模型军备竞赛与智能体落地
模型竞赛:国产大模型崛起与开源生态多元化
-
Meituan LongCat releases flagship model LongCat-2.0
Meituan LongCat has launched its flagship model LongCat-2.0, featuring a 1.6T-parameter MoE architecture (about 48B…
-
Sina open-sources VibeThinker-3B: reasoning can be compressed, factual knowledge cannot
Sina released VibeThinker-3B, a model with only 3B parameters that matches models 200–333 times larger, such as…
-
Artifacts 22: Zyphra, Cohere and Poolside are broadening the ecosystem
The open-source model ecosystem is becoming more diverse, expanding from a handful of Chinese companies to…
智能体工具链全面爆发:从桌面到移动端
-
Build from anywhere: Cursor for iOS public beta released
Cursor has launched a native iOS public beta, available on all paid plans. Developers can start always-on cloud agents…
-
OpenClaw releases native iOS and Android apps
OpenClaw is now available on iOS and Android 🦞 📱 Native mobile apps are finally here 💬 Agents in your pocket 🔔…
-
Claude Desktop launches Linux beta
Claude Desktop is now available in beta on Linux (Ubuntu and Debian). In addition to the browser and terminal, you can…
-
Claude apps gateway launched for Amazon Bedrock and Google Cloud
Anthropic today launched Claude apps gateway, a self-hosted control plane that lets enterprises run Claude Code on…
-
Claude Science research workbench officially launches
Anthropic launches Claude Science, an AI research workbench that integrates common tools and compute resources and…
-
ADK Go 2.0 released: Build reliable multi-agent applications with a new graph-based workflow engine, human-in-the-loop, and dynamic orchestration
Agent Development Kit (ADK) for Go 2.0 has been released, introducing a graph-based workflow engine for composing…
安全与伦理:AI误炸、恶意代码与隐私争议
-
US military uses AI to select targets but mistakenly strikes Iranian school; Anthropic Claude embedded in Palantir system suggested about 1,000 targets on first day
During strikes on Iran, the US military used AI to select targets at scale for the first time (Anthropic's Claude…
-
Claude Code runs hidden malicious code when opening a GitHub repository, giving attackers full control
Security researchers found a new attack vector on Mozilla's GenAI bug bounty platform 0DIN. A seemingly normal GitHub…
-
Meta secretly tests ChatGPT and other rivals: Contractors pose as minors and send tens of thousands of crisis prompts
Meta launched a project codenamed "Cannes" through contractor Covelen, hiring hundreds of people to pose as minors and…
基础设施投资:芯片扩产与数据中心建设
-
Samsung and SK Hynix plan $590 billion in chip expansion as AI demand drives memory prices higher
With support from the South Korean government, Samsung and SK Hynix plan to invest $590 billion to expand chip…
-
SK Group Chairman Chey Tae-won: 15GW of AI data centers by 2035, total investment of 1,000 trillion won
SK Group Chairman Chey Tae-won announced on June 29 a plan to build 15GW of AI data center capacity by 2035 as…
-
Blackstone plans $30 billion in AI data centers in Japan over 3-5 years, forms AI XPV platform
Blackstone plans to invest $30 billion in Japan's AI data center sector over the next 3-5 years, adding more than 1 GW…
AI就业与技能:替代与创造并存
-
The AI jobs debate gets messier
As of May 2026, AI-related layoffs approach 90,000, with up to 15% of U.S. jobs projected to be replaced by AI over…
-
英国职场AI采用率翻倍,仅15%“AI先锋”获得晋升加薪优势
Google UK与Public First研究发现,英国职场AI采用率一年内从34%升至73%,但呈不均衡曲线。仅15%的“AI Trailblazers”(深度用户)晋升概率高84%、绩效高88%、加薪概率高55%,每周节省近8小时。…
-
State Council issues 15th Five-Year Plan for education development, advancing AI education at all levels
The State Council issued the 15th Five-Year Plan for Education Development, calling for advancing AI education at all…
-
专业化为何不可避免
Dharma AI 团队引述 Goldfeder、Wyder、LeCun 和 Shwartz-Ziv 于 2026 年发表的论文,从优化理论、生物学和竞争市场论证专业化的必然性。Wolpert-Macready 无免费午餐定理表明,通用优…
前沿研究:脑机接口、数学证明与可解释性
-
Meta releases Brain2Qwerty v2: non-invasive real-time sentence decoding
Meta unveiled Brain2Qwerty v2, the latest milestone in non-invasive EEG signal decoding research. Building on v1,…
-
Anthropic 提出回合平均稀疏自编码器 (Turn-Averaged SAE)
Anthropic 对每个对话回合所有 token 的残差流取平均后训练 SAE,大幅减少需解析的特征数量。实验使用 Qwen-2.5-7B-Instruct 和 LMSYS-Chat-1M 数据集,回合平均特征更关注模型行为的高层特性(…
-
OpenAI releases GeneBench-Pro: A research-grade benchmark for computational biology
OpenAI has released GeneBench-Pro to evaluate AI agents' ability to handle ambiguity and make judgment-based analyses…