SetupAIBusiness
Business
OpenAI Backs PORTS-Pike in Ohio
OpenAI has joined the PORTS-Pike project, expanding its community investment. The initiative also supports thousands of jobs in Southern Ohio.
OpenAI News · 37d ago
Full archive
Every story we have kept, newest first.
Looking for the daily editions? → Past editions
Page 159 of 1375
August 17, 2026
SetupAIBusiness
Business
AI Leaders Must Build More Agile Organizations
Harvard Business School Professor Linda Hill says the next generation of CEOs must prioritize flexible organizational structures as AI reshapes business operations. Leaders will need to adapt how teams collaborate, innovate, and deploy AI across the enterprise.
Bloomberg Technology · 37d ago
Infrastructure
Domestic GPUs Enter the Big-Tech Finals
Chinese GPU companies are showing rapid revenue growth and narrowing losses, but rising inventories and cash outflows reveal an aggressive race for wafer capacity. The industry is shifting from government-led domestic substitution toward demanding, recurring purchases from major internet companies.
虎嗅 · 37d ago
Applications
AI Personas Solve Three Marketing Gaps
味之素冷凍食品透過 NEC 的 BestMove 導入 AI Persona，改善行銷團隊缺乏資料、技能與創意延伸能力的問題。這項案例展示了 AI Persona 如何支援行銷洞察與企劃發想。
ITmedia AI+ (日本) · 37d ago
Video
Higgsfield Raises $400M at $5.4B Valuation
AI video generation platform Higgsfield has raised $400 million from DST Global, Goldman Sachs, Intel and other investors, reaching a $5.4 billion valuation. The company plans to use the funding to expand enterprise subscriptions, secure compute capacity and strengthen enterprise-grade products and security.
IT之家 · 37d ago
Applications
Chinese Robots Advance Despite Tech War Barriers
Chinese robotics companies showcased advanced prototypes at CES 2026, including stair-climbing and flying robot vacuums and Xpeng’s highly realistic Iron humanoid robot. However, export restrictions and geopolitical tensions may prevent American consumers and businesses from accessing these products.
SCMP Technology · 37d ago
Research
E-Commerce Models Face a Tougher Test
The article presents a rigorous evaluation philosophy for e-commerce AI models. It proposes defining tasks from real user needs and validating outcomes against states in simulated environments rather than accepting approximate results.
Ifanr (爱范儿) · 37d ago
Infrastructure
Memory, Not GPUs, Is Driving the Graphics Card Surge
The article argues that rising graphics card prices are being driven primarily by memory costs and constrained supply rather than GPU pricing. IDC estimates that AI data centers will consume 70% of global memory capacity in 2026, while major memory manufacturers are allocating 70%–90% of advanced capacity to HBM.
钛媒体 · 37d ago
Infrastructure
Microsoft’s AI Ambitions Face a Chip Capacity Question
A Guardian investigation identifies an apparent discrepancy between Microsoft’s claims about its AI capacity and the number of advanced AI chips it reportedly operates. The finding raises questions about whether chip availability could constrain the company’s AI development and deployment plans.
The Guardian Technology · 38d ago
Research
Why Confidently Wrong LLMs Can Stay Stable
This study examines stable miscalibration, where large language models produce confident wrong answers that remain stable under small prompt perturbations. Across multiple domains and open-weight models, self-critical prompting reduced hidden-state sensitivity but did not reliably distinguish confident errors from correct answers.
ArXiv AI · 38d ago
Research
RubricForge Cuts False Passes in Agent Evaluation
RubricForge automatically induces a human-readable judging rubric from a small set of ground-truth-labeled agent trajectories, then applies the frozen rubric without environment access. On tau-bench and WebShop, it substantially reduced false passes for failed trajectories and improved ranking of graded outcomes, although overall agreement was not significantly better than G-Eval.
ArXiv AI · 38d ago
Research
Robots Learn to Look Before Asking
This research introduces an active-perception framework that helps robots resolve ambiguous targets by changing their viewpoint and gathering missing visual evidence. A vision-language model decides whether to keep observing, ask the user for clarification, or finalize the target selection, with the approach validated on real robots.
ArXiv AI · 38d ago
Research
Rethinking AI Evaluation Around Human-AI Teams
This arXiv position paper argues that AI evaluation focused on autonomous, superhuman performance implicitly prioritizes replacing humans. It proposes evaluating human-AI teams instead, with the goal of building systems that complement human capabilities and produce better societal outcomes.
ArXiv AI · 38d ago
Models
No Single Signal Catches Every LLM Regression
This study finds that no inference-time signal universally predicts when an updated LLM will produce a worse answer than its predecessor. Confidence works best for multiple-choice and simpler math, while likelihood and KL-based signals are more useful for difficult math and code, enabling selective fallback to the older model.
