模型发布/更新
-
Kimi K3 released: 2.8T-parameter open-source model with native vision and million-token context window
Moonshot AI has released Kimi K3, a 2.8T-parameter open-source model built on the Kimi Delta Attention and Attention…
产品发布/更新
-
Grok launches Automations: scheduled or email-triggered task execution with results reporting
xAI has introduced Automations for Grok, allowing users to describe a task once and have Grok run it automatically on…
-
ChatGPT workspace supports editing documents, spreadsheets, and slides
Create and edit polished documents, spreadsheets, and slides in the ChatGPT workspace. @nickbaumann_ demonstrates how.
-
Tiangong Short Drama Workbench launches dual-track creation mode with Agent storyboarding and infinite canvas
Tiangong Short Drama Workbench has launched a dual-track creation mode, in which a director Agent automatically parses…
-
Miaoda 3.5 debuts iOS App globally, no-code development and multi-app shared backend capabilities upgraded
Baidu AI Cloud released Miaoda 3.5 at WAIC2026, adding iOS packaging capability so users can package apps into IPA…
-
ModelBest open-sources enterprise AI digital employee platform StaffDeck
ModelBest, together with multiple teams, has open-sourced StaffDeck, an enterprise platform for building and managing…
-
MiniMax Code 2.0 desktop revamped: underlying architecture fully upgraded, finance module coming soon
MiniMax Code 2.0 desktop has been released, rebuilding its underlying architecture on the Pi Agent framework and…
-
Using Claude Fable 5 in Claude Cowork
Anthropic has released Claude Fable 5, its most capable general model, designed for long-running, multi-step complex…
行业动态
-
xAI sues Grok user over child sexual abuse content, no longer denies model misuse
xAI has filed its first lawsuit against a Grok user, alleging that the user used the model to create child sexual…
-
EU rules Google must open Android and Search to rivals, affecting Gemini and other AI services
The EU has ruled under the Digital Markets Act (DMA) that Google must open key parts of Android and Google Search to…
-
TSMC raises 2026 capex forecast to $60-64 billion, A14 process on track
At its 2026 Q2 earnings call, TSMC raised its 2026 capital expenditure forecast to $60-64 billion, up from a previous…
-
Signing ceremony for World AI Cooperation Organization agreement held in Shanghai, headquarters in Shanghai, China
On July 16, the signing ceremony for the agreement establishing the World AI Cooperation Organization was held in…
-
54%企业已遭遇AI智能体安全事件,多数仍共享凭证
VentureBeat调查107家企业发现,54%已遭遇AI智能体安全事件(18%确认事故,36%险些酿祸)。仅32%为每个智能体分配独立身份凭证,30%将高风险智能体隔离在沙箱中。安全工具主要依赖模型提供商原生方案,专用智能体安全产品渗…
-
Google DeepMind 与 Isomorphic Labs 公布生物弹性联合方案:用 AI 预防、检测和应对疫情
Google DeepMind 与 Isomorphic Labs 公布生物弹性联合方案,通过防止模型滥用、利用 AlphaFold 和 IsoDDE 等技术加速疫苗与药物设计,以及借助 AlphaEvolve 优化病原体监测,在预防、检…
论文研究
-
HYPIC: Xiaohongshu, Peking University, and SJTU propose first position-independent cache system for hybrid-attention large models
HYPIC implements position-independent caching on hybrid-attention large models, reducing first-token latency by an…
-
Moonshot AI releases PerceptionBench: a diagnostic benchmark for visual perception in multimodal models
Moonshot AI has released PerceptionBench, a visual perception benchmark derived from cases where models actually…
-
Apple 研究:Embarrassingly Simple Self-Distillation 提升代码生成能力
Apple 机器学习研究提出一种极简的自我知识蒸馏方法,让大语言模型仅使用自身原始输出即可提升代码生成能力,无需验证器、教师模型或额外数据。该方法在多个代码生成基准上取得显著改进,且计算成本极低。研究展示了自我蒸馏在代码任务中的有效性,为…
-
Apple 用混合文本与 ID 嵌入实现增量视频搜索个性化排序
Apple 提出一种混合文本与 ID 嵌入方法,用于增量视频搜索中每次按键后的高质量排序。该方法在意图不明确(如仅输入 1–3 个字符)时仍能提升个性化排序效果。研究来自 Apple Machine Learning Research,未…
技巧与观点
-
企业AI智能体评估存在“现实对齐”缺口:半数组织曾将通过内部测试的智能体部署到生产环境后导致客户故障
对157家企业的调查显示,50%的组织在过去一年曾部署通过内部评估但导致客户故障的AI智能体或大语言模型功能,5%的企业完全信任自动化评估,29%认为评估与现实结果对齐不佳是最大局限。尽管信任度低,66%的企业已允许或正计划在12个月内实…
-
Patter SDK tutorial: building a restaurant reservation phone agent with dynamic variables, guardrails, latency dashboard, and evaluation checks
Patter SDK has released a tutorial demonstrating how to build a voice agent workflow for a restaurant reservation…
-
Mermaid-to-Unicode diagram tool found in xAI's open-source Grok CLI codebase
The codebase of xAI's open-source Grok CLI coding agent includes a Mermaid diagram sample terminal renderer written in…
-
Remote agent operation setup: Codex as primary, UU Remote as fallback
The author shares a combined setup for using agents remotely: Codex's remote control as the primary method, connecting…
-
Anthropic migrates code at scale with Claude Code: Bun's million lines of Zig to Rust in two weeks
Anthropic engineers used Claude Code to migrate Bun's million lines of Zig code to Rust in two weeks, with 100% of…
-
Qwen APP and Wuhan Release hold AI job-hunting practical course, demonstrating resume diagnosis and office automation
Qwen APP and Wuhan Release held an AI job-hunting practical course in Wuhan, demonstrating resume diagnosis, PPT…
-
Google 提出模块化提示词转译方案,构建可扩展 AI 智能体
Google 提出将系统提示词模块化为可复用的“技能文件”,通过转译器进行静态验证并在构建时捕获缺失依赖,从而解决单体提示词导致的扩展瓶颈与运行时错误。该方法将提示词生成集成到 CI/CD 流水线中,防止代码漂移,并允许智能体通过标准 p…
-
OpenAI 强调青少年应获得安全 AI 访问权,推出 Study Mode 等保护措施
OpenAI 发文主张青少年应获得安全的 AI 访问权,并分享了针对 teen 用户的保护措施。目前每周有近 9 成 teen 用户使用 ChatGPT 进行学习、信息获取或技能提升。OpenAI 已推出 Study Mode(默认开启引…