模型发布/更新
-
Ant Lingbo open-sources real-time interactive world model LingBot-World 2.0
Ant Lingbo open-sourced the new-generation real-time interactive world model LingBot-World 2.0 (14B parameters),…
-
OpenAI 发布最强模型与最佳博文
显然是我们有史以来最好的模型,也是我们写得最好的博文之一: https://openai.com/index/gpt-5-6/
-
Ant Lingbo open-sources LingBot-Video, the world's first MoE video foundation model for embodied intelligence
Ant Lingbo Technology officially open-sourced LingBot-Video, the world's first MoE-architecture video generation…
-
NVIDIA releases Nemotron-Labs-3-Puzzle-75B-A9B: compressed hybrid MoE model with 2.03x higher server throughput
NVIDIA released Nemotron-Labs-3-Puzzle-75B-A9B, a compressed variant of Nemotron-3-Super, reducing total parameters…
-
Robbyant releases LingBot-VLA 2.0: an open-source 6B cross-embodiment robot vision-language-action model
Robbyant released LingBot-VLA 2.0, a 6B-parameter open-source vision-language-action (VLA) foundation model. It uses…
-
Meta releases Muse Spark 1.1 model
From @finkd — Muse Spark 1.1 is now live.
产品发布/更新
-
Claude launches reflection feature (Beta)
Anthropic launched a reflection feature (Beta) for Claude to help users track usage patterns. Users can review…
-
ChatGPT Sites Turns Ideas into Publishable Websites
Turn an idea into a live website that can be published and shared. Here are some examples built with Sites by members…
-
Google Launches LiteRT.js: High-Performance Web AI Inference Runtime
Google released LiteRT.js, the newest member of the LiteRT cross-platform edge AI runtime, designed for JavaScript…
-
Mistral 推出 Studio,为 AI 提示词和技能提供系统记录
Mistral 今日推出 Studio,为 AI 提示词和技能提供集中式系统记录。平台将 prompts 和 skills 视为生产资产,支持不可变版本、回滚、明确所有权、分类标签和审计日志,保证变更可追溯。非开发者可直接编辑测试,通过标…
-
OpenAI launches ChatGPT Work: an AI agent that can work autonomously across apps
OpenAI released ChatGPT Work, an AI agent that can gather information across apps and files, break complex projects…
行业动态
-
Elon Musk称赞Anthropic并承诺不切断其算力
Elon Musk近日在X上承认此前对Anthropic的判断有误,称其“显然是当前AI领域的领导者”,盛赞Mythos/Fable模型“目前最好”,并承诺不会恶意切断其计算资源。2026年7月起,Anthropic成为SpaceX最大客…
-
Report: Tesla Optimus Gen 3 design preliminarily finalized; Musk threatens to fire entire procurement team if production targets are missed
According to LatePost, Tesla's Optimus Gen 3 passed Musk's review and is about to enter mass production. The supply…
-
Ollama reaches 8.9 million developers, Series B led by Theory
Ollama makes it easy to run open-source models locally or in the cloud with a consistent experience. It now has 8.9…
-
Anthropic launches 'Hard Questions' initiative, inviting the public to raise sharp AI-related questions
As a public benefit corporation, Anthropic launched the 'Hard Questions' initiative, inviting the public to raise the…
-
Anthropic's Long-Term Benefit Trust appoints Ben Bernanke as trustee
Anthropic's Long-Term Benefit Trust (LTBT) appointed former Federal Reserve Chair and 2022 Nobel Economics Prize…
-
GPT-5.5 生物漏洞赏金计划
OpenAI 将 GPT-5.5 Bio Bug Bounty 升级为持续私密项目 OpenAI Bio Bounty Program,以 GPT-5.6 为起点并覆盖后续前沿模型,旨在发现能突破预设生物安全挑战的通用越狱攻击。奖励从 $…
论文研究
-
揭密在线策略蒸馏:何时有益、何时有害及原因
Apple机器学习研究团队提出训练无关诊断框架,以每个token、每个问题、每个教师的分辨率分析on-policy蒸馏。通过可扩展targeted-rollout算法估计理想梯度,并计算蒸馏梯度与理想梯度的余弦相似度(梯度对齐分数)。实验…
-
TGPO:通过可验证奖励强化学习激励第一人称视频时序感知
多模态大语言模型(MLLM)在第一人称视频理解中缺乏时序感知,常依赖空间捷径。为此,研究者提出 Temporal Global Policy Optimization(TGPO),一种基于可验证奖励的强化学习算法。TGPO 通过对比模型在…
-
Apple 提出 SRLM:自反思程序搜索提升长上下文处理能力
Apple 机器学习研究团队提出 SRLM 框架,利用自一致性、推理链长度和口头置信度三种内在信号,让模型在推理时评估候选长上下文交互程序。实验表明,在相同时间预算下,SRLM 较传统递归语言模型(RLM)最高提升 22%。分析发现,递归…
技巧与观点
-
Can AI Answer the $3 Trillion Question?
Sequoia partner David Cahn updated his AI infrastructure spending estimates: global investment will reach $1.5…
-
TeXada: local math agent based on MiniCPM released
Community developers built TeXada, a local-first math agent based on MiniCPM5-1B and MiniCPM-V 4.6. The agent supports…