ArXiv AI · 38d ago
Research
MobileMem Brings Long-Term Memory to Mobile AI
MobileMem is a benchmark and framework for evaluating on-device, long-term memory in personal AI assistants using year-scale mobile experiences. It tests multimodal and text-based agents on temporal reasoning, knowledge updates, multi-hop recall, and implicit preference inference.
ArXiv AI · 38d ago
Research
LLMs Develop Brain-Like Cognitive Modules
A circuit analysis of 46 tasks finds that large language models develop modular neuron organization across language, formal reasoning, social reasoning, and physical reasoning. Tasks associated with the same human brain network activate overlapping LLM neurons, while different domains recruit more distinct populations.
ArXiv AI · 38d ago
Infrastructure
Agentao Makes LLM Agents Governable Locally
Agentao is a local-first runtime that separates LLM-generated action proposals from host-authorized tool execution. It provides explicit permissions, state management, protocol boundaries, and execution traces to make tool-using agents more inspectable and controllable.
ArXiv AI · 38d ago
Infrastructure
A Year of Real-World LLM Serving Insights
A new study analyzes a full year of production LLM-serving traffic from Chutes across models, users, time periods, and interaction patterns. The researchers will release the complete trace to support more realistic benchmarking and serving-system research.
ArXiv AI · 38d ago
Research
A New Taxonomy of Misunderstanding in AI Agents
This arXiv paper presents a cross-disciplinary framework for understanding how communication divergences are generated, amplified, detected, and repaired in AI-mediated interactions. It identifies 11 failure modes across eight analytical layers and formalizes the process from signal transmission to meaning reconstruction.
ArXiv AI · 38d ago
Research
A New Metric for Consistent AI Agent Behavior
Researchers introduce the Behavioral Consistency Metric (BCM), which measures how consistently language model agents behave across tasks rather than only whether they succeed. Evaluations across roughly 9,000 software-engineering trajectories show that cross-task consistency and repeated-attempt consistency are distinct properties.
ArXiv AI · 38d ago
Image
AI Image Errors Trigger Product Label Apology
味之素為「味之素公園」官方 X 於 8 月 14 日發布的圖片致歉。該圖片經生成式 AI 加工後，造成「ほんだし」產品標籤與部分文字變形或錯亂。
ITmedia AI+ (日本) · 38d ago
Infrastructure
Maniformer Raises Multi-Billion RMB Funding
Maniformer completed a new multi-billion RMB financing round led by China Telecom, with Zhangjiang Group and existing investors participating. The funding will scale MEgo hardware production and expand its physical-AI data collection, governance, evaluation, and global delivery capabilities.
雷峰网 · 38d ago
Applications
AI Drama Rights Protection Hits a Wall
AI-generated dramas are increasingly accused of cloning voice actors, game characters, celebrities, and ordinary users without authorization. Rights holders face opaque training data, fragmented industry responsibility, difficult evidence collection, and compensation that often fails to cover litigation costs.
虎嗅 · 38d ago
Applications
Lagrange Rebuilds Factory Robotics Around Tasks
Lagrange is testing its Agentic OS in an automotive logistics line with wheeled dual-arm robots, a quadruped robot, and an AGV. Its approach combines task-centric process redesign with layered intelligence so robots can monitor results, recover from failures, and complete end-to-end workflows instead of merely executing isolated commands.
雷峰网 · 38d ago
Research
Being-H0.8 Brings Touch Into Latent Robotics
BeingBeyond, founded by Peking University professor Lu Zongqing, is pursuing a latent world-action model that predicts robot actions and future world states in embedding space rather than generating pixels. Its Being-H0.8 model adds tactile sensing alongside vision, action, and future-state prediction for physical interaction reasoning.
雷峰网 · 38d ago
Research
World Models Enter Their Foundational Year
The article declares that the era of world models has arrived, signaling growing momentum around this AI research direction. However, the provided excerpt offers no specific models, companies, benchmarks, or technical evidence.
钛媒体 · 38d ago
Applications
Ten Unusual Embodied AI Bets
Digu Robotics’ closed-door DemoDay showcased ten early-stage projects spanning embodied AI, intelligent hardware, assistive robotics, and AI companions. The event highlighted products that are already testable or shipping, including an AI golf coach, a jumping robot, a knee exoskeleton, a following mini-truck, and home robots.
极客公园 · 38d ago
SetupAIResearch
Research
OpenAI Funds 14 AI Policy Projects
OpenAI is funding 14 independent projects that explore new policy ideas for the Intelligence Age. The initiative aims to expand economic opportunity and strengthen societal resilience as AI adoption grows.
OpenAI News · 38d ago
Infrastructure
Intel Arc Pro B70 Prices Surge 48%
Intel Arc Pro B70 professional GPUs have risen 26% to 48% in retail channels across the US, Germany, and South Korea. The AI workstation GPU is gaining attention as a scarce accelerator with 32GB of memory and 367 TOPS of INT8 performance.
cnBeta (Full RSS) · 38d ago