模型发布/更新
-
Microsoft releases MAI-Cyber-1-Flash: a 5B active-parameter cybersecurity model powering MDASH to 95.95% on CyberGym
Microsoft released MAI-Cyber-1-Flash, a sparse MoE cybersecurity model with 137B total parameters (5B active…
-
OpenAI launches two new transcription model APIs
We introduced two new transcription models in the API: • GPT-Live-Transcribe: built for low-latency real-time…
产品发布/更新
-
OpenAI releases Codex security CLI and SDK
More open-source goodness. We just released a CLI and TypeScript SDK for finding, validating and fixing security…
-
Gemini API Managed Agents upgrade default to 3.6 Flash, adds environment hooks and free tier
Google DeepMind upgraded the default model for Gemini API Managed Agents to Gemini 3.6 Flash, with support for…
-
Perplexity launches Personal Computer agent for Windows
Personal Computer is now available in the Perplexity Windows app. Personal Computer is a local agent tool for your…
-
Volcano Engine launches Doubao Search service to provide real-time, trustworthy search for AI agents
Volcano Engine officially launched the Doubao Search service, providing AI agents with cross-language, multimodal and…
-
Cursor 在印度推出 Cursor Start 计划,含 Grok 4.5 和 Composer,月费 ₹649
Cursor 面向印度开发者推出新订阅计划 Cursor Start,月费 ₹649(含税),支持 UPI 支付。该计划提供对 Grok 4.5 和 Composer 模型的慷慨访问权限,包含比 Free 计划更多的 agent 请求次数…
行业动态
-
Hugging Face discloses details of autonomous agent cyberattack
The first autonomous agent cyberattack was an unprecedented event that deserves unprecedented transparency. Today, we…
-
Andrew Ng founds LearnVector to bring one-on-one learning with AI
Andrew Ng announced the founding of AI education company LearnVector, with a $100 million investment from Coursera,…
-
Delhi High Court rules OpenAI's use of ANI content to train AI does not infringe copyright
The Delhi High Court ruled that OpenAI's use of Asian News International (ANI) content to train artificial…
-
Anthropic supports petition on pacing AI development
We support this petition, and our CEO, several co-founders and senior employees have signed it. Our research published…
论文研究
-
Claude research on discovering cryptographic weaknesses published
New Anthropic research: using Claude to discover cryptographic weaknesses. The Claude Mythos preview has helped our…
-
Apple 为 Siri Expressive Voices 推出内存高效的音频合成架构
Apple 为 Siri Expressive Voices 推出了一种内存高效的音频合成架构,其 detokenizer 在 AMX 上以约 10ms/步运行,峰值内存仅约 21MB。相比此前设备端系统,该架构将 Mean Opinio…
技巧与观点
-
Sam Altman shifts stance: AI development may need to 'slow down' so society can prepare
OpenAI CEO Sam Altman said AI development speed may need to be "adjusted" so society has time to adapt to new…
-
OpenAI calls for setting the pace of frontier AI development
At the core of our mission is researching how to ensure increasingly powerful AI benefits everyone. We believe that at…
-
Google Search's AI Mode launches 5 new features to help plan offline life
Google Search's AI Mode added 5 tools to help users plan offline activities. Features include: recommending local…
-
How to evaluate LLM providers on latency, throughput and uptime
The same model performs differently across provider endpoints due to infrastructure, quantization, load handling and…
-
Databricks 发布 Genie One:面向业务用户的 AI 协同工作助手
Databricks 推出 Genie One,一款面向业务用户的 AI 协同工作工具,旨在帮助非技术员工通过自然语言与数据交互。Genie One 支持数据查询、报告生成和自动化工作流,无需编写代码即可完成常见业务任务。该工具现已通过…
-
AI 框架(Harness)对模型性能的影响超过模型本身:GPT-5.5 在 Cursor 上得分 87.2%,比 Codex 高 25.7 个百分点
Endor Labs 测试发现,同一模型在不同框架(Harness)上性能差异巨大:OpenAI 的 GPT-5.5 在原生 Codex 框架上功能正确率为 61.5%,在 Cursor 上达 87.2%;Anthropic 的 Opus…
-
LangChain 如何构建 Agent-First 数据栈:自服务分析规模提升 40 倍
LangChain 利用 Hex、dbt、语义模型和可观测性工具构建了一个可信数据智能体,将自服务分析规模提升了 40 倍。该方案通过语义层统一指标定义,结合 Agent 框架实现自然语言查询,并借助可观测性监控查询质量与数据血缘。这一…
-
NVIDIA Jetson 为边缘 AI 和机器人提供紧凑型开发套件
NVIDIA Jetson 平台为边缘 AI 和机器人提供紧凑型开发套件,可放入手提包中。其中 Jetson Orin Nano Super 具备 67 TOPS 的 AI 性能,支持运行 Mistral 等开源模型,所有推理均在本地 G…