模型发布/更新
-
FastWan-QAD: 5-second video generated in 1.8 seconds on a single 5090
Sky Computing Lab released the FastWan-QAD video generation model series, trained with FastVideo's quantization-aware…
-
JD open-sources full-stack JoyAI-VL-Interaction, moving from Q&A to watching while talking
JD recently open-sourced JoyAI-VL-Interaction, the world's first full-stack interaction model, with native support…
-
NetEase Youdao releases Confucius4-TTS: open-source 14-language cross-lingual accent-free voice cloning model
NetEase Youdao launched the Confucius4-TTS engine of its Ziyue 4.0, claiming it is the industry's first open-source…
-
Krea 2 technical report officially released
Our technical report has been published. It takes a deep dive into the data, architecture and training techniques used…
-
Doubao audio generation model 1.0 released, redefining AI audio creation
Volcano Engine officially released the Doubao audio generation model 1.0 (Doubao-Seed-Audio 1.0), which supports text…
-
Mistral releases OCR 4 document understanding model with bounding boxes, block classification and confidence scores
Mistral released OCR 4, which returns bounding boxes, typed block classifications and per-word confidence scores in…
-
Seed2.1 officially released, deepening AI productivity
ByteDance Seed released the Seed2.1 series, agents for real productivity scenarios that strengthen general agent…
产品发布/更新
-
IBM open-sources CUGA: lightweight agent framework with over 20 single-file example apps
IBM open-sourced CUGA (Configurable Generalist Agent), a lightweight agent framework that handles planning, execution…
-
China's first AI evaluation for college entrance exam applications released, Qwen outperforms senior consultants in multiple areas
Yousong Lab released China's first AI capability evaluation report for college entrance exam application, testing four…
-
Anthropic launches Claude Tag: collaborate in Slack via @Claude
Anthropic launched Claude Tag, a new collaboration method for delegating tasks in Slack channels via @Claude. Claude…
-
Claude Code v2.1.187 发布
Claude Code v2.1.187 新增 `sandbox.credentials` 设置,可阻止沙箱化命令读取凭证和秘密环境变量;模型选择器及相关参数现已支持组织配置的模型限制,选中受限模型时显示“受组织设置限制”提示。全屏模式下…
行业动态
-
Oracle cuts 21,000 jobs due to AI adoption, debt-funded cloud infrastructure investment
Oracle cut 21,000 jobs in the fiscal year ended May 31, reducing its headcount to 141,000, a 12.9% decline. The…
-
GitHub and open-source coalition call for changes to California AI transparency bill to protect open source
GitHub, together with Black Forest Labs, Hugging Face, and Mozilla Corporation, formed an open-source coalition…
-
OpenAI 助力 Appia Foundation 推动先进 AI 共享标准建设
OpenAI 通过 Appia Foundation 支持制定先进 AI 的共享标准,涵盖评估框架、安全实践与全球合作。
-
Omio 如何构建对话式旅行的未来
Omio 利用 OpenAI 技术打造对话式旅行体验,加速产品开发进程,并推动自身向 AI 原生公司转型。
论文研究
-
Nine judges, two effective votes: correlation errors weaken LLM review panels
Apple's machine learning research team found that LLM-as-a-judge panels are severely limited by high correlation among…
-
基于指标依赖的标注饱和:从标签分布中学习
在ChaosNLI数据集(每项100个标注)上微调NLI模型,发现所需标注人数因评估指标而异:熵相关(识别分歧项)需约20-50个标注者收敛,KL散度(分布匹配)约10个标注者即饱和(达全量效果的87%-95%)。软标签的熵相关r=0.6…
技巧与观点
-
huggingface_hub achieves weekly releases: AI, open-source tools and human review in a closed loop
Hugging Face shortened the release cycle of huggingface_hub from every 4–6 weeks to weekly, fully automated by a…
-
Experimenting with a proposed Cross-Origin Storage API in Transformers.js
When Transformers.js runs AI models in the browser, web apps from different origins repeatedly download and cache the…
-
Meta 如何为 AI 眼镜设计超窄钢壳电池
Meta 工程团队为 Ray-Ban Meta 等智能眼镜开发了宽度仅 7mm 的钢壳电池。传统软包电池难以塑形且空间利用率低,Meta 改用叠片式电极结构以降低阻抗、避免多任务时电压骤降,并将公差控制在约 100 微米以释放更多体积。G…
-
Claude Tag 的 Agent Identity 访问模型
Claude Tag 推出 agent identity(智能体身份)访问模型,让 Claude 在共享频道中以独立身份工作,而非模拟某个用户。管理员在工作区级配置连接器、仓库访问、技能插件和固定指令等权限,每个频道可覆盖继承的基线设置。…
-
GPT-5 帮助免疫学家 Derya Unutmaz 解开三年未解之谜
GPT-5 Pro 帮助免疫学家 Derya Unutmaz 解决了一个长达三年的免疫学谜团,揭示了 T 细胞行为的新见解。这一突破可能为癌症和自身免疫疾病研究提供支持。