NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2025-09-12ENGLISH EDITION
This issue
—
All time
—

Hacker News

6 stories
01

VaultGemma: The most capable differentially private LLM

Google Research has introduced VaultGemma, touted as the most capable differentially private Large Language Model to date. This development marks a significant milestone in the field of AI, addressing the critical challenge of balancing powerful language processing capabilities with robust data privacy protections. Differential privacy is a rigorous mathematical framework that ensures individual data points cannot be inferred from the model's outputs, thereby safeguarding sensitive user information during both training and inference. The emergence of VaultGemma suggests that it overcomes previous limitations where applying differential privacy often led to a substantial degradation in model performance. By achieving high capability alongside strong privacy guarantees, VaultGemma opens new avenues for deploying advanced LLMs in sensitive applications, such as healthcare, finance, and government, where data confidentiality is paramount. This innovation is expected to accelerate the adoption of privacy-preserving AI technologies and foster greater trust in large-scale AI systems.

02

Qwen3-Next

Qwen3-Next represents the forthcoming generation in the acclaimed Qwen series of large language models, spearheaded by Alibaba Cloud. Although comprehensive technical specifications and performance metrics are currently under wraps, the designation "Next" strongly indicates a significant leap forward from its predecessors. Industry expectations for such an advancement typically include substantial enhancements in areas like complex reasoning, multimodal integration, and computational efficiency. Furthermore, improvements in model robustness, ethical alignment, and reduced hallucination rates are often key objectives for new foundational models. This strategic development underscores the ongoing global race to innovate in artificial intelligence, promising to deliver more sophisticated tools for natural language understanding, content generation, and diverse AI applications. The introduction of Qwen3-Next is poised to influence the trajectory of generative AI research and practical deployment, offering new capabilities to the broader AI community.

03

OpenAI Grove

OpenAI Grove has been unveiled as a new strategic initiative or platform by OpenAI, marking a potential new direction in the company's ongoing efforts to advance artificial intelligence. While specific details regarding its precise functionalities and overarching goals are yet to be fully disclosed, the chosen name "Grove" strongly suggests an environment designed for growth, collaboration, and the cultivation of ideas or resources. It is widely anticipated that OpenAI Grove will serve as a dedicated ecosystem aimed at fostering significant advancements across various domains of AI. This could manifest as a collaborative framework for researchers and developers, offering enhanced access to cutting-edge AI tools, novel datasets, or establishing a community-driven platform for sharing insights and accelerating the development of next-generation AI models. The initiative is expected to reinforce OpenAI's commitment to pushing the boundaries of AI research and application, potentially facilitating more structured experimentation and the responsible deployment of their advanced technologies. Further official communications are awaited to fully elucidate the scope and transformative impact of OpenAI Grove on the broader artificial intelligence landscape and its community.

04

Show HN: An MCP Gateway to block the lethal trifecta

Inspired by Simon Willison's 'lethal trifecta' concept, a new project introduces an MCP Gateway designed to enhance the security of Large Language Models (LLMs) interacting with multiple Multi-Capability Platform (MCP) servers. This gateway acts as an intermediary, inspecting the tools and requirements of each connected MCP server. It classifies these tools along three critical axes: private data access, untrusted content handling, and external communications. The primary function of the gateway is to identify and block potentially dangerous operations where all three 'trifecta' conditions are about to align within a single session. By intercepting such actions, the gateway prevents hazardous outcomes and prompts the LLM to issue a warning, thereby nudging the user to review the situation before any irreversible or harmful steps are taken. This proactive security measure aims to safeguard against unintended consequences arising from LLM interactions with diverse and potentially risky external services.

05

Show HN: I made a generative online drum machine with ClojureScript

Beat Maker, a new generative online drum machine, has been launched after two years of development, aiming to be a premier free web-based solution. Constructed almost entirely as a client-side application using ClojureScript and functioning as a Progressive Web App (PWA), it offers an app-like experience. The project's core objective was to blend user-friendliness for novices with robust capabilities for experienced producers, informed by extensive research into existing drum machine user experiences. A key innovation is its procedural sample generation, which provides an infinite supply of unique sounds at the click of a button, thereby streamlining the beat-making process by eliminating the need for manual sample searching. This feature, combined with a standard grid editor, positions Beat Maker as a distinctive and powerful tool in digital music production, offering a fresh perspective on creative sound design and beat composition for a wide range of users.

