Full archive

Textbook AI Agent Lands on Xiaoyuan Learning Machines

人民教育音像數字出版社與小猿達成合作，推出「中小學課本學習智能體」。該智能體於2026年8月21日首次上線小猿AI學習機。

量子位 · 25d ago

How Close Are World Models to General Intelligence?

The article examines the gap between current world-model capabilities and a genuinely general world model. It argues that such a system must integrate generation, simulation, robotic action, and policy capabilities into a closed-loop feedback system rather than treating them as isolated modules.

钛媒体 · 25d ago

Data Center Boom Fuels U.S. Gas Power Buildout

The rapid growth of data centers is driving a wave of natural-gas power projects across the United States. Many projects have already been proposed or are under construction, highlighting the infrastructure and environmental costs of expanding AI and cloud capacity.

Wired · 25d ago

Vercel Connect Ends Agent Credential Sprawl

Vercel Connect is now generally available, replacing long-lived provider tokens with short-lived, task-scoped credentials requested at runtime. It supports more than 100 connectors and uses Vercel OIDC identity to authenticate deployments without storing provider secrets.

Vercel News · 25d ago

Teaching Agents to Discover Hidden Rules

This arXiv report examines reinforcement learning agents that infer hidden rules through trial-and-error feedback in the Game of Hidden Rules (GOHR). It analyzes representation design, rule difficulty, transfer learning, generalization, and how pseudo-bots may support human learning.

ArXiv AI · 25d ago

Secure RAG Inference on IBM LinuxONE

這篇論文提出一套完全運行於 IBM LinuxONE 的六子系統 RAG 架構，結合 IBM Spyre、Telum II 與 Red Hat OpenShift。初步分析顯示，架構可達成低於兩秒的端到端延遲，並較平台外推論降低最多 20 倍延遲，同時維持資料隔離與合規性。

ArXiv AI · 25d ago

SchemaRouter Cuts Agentic RAG Context by 89%

SchemaRouter is a schema-graph routing layer that selects tools and response fields for heterogeneous agentic RAG systems. On a 110-query materials-science benchmark, it matched fetch-everything accuracy while using 227 instead of 2,066 context tokens and achieving lower latency than prompt-all.

ArXiv AI · 25d ago

RIACT Spots Early Student Burnout Signals

RIACT is a web-based AI system that tracks students’ self-logged study sessions and identifies potential early burnout signals. It combines transparent week-over-week behavioral rules with a schema-constrained large language model to provide personalized observations and recommendations without making diagnoses.

ArXiv AI · 25d ago

LLM Leaderboards Depend on Fragile Evaluation Harnesses

A new study shows that LLM benchmark rankings can change dramatically based solely on evaluation-harness choices, including prompt wording, option order, and scoring method. Across 12 open-weight models, four reached first place under different configurations, while fragile items explained 95.7% of model gaps on average.

ArXiv AI · 25d ago

LitReview Arena Benchmarks AI Literature Reviews

LitReview Arena introduces a battle-style, expert-reviewed platform for evaluating AI-generated literature reviews across five quality dimensions. Based on roughly 3,000 expert judgments, the study finds leading systems win only 23% of decisive matches against human drafts, while its LitJudge evaluator raises agreement with experts to Spearman’s rho 0.78.

ArXiv AI · 25d ago

KVBoost Makes LLM Prefill 4.49× Faster

KVBoost is a chunk-level KV cache reuse system for HuggingFace-compatible decoder models, allowing cached content to be reused even when it appears at arbitrary positions in prompts. On Qwen2.5-3B bug-localization workloads, it reduced time-to-first-token by 4.49× while maintaining comparable accuracy.

ArXiv AI · 25d ago

How Synthetic Data Causes Model Collapse

This arXiv review examines model collapse, a failure mode that can emerge when AI-generated data is repeatedly used to train successor models. It surveys research across application scenarios and consolidates proposed countermeasures, challenges, and future directions.

ArXiv AI · 25d ago

Grounded AI Generates and Repairs Robot Programs

This research presents a workflow that generates ABB RAPID programs from natural-language instructions, grounds them in technical documentation and production templates, and iteratively corrects them through RobotStudio simulations. An MCP server connects the language-model client to RobotStudio, exposing execution failures that static code checks may miss.

ArXiv AI · 25d ago

CoreSec Brings Abstention to Network RCA

CoreSec is a production root cause analysis system for hyperscale Clos networks that uses a PAM-style abstention algebra instead of weighted score fusion. It produces deterministic, explainable decisions and explicitly abstains when telemetry is ambiguous or incomplete.

