产品发布/更新
-
AI version of Alipay opens public beta, Ant Abao available without invitation code
Alipay's Abao AI assistant officially opened its public beta today; iOS and Android users can search for "Abao" or…
-
Google Health API launches CLI: ghealth is an open-source tool for Fitbit data
ghealth is an open-source command-line tool wrapping Google Health API v4, released as a single Go binary (Apache 2.0…
-
Alibaba releases Page Agent: an open-source JavaScript library for natural-language DOM control
Alibaba released Page Agent, an open-source JavaScript client library that, once embedded in a webpage, can directly…
-
Kunlun Wanwei's Skywork 3.2 releases Skywork Tags, bringing AI agents into work group chats
Kunlun Wanwei's Skywork 3.2 released Skywork Tags, which brings AI agents into instant messaging tools such as Slack,…
-
Claude Enterprise adds usage and cost analytics and spend controls
Claude Enterprise introduces richer management analytics tools and cost controls. The dashboard can now analyze usage…
行业动态
-
Microsoft forms Frontier Company, investing $2.5 billion to place 6,000 AI engineers at enterprise customer sites
Microsoft has set up a new business unit, Frontier Company, allocating $2.5 billion to place 6,000 industry and…
-
Citi, Adobe, and other enterprises limit employee use of flagship AI models to control costs
According to internal materials obtained by 404 Media, six enterprises including Atlassian, Adobe, and Amazon are…
-
Kuaishou Kling AI receives $2.028 billion from initial investors, $18 billion post-money valuation
Kuaishou announced on the Hong Kong Stock Exchange that 21 initial investors agreed to inject 13.824 billion yuan…
-
Google AI buildout drives 37% electricity growth in 2025
In 2025, Google's annual electricity consumption rose 37% year over year, the largest increase in its history. Data…
论文研究
-
Multi-agent teams hinder expert performance
In self-organizing multi-agent LLM systems, teams fail to effectively leverage the expertise of expert members. Across…
-
Study on robustness and chain-of-thought consistency of RL-fine-tuned VLMs
Reinforcement learning (RL) fine-tuning has been extended to vision-language models (VLMs). The study finds that…
-
VideoFlexTok: variable-length coarse-to-fine video tokenization
VideoFlexTok proposes a variable-length token sequence video representation with a coarse-to-fine structure: the first…
技巧与观点
-
Fable 5 reaches 16.1% automation rate on RLI benchmark, six times higher than eight months ago
The Remote Labor Index (RLI) measures the share of 240 paid freelance projects (worth $144,000 in total) that AI…
-
Fable 5 builds Rube Goldberg machine for just $4.44
A Rube Goldberg machine built with Fable 5 costs just $4.44 👀 Prompt here:…
-
Qwen team's Zhu Da: the "more, faster, better, cheaper" engineering philosophy of C-end Agent Harness and exploration of proactive service
The Qwen team launched a general complex-task Agent in January 2026 (the capsule entry in the Qwen App) and summarized…
-
Agent-assisted SGLang development: an initial exploration
The SGLang team encoded workflows such as LLM serving, distributed runtime, GPU kernels, and diffusion pipelines into…
-
借用夜晚:将闲置推理GPU回收用于研究
Runway 开发了名为 deckard 的容量控制器,在生产推理集群与研究集群间动态重分配 GPU。生产流量在北美工作日上午 9 点 ET 达峰,晚 8 点 ET 跌至不足一半。控制器基于预计算的时间窗口(如工作日 8:30–12:30…