模型发布/更新
-
Fun-ASR-Realtime released: a single model supports 30 languages and 16 dialects with leading recognition accuracy
Tongyi Lab released the Fun-ASR-Realtime real-time speech recognition model. A single model covers 30 languages and 16…
产品发布/更新
-
Synthetic Sciences releases OpenScience: an open-source, model-agnostic AI workbench for machine learning, biology, physics and chemistry research
Synthetic Sciences launched OpenScience, an open-source (Apache 2.0) AI research workbench covering machine learning,…
-
xAI adds 21 flagship voices to Grok Voice
xAI released 21 new flagship voices, joining the original 5. All new voices are multilingual and are available in the…
-
Claude Code v2.1.202 released
Claude Code v2.1.202 adds a "Dynamic workflow size" setting in `/config` that controls the scale of agents in dynamic…
-
OpenClaw lands on Hugging Face local app
OpenClaw lands on the @huggingface local app 🦞🤝🤗 1. Pick any GGUF/MLX model on hf 2. Copy the openclaw onboard setup…
-
SGLang integrates DSpark speculative decoding with confidence-driven variable-length verification
The SGLang team has integrated the DSpark speculative decoding algorithm into its open-source inference engine. The…
行业动态
-
2026 tech company AI layoffs: Microsoft, Oracle, GitLab and seven others cut thousands of jobs
Since the start of 2026, multiple tech companies have carried out large-scale layoffs citing AI. Microsoft cut about…
-
Meta reportedly had contractors pose as minors to lure rival AI into sensitive topics
According to Wired, Meta carried out a project codenamed "Cannes" through outsourcing company Covalen, having hundreds…
-
SK Hynix to launch $28 billion US listing, potentially the second-largest IPO ever
SK Hynix launched a US listing plan of about $28 billion on Monday, issuing 17.79 million new shares through…
-
Runway announces Paris office
The original body contains only website cookie settings and provides no specific information about the Paris office's…
-
阿尔伯塔省用Claude进行政府系统安全审查
阿尔伯塔省技术与创新部自2025年起使用Claude Code扫描全部27个省级部门的系统。50个智能体在20小时内并行审查4.66亿行代码,发现传统自动化工具遗漏的漏洞并直接生成修复方案,传统方法估计需约6.5年。团队还构建了红队/蓝队…
-
Cursor 成立 CFO Council,聚焦 AI 支出新经济学
2025 年全球 AI 支出达 1.5 万亿美元,但麦肯锡研究显示仅 39% 组织能追踪至 EBIT 影响。Cursor 成立 CFO Council,旨在建立让 AI 支出可衡量、可预测且与价值挂钩的框架。BCG 分析发现,token…
论文研究
-
Using interpretability to understand annotator safety policies
Annotation disagreement can stem from operational failures, policy ambiguity, or value pluralism. Annotator Policy…
-
Apple 提出专用小型 seq2seq 模型用于 ASR 纠错
Apple 机器学习研究团队采用紧凑的 seq2seq 模型进行 ASR 纠错,训练数据来自真实和合成音频的 ASR 错误。通过级联 TTS 和 ASR 构建合成语料,关键在于匹配真实错误分布的多样性。模型采用 correction-fi…
-
TopoPrimer:预测模型中缺失的拓扑上下文
TopoPrimer 框架将序列群体的全局拓扑结构作为显式输入加入预测模型。通过持续同调与谱坐标预计算,可部署为全训练模型的 per-token 输入或预训练骨干的轻量适配器。在 Chronos 和 TimesFM 的四个基准上,Topo…
技巧与观点
-
Google updates privacy settings to train AI on media data by default, with manual opt-out
In June, Google quietly updated the privacy settings for its search services via customer emails, adding two new…
-
Sharing 8 highly practical prompts to run before Claude Fable 5 goes offline
Claude Fable 5 is about to go offline, and the author compiled 8 battle-tested prompts: a /goal prompt that lets the…
-
Claude Code team details four types of agent loops
The Claude Code team defines a "design loop" as an agent repeating work until a stop condition is met, and divides it…
-
Claude Fable field guide: discovering your unknowns
Claude Fable is the first model that requires users to actively clarify unknowns to get high-quality work.…
-
Google MaxText 弹性训练:训练中途终止 TPU,数秒内恢复
分布式 AI 训练常因单台机器故障导致整个多节点作业崩溃,需耗时重启。Google 的 JAX 生态通过 Pathways 实现弹性训练,将硬件故障转化为可捕获的 Python 异常,使运行进程存活。测试中,训练中途终止一个 TPU 后,…