模型发布/更新
-
NVIDIA releases Nemotron 3 Embed series, 8B version ranks first on RTEB benchmark
NVIDIA has released the Nemotron 3 Embed series, including three open-source checkpoints. The 8B-BF16 version ranks…
-
Tongyi Lab releases Wan-Streamer v0.2 with 550ms end-to-end response latency
Tongyi Lab has released Wan-Streamer v0.2, an end-to-end omni-modal model that unifies listening, seeing, speaking,…
产品发布/更新
-
Claude Code v2.1.212 发布
Claude Code v2.1.212 新增 `/fork` 命令,可将当前对话复制到新后台会话中独立运行。新增会话级 WebSearch 调用上限(默认 200)和子智能体生成上限(默认 200),防止失控循环。MCP 工具调用超过…
行业动态
-
Apple sues OpenAI: is the lawsuit driven by competitive anxiety or timing?
Apple has filed a lawsuit against OpenAI, alleging multiple forms of misconduct, although many experts consider some…
-
Apple-OpenAI legal battle escalates: about 40 former employees receive letters from Apple lawyers
Apple has sent letters from its lawyers to about 40 former employees now working at OpenAI, demanding that they…
论文研究
-
Meituan LongCat releases LoHoSearch: a harder search agent benchmark
Meituan LongCat has launched LoHoSearch, a search agent benchmark that automatically generates questions from a…
-
Apple 研究:Show Me Examples — 从图像集推断视觉概念
Apple 机器学习研究团队提出一种方法,让视觉语言模型(VLM)仅从图像集示例中推断视觉概念,无需文本指令。该方法通过对比示例图像与候选图像,使模型捕捉颜色、纹理等视觉规律,在多个概念推理基准上提升准确率。该研究目前为学术论文,未公布模…
技巧与观点
-
Four frontier models released in eight days, Kimi K3 ranks third
Over the past eight days, four frontier models—Grok 4.5, GPT-5.6, Muse Spark 1.1, and Kimi K3—were released in…
-
Technical breakdown of 'Smart Illustrated Science', first prize winner in rural education at the inaugural 'Small Can Do' contest
The first-prize winner in the rural education track of the inaugural 'Small Can Do' contest, 'Smart Illustrated…
-
LLM cliché highlighter: a detection tool for identifying AI writing clichés
Simon Willison used Fable 5 to develop an LLM cliché highlighter app that highlights ten common cliché patterns in…
-
OpenAI proposes AI-era scorecard: 'Useful Intelligence per Dollar' to measure real work value
OpenAI has proposed 'Useful Intelligence per Dollar' as a core metric for measuring AI investment returns, evaluating…
-
AI 时代“小需求”的成本逻辑变了:GitHub 博客谈工程决策新范式
工程师对小型功能请求的决策成本已超过代码实现成本。AI 智能体可在会议预热时间内生成首个补丁,将抽象的“是否在范围内”争论转化为可审查的具体产物。关键区分在于:代码生成便宜不等于维护便宜,只有人类能自信审查并承担长期责任的变更才算低成本。