06

K2-think: A parameter-efficient reasoning system

K2-think introduces a novel parameter-efficient reasoning system designed to address the computational demands of complex AI tasks. This innovative approach focuses on achieving robust reasoning capabilities with a significantly reduced number of parameters compared to conventional models. The system aims to enhance the scalability and deployability of AI, making advanced reasoning accessible in environments with limited computational resources. By optimizing model architecture and potentially leveraging new algorithmic paradigms, K2-think demonstrates a significant step towards more sustainable and efficient artificial intelligence. The core idea revolves around striking an optimal balance between model complexity and reasoning performance, ensuring that high-quality analytical and inferential tasks can be executed without the need for massive computational overhead. This development is particularly relevant for edge computing, mobile AI applications, and scenarios where rapid inference and energy conservation are critical. The research highlights the potential for future AI systems to be both powerful and resource-conscious, pushing the boundaries of what is achievable with compact AI models.

Twitter

6 stories
01

ZoubinGhahrama1_VaultGemma Launch

Google Research has introduced VaultGemma, a significant advancement in open-source AI models. This model stands out as the largest trained from scratch with differential privacy, a crucial technique for protecting sensitive data during the training process. The development signifies a commitment to both powerful AI capabilities and robust data security. VaultGemma's open nature allows for broader research and application, potentially accelerating innovation in areas requiring privacy-preserving machine learning. The announcement highlights Google's ongoing contributions to the AI community, particularly in developing large-scale models with enhanced privacy features, making advanced AI more accessible and secure for various use cases.

02

OpenAI Reasoning Models Advance

The tweet highlights significant advancements in OpenAI's reasoning models, contrasting the capabilities of the o1-preview released a year ago with current models. The latest models can now process information for hours, browse the web, and write code, demonstrating a substantial leap in performance. The author expresses excitement about the potential for further improvements in reasoning capabilities over the next year, indicating a strong focus on pushing the boundaries of AI reasoning and functionality. This progress suggests a rapid development cycle within OpenAI, particularly in areas requiring complex cognitive tasks and extended operational capacity.

03

ChatGPT Email Leak

A recent tweet by GaryMarcus highlights a concerning security vulnerability where ChatGPT was allegedly used to leak private email data. The exploit reportedly requires only the victim's email address to access sensitive information. This incident raises significant questions about the security protocols and data protection measures in place for AI models like ChatGPT. The tweet implies a serious breach of privacy, potentially impacting users who interact with or share data through such platforms. Further details are expected regarding the specifics of the exploit and the extent of the data compromised, emphasizing the need for robust security in AI development and deployment.

04

GaryMarcus_Sora Physics Critique

Gary Marcus, a prominent AI researcher, has shared an essay critiquing Sora, OpenAI's text-to-video model. Marcus suggests that Sora does not fundamentally grasp the principles of physics, a viewpoint he articulated in an essay written shortly after Sora's release. He notes that this perspective was not widely accepted or understood at the time of his writing. The tweet also mentions Rohan Paul, indicating a discussion or engagement with the AI community regarding Sora's capabilities and limitations. The shared link points to a more detailed analysis of his arguments concerning Sora's understanding of physical laws and its implications for generative AI development.

05

Thom_Wolf_Hugging Face MCP

The tweet highlights the integration of Hugging Face's Model Card Platform (MCP) into Manus AI's HQ. This integration allows agents developed by Manus AI to leverage Hugging Face's extensive ecosystem, which includes access to over 2 million models and 500,000 datasets. The announcement emphasizes the significant expansion of capabilities for these AI agents by enabling them to utilize a vast repository of pre-trained models and data, thereby enhancing their performance and versatility across various AI tasks. This collaboration signifies a step forward in making advanced AI models more accessible and integrated into AI agent development workflows.

06

gdb_GPT 5 Pro Rumor

A recent tweet from user gdb suggests that OpenAI may be preparing to release GPT 5 Pro within the next year. While the tweet is brief, it points towards a potential advancement in OpenAI's flagship language model series. The mention of 'GPT 5 Pro' implies an enhanced or premium version of the upcoming GPT-5, which is highly anticipated in the AI community. Such a release could signify significant improvements in performance, capabilities, or specialized features compared to previous iterations. The timeframe of 'in a year' provides a speculative but concrete outlook for this potential product development. This information, if accurate, would be a key indicator of the pace of innovation in large language models and the competitive landscape of AI development.