ArXiv AI · 25d ago

Open Source

Chinese Models Lead the Open-Weight AI Race

Mozilla has published The State of Open Source AI, examining adoption and performance trends for open-weight AI models. The report indicates that Chinese models occupy many top positions, with a significant lead over US models, based on SlashData research and public data from Chatbot Arena and OpenRouter.

ITmedia AI+ (日本) · 25d ago

ChatGPT Desktop App Arrives on Linux

OpenAI has released a preview version of its ChatGPT desktop application for Linux. The app provides access to ChatGPT, ChatGPT Work, and Codex, while the article examines its supported and unsupported capabilities.

ITmedia AI+ (日本) · 25d ago

Benchmarking Efficient Hate Speech Detection

This arXiv study compares zero-shot inference, LoRA-based parameter-efficient fine-tuning, prompt tuning, and prompt engineering for hate speech classification in low-resource Roman Urdu. It evaluates how different training and prompting strategies perform under limited-data conditions.

ArXiv AI · 25d ago

AIREP Brings Per-Decision Evidence to AI Governance

AIREP is a protocol for creating signed, independently verifiable records whenever an AI runtime releases, blocks, defers, redacts, or escalates an output. It uses hashed references, explicit evidence boundaries, and chained records to make governance decisions auditable across runtimes.

ArXiv AI · 25d ago

Models

Kimi Raises Prices as US AI Giants Cut Costs

OpenAI and Google have announced substantial API or token-price cuts, while Anthropic canceled a planned price increase. In contrast, Kimi K3 has raised output pricing and DeepSeek has signaled a major increase, highlighting intensifying competition between cost-efficient Chinese models and high-cost US providers.

虎嗅 · 25d ago

Zhipu AI Members Face Sudden Account Bans

The article examines sudden account bans affecting Zhipu AI members and questions how many users may have been wrongly penalized. It raises concerns about moderation transparency and appeal mechanisms.

钛媒体 · 25d ago

Doubao Work Becomes a Feishu Teammate

A hands-on review explores how Doubao Work changes after connecting to Feishu. The integration gives its Agent access to people, projects, context, and workflows, making it more useful for workplace collaboration.

Ifanr (爱范儿) · 25d ago

Business

Alibaba Raises HK$80 Billion for AI

Alibaba announced a new share placement to raise HK$80 billion, described as the largest follow-on share issue in Hong Kong market history. The fundraising signals a major commitment to expanding its AI strategy and capabilities.

钛媒体 · 25d ago

Figure Robots Turn Headquarters Into a Sci-Fi Scene

A video showing Figure AI humanoid robots walking outdoors around the company’s headquarters has attracted attention on X. CEO Brett Adcock described the Figure campus as resembling a science-fiction movie.

ITmedia AI+ (日本) · 25d ago

NVIDIA AI Servers Face a 15% Price Hike

NVIDIA AI-server systems are reportedly set to rise by more than 15% for shipments beginning early next year, largely because memory costs have surged. HBM and DRAM shortages are shifting pricing power toward Samsung, SK hynix, and Micron while increasing pressure on cloud providers and server ODMs.

虎嗅 · 25d ago

SetupAIModels

Models

Why I Stopped Talking to AI

The author explains why they have largely stopped using AI for open-ended conversation, arguing that fluent answers often create a temporary sense of insight without durable understanding. The essay calls for slower, independent learning and warns that AI's apparent omniscience can erode personal experience, judgment, and intellectual grounding.

虎嗅 · 25d ago

ASD Warns of Attacks on TeamCity Servers

Australia’s Signals Directorate (ASD) warns that Australian JetBrains TeamCity servers are being targeted. The warning follows a critical authentication bypass vulnerability that was patched in late July.

iTNews Australia · 25d ago

YMTC Files for A-Share Listing

Changjiang Storage, also known as YMTC, has reportedly filed for an A-share listing. The company is highlighted as generating approximately RMB370 million in daily profit, with Wuhan state-owned capital positioned as a major beneficiary.

钛媒体 · 25d ago

Thunderbird Bets on AI Plus AR

雷鳥創新表示，智能眼鏡的長期方向仍是 AI 與 AR 的結合，目前最重要的工作是推進 AI 能力。公司強調 AI 應該用來增強人類，而不是取代人類。

Ifanr (爱范儿) · 25d ago