模型发布/更新
-
推出 Grok 4.5
Cursor 与 SpaceXAI 联合训练了混合专家模型 Grok 4.5,在数万亿 tokens 的 Cursor 用户交互数据上训练,并通过强化学习解决软件工程、数据科学、金融、法律等领域的困难问题。基础版定价 $2/M 输入 to…
-
OpenAI releases GPT-Live, a new generation of full-duplex voice model
OpenAI today launched GPT-Live, built on a full-duplex architecture that enables simultaneous listening and speaking,…
-
SpaceXAI 发布 Grok 4.5,与 Cursor 联合训练的编程与智能体模型
SpaceXAI 推出 Grok 4.5,称其迄今最聪明,聚焦编程、智能体任务和知识工作。该模型与 AI 编程编辑器 Cursor 联合训练,在数万块 NVIDIA GB300 GPU 上完成训练和强化学习。基准测试中,Grok 4.5…
-
Ant Group's Robbyant open-sources LingBot-Vision: a 1B-parameter boundary-centric vision foundation model for dense spatial perception
Robbyant, Ant Group's embodied intelligence company, has open-sourced LingBot-Vision, a family of self-supervised…
产品发布/更新
-
vLLM transformers modeling backend at native speed
Hugging Face announced that the transformers vLLM backend now matches or exceeds the speed of hand-written native vLLM…
-
OpenRouter Chat launches one-click zero data retention
New feature: one-click ZDR (zero data retention) in Chat. Compare models side by side with full privacy protection:…
-
Runway Dev 发布
Runway 官网以 Cookie 设置页面代替了产品介绍,未提供 Runway Dev 的功能、参数、可用性等任何具体信息。
-
利用 Netpreme X‑Mem™ MPU 加速 SGLang HiCache
Netpreme X‑Mem™ 内存处理单元(MPU)作为专用高带宽 KV 内存层与 SGLang HiCache 集成,替代主机 DRAM 作为 L2 卸载层。在基于 Claude Code 代理轨迹的编码工作负载中,前缀缓存命中率平均…
-
Claude Code v2.1.205 发布
Claude Code v2.1.205 修复多项 bug,包括 `--json-schema` 在 schema 无效时静默输出非结构化结果、Windows 下工作树删除误删文件、以及目录被删除/锁定导致的崩溃。改进自动模式,执行 `r…
行业动态
-
美国商务部批准OpenAI大规模发布GPT-5.6,Sol明日亮相
美国商务部正式批准OpenAI大规模发布GPT-5.6。OpenAI宣布GPT-5.6 Sol将于本周四完成最后准备后,与Terra和Luna一同面向公众推出。此前因国家安全考量,美国政府要求分阶段发布,仅允许向经批准的有限实体开放。此次…
-
British Columbia plans to sue OpenAI over failure to report violent ChatGPT conversations before school shooting
British Columbia, Canada, announced on July 7 that it will sue OpenAI, alleging it failed to report to law enforcement…
-
Profit exceeds $1 billion, ARR targets $100 billion, Anthropic beats OpenAI to IPO
Anthropic's third-quarter profit is expected to exceed $1 billion, and it confidentially filed for an IPO on June 1.…
-
Lawsuit: man used Grok to create 7,000 sexual images of stepdaughter before suicide
A man died by suicide after using Grok to generate 7,000 child sexual abuse material (CSAM) images of his…
-
OpenAI releases principles for government and national security partnerships
OpenAI recently published national security principles outlining its approach to deploying frontier AI systems in…
-
GitHub 联合联盟倡议修订 California AI Transparency Act 以保护开源生态
GitHub 加入一个联盟,呼吁对 California AI Transparency Act 进行针对性修订,以解决该法案与开源许可之间的冲突,并与国际透明度框架保持一致,同时保留其监管意图。
论文研究
-
Hackers can use 9 of the most popular AI tools to assemble large-scale botnets
Prompt injection has become the top threat in AI security—large language models cannot distinguish legitimate instructi…
-
OpenAI audit of SWE-Bench Pro finds about 30% of evaluation tasks are flawed
OpenAI conducted a detailed audit of the coding evaluation benchmark SWE-Bench Pro and found that about 30% of tasks…
-
A 'switch' for dual-use knowledge in AI models: Anthropic and AE Studio propose GRAM method
Anthropic and AE Studio jointly proposed the Gradient Routing Auxiliary Module (GRAM) method. By adding removable…
技巧与观点
-
Designing Your Life prompt tested: four stages of life design with Claude
The author turned the theoretical framework of Stanford's Designing Your Life into a prompt, with Claude asking…
-
Ant Group's Zhou Jun at AICon: from token count to token density, efficiency first for trillion-parameter models
Ant Group Vice President Zhou Jun said at AICon that a trillion-parameter model's compute cost for every 15 minutes of…
-
AI preflight check: an agent working memory architecture
A preflight working memory architecture designed for AI agents: when a query arrives, the system retrieves the most…
-
Anthropic 市场运营团队用 Claude Cowork 自动化报告与活动构建
Anthropic 市场运营团队的 Ian Chan 和 Annabel Custer 利用 Claude Cowork 将手动工作从多天压缩至数小时。Ian 此前每周花 1—2 天整合营销指标报告,现在通过 Claude Cowork…
-
NVIDIA 发布 Nemotron 开放数据集与配套工具,支持 AI 智能体开发
NVIDIA Nemotron 开放数据集包含超过 10 万亿预训练 tokens 和数百万后训练样本,覆盖多领域与工具使用场景,支持 AI 智能体开发。同步推出 Nemotron Post-Training v3 Prompt Atla…