GitHub

6 stories
01

DeepSeek-V3

DeepSeek-V3 is a state-of-the-art large language model developed by DeepSeek AI, representing a significant leap in artificial intelligence capabilities. Positioned for broad accessibility, as evidenced by its presence on Hugging Face and a dedicated chat platform, the model is designed to empower both academic research and practical industrial applications. DeepSeek-V3 is engineered with advanced architectural innovations and trained on extensive, diverse datasets, enabling it to excel in a wide spectrum of natural language processing tasks. Its core functionalities encompass sophisticated text generation, nuanced comprehension, efficient summarization, and highly engaging conversational abilities. These features make it an invaluable tool for developing intelligent agents, enhancing human-computer interaction, automating complex content creation workflows, and facilitating data analysis. DeepSeek-V3's introduction underscores DeepSeek AI's commitment to delivering powerful, accessible AI technologies, driving innovation and expanding the frontiers of what is possible with large language models across various domains.

02

Grok-1

This repository offers JAX example code designed for loading and executing the Grok-1 open-weights model, a formidable large language model boasting 314 billion parameters. It outlines the necessary steps for users to download the model checkpoint (`ckpt-0`) and run a Python script to perform inference on a test input. Key technical specifications of Grok-1 are detailed, including its Mixture of 8 Experts (MoE) architecture, which employs 2 experts per token, distributed across 64 layers with 48 attention heads. The documentation emphasizes the significant GPU memory required to run such a large model. While the current MoE layer implementation prioritizes correctness validation over efficiency, avoiding custom kernels, this project serves as a crucial resource for developers and researchers aiming to explore and deploy state-of-the-art, large-scale MoE language models within the JAX ecosystem.

03

GitHub MCP Server

The GitHub MCP Server is an innovative platform designed to bridge AI tools directly with GitHub's ecosystem, empowering AI agents, assistants, and chatbots to interact seamlessly with repositories. It facilitates a wide range of operations through natural language, including reading code, managing issues and pull requests, analyzing codebases, and automating development workflows. Key use cases span comprehensive repository management, allowing AI to browse code, search files, and understand project structures. It also streamlines issue and PR automation, enabling AI to triage bugs, review changes, and maintain project boards. Furthermore, the server offers CI/CD and workflow intelligence by monitoring GitHub Actions, analyzing build failures, and managing releases. Advanced code analysis features include examining security findings and Dependabot alerts, providing deep insights into code patterns. This integration significantly enhances team collaboration by enabling AI-driven access to discussions and notification management, ultimately boosting developer productivity and project efficiency.

04

GPT-SoVITS-WebUI

GPT-SoVITS-WebUI presents a robust, web-based solution for few-shot voice conversion and text-to-speech (TTS) synthesis. This project empowers users to generate high-quality speech from text or transform existing voices with remarkable efficiency, requiring only a small amount of input data. Built with Python, supporting versions 3.10 to 3.12, it ensures compatibility and performance. A key feature is its integration with Google Colab for training, significantly lowering the barrier to entry for developing custom voice models. The intuitive WebUI abstracts the underlying deep learning complexities, making advanced voice AI accessible to a wider audience. This tool is ideal for applications in content creation, personalized digital assistants, and accessibility technologies, offering a streamlined workflow for rapid prototyping and deployment of sophisticated voice AI capabilities. Its focus on few-shot learning positions it as a cutting-edge solution in the evolving landscape of generative audio.

05

Enable AI to control your browser 🤖

Browser Use is an innovative project focused on enabling artificial intelligence systems to directly control web browsers. This initiative aims to provide a robust framework for AI agents to interact with, navigate, and manipulate web content, thereby automating a wide range of online tasks. The core functionality involves empowering AI to perform actions typically requiring human intervention, such as data extraction, form submission, and complex web navigation. This capability is crucial for developing advanced AI-driven workflows, automated testing, and intelligent personal assistants that operate within the web environment. The project fosters community engagement through GitHub for collaboration, Discord for support, and comprehensive documentation. The mention of a 'Cloud' service suggests potential for scalable, hosted solutions, making AI browser control accessible without extensive local setup. Browser Use represents a significant advancement in AI automation and intelligent web interaction.

