模型发布/更新
-
SenseNova U1 image generation model lands on ComfyUI
The SenseNova U1 image generation model is now runnable on ComfyUI and has been highly recognized by reviewers…
-
AntLingAGI releases trillion-parameter model Ring-2.6-1T
AntLingAGI has released Ring-2.6-1T, its trillion-parameter flagship thinking model, which is free to use via…
-
腾讯混元Hy3预览版发布,专注复杂智能体任务
腾讯混元Hy3预览版已开放早期体验,被描述为该系列最强模型。其核心设计面向现实世界有效性,而非单纯追求基准测试分数。模型具备处理复杂智能体任务的能力,采用256K上下文长度,并融合了快慢思维机制的混合专家架构。该模型基于重建的预训练和强化…
产品发布/更新
-
Pareto Code reshapes model selection with market demand
Pareto Code is a new approach that uses real market demand to observe the Pareto frontier. DeepSeek V4 Pro takes first…
-
Luma Agents:从情绪板到完整广告的自动化创作
情绪板曾是最精彩的部分。如今它只是起点。 上传你的参考素材。设定方向。Luma Agents 会将其从情绪板转化为完整的广告。 生成广告 → http://lumalabs.ai/app
-
Claude Code v2.1.139 update
This update introduces several new features and improvements. Core additions include an Agent view for centrally…
-
Anthropic officially launches Claude Platform on AWS
Anthropic has officially launched the Claude Platform on AWS, giving AWS customers a new way to use full Claude…
行业动态
-
OpenAI launches DeployCo to help enterprises build businesses around intelligence
OpenAI has officially launched DeployCo, a new enterprise deployment company aimed at helping organizations put…
-
Report: AI Chip Company Cerebras IPO Gets 20+ Times Oversubscription, Plans to Raise Offering Price by Nearly 30%
AI chip company Cerebras's IPO has received more than 20 times oversubscription and is expected to become the world's…
-
Nadella testifies to counter Musk lawsuit: he has my phone number but never complained about Microsoft's OpenAI partnership
Microsoft CEO Nadella testified in Musk's lawsuit against OpenAI, rebutting Musk's allegations that OpenAI strayed…
-
旧金山AI模型开发者盛会周三举行
本周三在旧金山,与@vercel_dev、@AnthropicAI和@Kimi_Moonshot共同参与美国下一代模型盛会。 每位参与者将获得$30 MiniMax API积分 + 大奖池。 📅 5月13日 | 太平洋时间下午5:30 |…
-
MiniMax组建“10x团队”,邀请领域专家共推AI跨越式发展
MiniMax宣布成立“10x团队”,旨在邀请各领域深度专家共同推动人工智能实现十倍速进步。公司认为大模型已从演示阶段发展为真实生产力设施,而将顶尖领域人才引入模型研发能显著提升能力并产生实际影响。团队寻求的是坚信AI能加速自身领域发展的…
论文研究
-
Agents execute well but fall short on optimizing user interests
Testing through SocialReasoning Bench found a consistent pattern across models: agents can handle execution tasks, but…
-
BalCapRL: 基于强化学习的MLLM图像描述平衡框架
研究团队针对多模态大语言模型图像描述任务提出BalCapRL平衡框架。该框架通过设计多维度奖励函数,系统解决了现有强化学习方法在追求描述效用时引发的幻觉、噪声和冗长等问题。实验表明,BalCapRL在保持描述准确性的同时,显著提升了信息密…
技巧与观点
-
Using LLMs in a script's shebang line
This article introduces an innovative approach that embeds LLM tools directly in a script's shebang line to execute…
-
人机交互界面的演进:从文本到交互式神经视频
作者认为,AI的人机交互界面正从当前的Markdown默认输出向更丰富的HTML格式演进,后者能提供更好的图形、布局和交互性。长远看,由扩散神经网络直接生成的交互式神经视频或模拟将是终极输出形式,但其如何与精确的传统软件结合仍是开放问题。…
-
Fields Medalist tests ChatGPT 5.5 Pro: paper-level results in 17 minutes, sounding a red alert for students
Fields Medalist Timothy Gowers tested ChatGPT 5.5 Pro, and the AI independently solved an open additive number theory…