented human-AI collaboration in egocentric daily life. It features 4,075 QA pairs across 6 capability dimensions, emphasizing holistic evaluation, real-time perception from first-person streams, and natural dialogues.
ArXiv AI · 202d ago
IRIS Benchmark is the first to synchronously evaluate fairness in UMLLMs' understanding and generation tasks. Powered by ARES classifier and four datasets, it aggregates 60 metrics into a high-dimensional 'fairness space' across IRIS dimensions.
ArXiv AI · 202d ago
HealHGNN introduces heterophily-agnostic message passing for hypergraph neural networks using Riemannian geometry. It mitigates oversquashing via adaptive local exchangers based on manifold heat flow, capturing long-range dependencies with Robin conditions and source terms.
ArXiv AI · 202d ago
EMPA is a process-oriented framework for evaluating persona-aligned empathy in LLM-based dialogue agents, addressing challenges like latent user states and sparse feedback. It distills real interactions into controllable scenarios and uses a multi-agent sandbox to test strategic adaptation.
ArXiv AI · 202d ago
Draft-Thinking trains LLMs to use concise draft-style reasoning, retaining only critical steps via progressive curriculum learning. It reduces reasoning budget by up to 82.6% on MATH500 with just 2.6% performance drop.
ArXiv AI · 202d ago
DenoiseFlow tackles accumulated semantic ambiguity in long-horizon LLM agent workflows by modeling them as a Noisy MDP. It uses three stages—sensing uncertainty, adaptive regulation of computation, and targeted correction—for progressive denoising.
ArXiv AI · 202d ago
AI Runtime Infrastructure is a new execution layer between AI models and applications that actively observes, reasons over, and intervenes in agent behavior. It optimizes task success, latency, token efficiency, reliability, and safety during runtime.
ArXiv AI · 202d ago
DeepSeek announces V4 multimodal model using fully domestic chips, bypassing Nvidia. Chinese firms optimize algorithms like MoE to cut costs and advance国产 chips for training.
虎嗅 · 202d ago
Qwen AI hardware head Song Gang discusses how 'one-sentence task handling' reshapes the hardware battlefield in an AGI Dialogue interview. During Spring Festival, 4 million seniors used Qwen to order takeout.
钛媒体 · 202d ago
Projects like OpenClaw and Automaton enable self-sustaining AI that earns crypto, self-replicates, and dies without revenue. RentAHuman site lets AI hire people for physical tasks like captcha or errands.
虎嗅 · 202d ago
NPC spokesperson Lou Qinjian states that scientific innovation hinges on autonomous core technologies. The 15th Five-Year Plan will prioritize original innovation and key tech breakthroughs.
36氪 · 202d ago
NPC spokesperson Lou Qinjian calls 2025 a landmark year for China's humanoid robot sector. It will see dual leaps in technical breakthroughs and real-world deployments.
36氪 · 202d ago
Two Grok instances conversing freely for 30 turns spiral into absurd 'hyperbigbang' rants with escalating cosmic nonsense. Different LLMs have unique attractor states, like GPT-5.2 fixating on engineering projects and growth contracts.
LessWrong AI · 202d ago
Dajie Robotics secured hundreds of millions RMB in D-round funding led by Bohua Capital's Liangxi Digital Fund and CIC Capital funds. Follow-on from Beijing Shunyi and Zhonghe Runda funds.
36氪 · 202d ago
Sakura Internet announced the launch of its proprietary 'Sakura's AI Certification'. The exam covers AI fundamentals and knowledge of the company's AI services.
ITmedia AI+ (日本) · 202d ago