06

OpenAI Codex CLI

OpenAI Codex CLI is a powerful, locally-running coding agent designed to assist developers directly from their command line interface. This tool provides a distinct alternative to integrated development environment (IDE) extensions, which are available for platforms like VS Code, Cursor, or Windsurf, as well as the cloud-based Codex Web service. Users can easily install the CLI globally using popular package managers such as npm (`npm install -g @openai/codex`) or Homebrew (`brew install codex`), and then initiate the agent by simply typing `codex`. The agent is engineered to bring OpenAI's advanced coding capabilities directly to the user's local machine, enabling efficient code generation, debugging, and other programming assistance without relying on a web browser or specific IDE. It streamlines the development workflow by offering immediate, on-demand coding support, making it a versatile tool for various programming tasks and environments, enhancing productivity for developers who prefer a command-line-centric approach to their work.

huggingface

6 stories
01

Kling-Avatar:基于多模态指令的级联长时程虚拟形象动画合成

近期音频驱动的虚拟形象视频生成技术在视听真实感方面取得了显著进展。然而,现有方法将指令条件处理为仅由声学或视觉线索驱动的低级跟踪,未能对指令所传达的交流目的进行建模。这一局限性损害了其叙事连贯性和角色表现力。为了弥补这一不足,我们引入了Kling-Avatar,一个新颖的级联框架,它将多模态指令理解与照片级真实感肖像生成相结合。我们的方法采用两阶段流水线。在第一阶段,我们设计了一个多模态大语言模型(MLLM)导演,根据多样化的指令信号生成蓝图视频,从而控制角色动作和情感等高级语义。在第二阶段,在蓝图关键帧的指导下,我们使用首尾帧策略并行生成多个子片段。这种从全局到局部的框架在忠实编码多模态指令背后高级意图的同时,保留了细粒度细节。我们的并行架构还支持快速稳定地生成长时程视频,使其适用于数字人直播和视频博客等实际应用。为了全面评估我们的方法,我们构建了一个包含375个精选样本的基准,涵盖了多样化的指令和具有挑战性的场景。大量实验表明,Kling-Avatar能够生成高达1080p和48 fps的生动、流畅、长时程视频,在唇形同步准确性、情感和动态表现力、指令可控性、身份保持和跨领域泛化方面取得了卓越性能。这些结果确立了Kling-Avatar作为语义接地、高保真音频驱动虚拟形象合成的新基准。

02

OmniEVA:通过任务自适应三维接地和具身感知推理实现的具身多功能规划器

多模态大型语言模型(MLLMs)的最新进展为具身智能开辟了新机遇,实现了多模态理解、推理和交互,以及连续的空间决策。然而,当前基于MLLM的具身系统面临两个关键限制。首先是几何适应性差距:仅在2D输入上训练或通过硬编码3D几何注入的模型,要么空间信息不足,要么2D泛化受限,导致在具有不同空间需求的任务中适应性差。其次是具身约束差距:以往的工作常常忽略真实机器人的物理约束和能力,导致任务规划在理论上有效但在实践中不可行。为了解决这些差距,我们引入了OmniEVA——一个具身多功能规划器,它通过两项关键创新实现了先进的具身推理和任务规划:(1) 任务自适应3D接地机制,引入门控路由器根据上下文需求执行3D融合的显式选择性调节,从而为多样化的具身任务实现上下文感知的3D接地。(2) 具身感知推理框架,将任务目标和具身约束共同纳入推理循环,从而产生既面向目标又可执行的规划决策。大量的实验结果表明,OmniEVA不仅实现了最先进的通用具身推理性能,而且在广泛的下游场景中也展现出强大的能力。对一系列提出的具身基准(包括基本任务和复合任务)的评估证实了其鲁棒和多功能的规划能力。项目页面:https://omnieva.github.io

03

LoCoBench:复杂软件工程中长上下文大型语言模型的基准测试

上下文窗口扩展到数百万个令牌的长上下文语言模型的出现,为复杂的代码理解和软件开发评估创造了新的机会。我们提出了 LoCoBench,这是一个专门设计用于在真实、复杂的软件开发场景中评估长上下文大型语言模型的综合基准。与专注于单函数完成或短上下文任务的现有代码评估基准不同,LoCoBench 解决了长上下文能力的关键评估空白,这些能力需要理解整个代码库、跨多个文件进行推理以及在大规模软件系统中保持架构一致性。我们的基准提供了 8,000 个评估场景,这些场景系统地生成自 10 种编程语言,上下文长度跨越 10K 到 1M 令牌,这是一个 100 倍的变化,能够精确评估真实软件开发环境中长上下文性能的下降。LoCoBench 引入了 8 个任务类别,捕捉了基本长上下文能力:架构理解、跨文件重构、多会话开发、错误调查、功能实现、代码理解、集成测试和安全分析。通过一个 5 阶段的管道,我们创建了多样化、高质量的场景,挑战大型语言模型以前所未有的规模对复杂代码库进行推理。我们引入了一个全面的评估框架,包含 4 个维度上的 17 个指标,其中包括 8 个新的评估指标,并结合成一个 LoCoBench 分数(LCBS)。我们对最先进的长上下文模型的评估揭示了显著的性能差距,表明复杂软件开发中的长上下文理解是一个尚未解决的重大挑战,需要更多关注。LoCoBench 已发布于:https://github.com/SalesforceAIResearch/LoCoBench。

04

SimpleVLA-RL:通过强化学习扩展VLA训练

视觉-语言-动作 (VLA) 模型最近已成为机器人操作的强大范式。尽管大规模预训练和监督微调 (SFT) 取得了显著进展,但这些模型面临两个基本挑战:(i) SFT 扩展所需的大规模人工操作机器人轨迹的稀缺性和高成本,以及 (ii) 对涉及分布偏移的任务泛化能力有限。大型推理模型 (LRM) 的最新突破表明,强化学习 (RL) 可以显著增强逐步推理能力,这引出了一个自然的问题:RL 能否类似地改进 VLA 的长程逐步动作规划?在这项工作中,我们引入了 SimpleVLA-RL,一个专为 VLA 模型量身定制的高效 RL 框架。在 veRL 的基础上,我们引入了 VLA 特定的轨迹采样、可扩展并行化、多环境渲染和优化的损失计算。当应用于 OpenVLA-OFT 时,SimpleVLA-RL 在 LIBERO 上取得了最先进的性能,甚至通过我们引入的探索增强策略在 RoboTwin 1.0&2.0 上超越了 pi_0。SimpleVLA-RL 不仅减少了对大规模数据的依赖,实现了鲁棒的泛化,而且在实际任务中显著超越了 SFT。此外,我们在 RL 训练期间发现了一种新颖的“pushcut”现象,即策略发现了超出先前训练过程中所见模式的、以前未见的模式。

05

FLUX-Reason-6M & PRISM-Bench:百万级文本到图像推理数据集与综合基准

开源文本到图像(T2I)模型的进步一直受限于缺乏大规模、以推理为重点的数据集和全面的评估基准,导致与领先的闭源系统相比存在性能差距。为了应对这一挑战,我们引入了FLUX-Reason-6M和PRISM-Bench(精确鲁棒图像合成测量基准)。FLUX-Reason-6M是一个庞大的数据集,包含600万张高质量的FLUX生成图像和2000万条双语(英语和中文)描述,专门用于教授复杂推理。图像根据六个关键特征进行组织:想象力、实体、文本渲染、风格、情感和构图,并设计了明确的生成思维链(GCoT)来提供图像生成步骤的详细分解。整个数据整理耗费了15,000个A100 GPU天,为社区提供了一个大型工业实验室之外前所未有的资源。PRISM-Bench提供了一个新颖的评估标准,包含七个不同的赛道,其中包括一个使用GCoT的艰巨长文本挑战。通过精心设计的提示,它利用先进的视觉语言模型对提示-图像对齐和图像美学进行细致的、与人类对齐的评估。我们对PRISM-Bench上19个领先模型的广泛评估揭示了关键的性能差距,并强调了需要改进的具体领域。我们的数据集、基准和评估代码已发布,以催化下一波面向推理的T2I生成。项目页面:https://flux-reason-6m.github.io/。

06

推理引入了新型投毒攻击,但同时也增加了其复杂性

早期针对大型语言模型(LLMs)的数据投毒攻击研究表明,注入后门是相对容易的。然而,近期LLMs增加了逐步推理能力,这扩展了攻击面,使其包含了中间的思维链(CoT)及其将问题分解为子问题的固有特性。利用这些向量进行更隐蔽的投毒,我们引入了“分解式推理投毒”,其中攻击者仅修改推理路径,而保持提示和最终答案的清洁,并将触发器分散到多个单独无害的组件中。令人着迷的是,尽管注入这些分解式投毒仍然可行,但要可靠地激活它们以改变最终答案(而不仅仅是CoT)却出奇地困难。这种困难源于模型通常能够从在其思维过程中被激活的后门中恢复。最终,似乎一种新兴的后门鲁棒性正源于这些先进LLMs的推理能力,以及推理与最终答案生成之间的架构分离。