NO/FOMO

每天一次，过滤 AI 噪音

值得打开的
AI 日报。

发布日期2025-09-23中文版本
本期阅读
—
累计阅读
—

Hacker News

6 stories
01

Launch HN: Strata (YC X25) – One MCP server for AI to handle thousands of tools

Klavis AI has launched Strata, an open-source Multi-Capability Platform (MCP) server designed to significantly improve how AI agents manage and utilize a vast number of API tools. Addressing critical challenges faced by AI systems, Strata prevents agents from being overwhelmed by thousands of options. Instead of presenting all available tools upfront, it introduces them incrementally, based on the agent's real-time needs. This innovative approach tackles common issues such as AI agents struggling to select the correct API from a large pool, the substantial token budget consumed by extensive tool descriptions, and the typical server limitations that cap tool usage at around 40-50. Developed by a former Senior SWE from Google Gemini's tool use team, Strata aims to enhance AI's capability to discover and employ relevant tools efficiently, mimicking human-like discovery processes to guide agents through tool categories.

AI AgentsAPI ToolsTool UseMCP ServerOpen SourceArtificial IntelligenceAI AgentLarge Language Model
02

Android users can now use conversational editing in Google Photos

Google has officially launched a significant update for its Google Photos application on Android, introducing a groundbreaking feature called "conversational editing." This innovative functionality empowers users to interact with their photos using natural language prompts, effectively allowing them to describe desired edits rather than navigating complex menus or sliders. For instance, users can command the app to "make the sunset more dramatic," "brighten the foreground subject," or "remove distractions in the background," and the AI-powered system will interpret these instructions and apply the relevant adjustments. This capability is underpinned by advanced artificial intelligence and natural language processing technologies, which interpret nuanced user commands and translate them into precise photo manipulation actions. The introduction of conversational editing is poised to democratize advanced photo enhancement, making sophisticated adjustments accessible and intuitive for all Android users. This update not only streamlines the photo editing workflow on mobile devices but also significantly enhances the user experience by offering a more natural and user-friendly interface for creative expression.

Conversational AIGoogle PhotosAndroidPhoto EditingNatural Language ProcessingArtificial IntelligenceNatural Language ProcessingComputer Vision
03

From MCP to shell: MCP auth flaws enable RCE in Claude Code, Gemini CLI and more

A critical security vulnerability, dubbed "MCP auth flaws," has been discovered, enabling Remote Code Execution (RCE) in several prominent AI development environments, notably Claude Code and Gemini CLI. These flaws pertain to authentication mechanisms within the system, which, if exploited, could allow unauthorized actors to execute arbitrary code, thereby compromising the integrity and security of applications built on or utilizing these platforms. The detailed analysis, presented by Veria Labs, highlights the technical specifics of how these vulnerabilities can be leveraged, emphasizing the severe implications for developers and organizations relying on such AI services. This discovery underscores a growing concern regarding the security posture of advanced AI tools, particularly those offering code generation and command-line interface functionalities. The potential for malicious code injection or data exfiltration through RCE necessitates immediate attention from platform developers to implement stronger authentication protocols and patching strategies. The findings stress the critical importance of secure development practices in the rapidly evolving AI landscape to mitigate risks associated with sophisticated cyber threats targeting AI infrastructure.

Remote Code ExecutionAuthentication BypassAI SecurityClaude CodeGemini CLIArtificial IntelligenceLarge Language ModelGenerative AI
04

Structured Outputs in LLMs

The ability of Large Language Models (LLMs) to generate structured outputs, such as JSON or XML, is becoming increasingly critical for their integration into automated systems and applications. While LLMs excel at producing natural language, ensuring they adhere to specific schemas and formats is challenging yet essential for tasks like data extraction, API calls, and database population. This capability transforms LLMs from mere text generators into powerful data processors, enabling seamless communication with downstream software. Current approaches to achieve reliable structured outputs often involve advanced prompt engineering, where clear instructions and examples guide the model's generation. Furthermore, techniques like constrained decoding are employed to enforce grammar and schema rules during the output process, significantly reducing errors and improving reliability. The development of robust methods for structured output generation is vital for maximizing the utility and operational efficiency of LLMs across a wide array of enterprise and developer-focused applications, addressing a key limitation in their practical deployment.

Large Language ModelStructured OutputPrompt EngineeringConstrained DecodingJSON GenerationLarge Language ModelArtificial IntelligenceMachine Learning
05

Getting AI to work in complex codebases

This analysis addresses the significant challenge of effectively integrating and optimizing AI capabilities, specifically coding agents, within large and intricate software codebases. It highlights that traditional AI methodologies often fall short when confronted with the scale and complexity characteristic of enterprise-level software development, underscoring the necessity for advanced context engineering strategies. The central theme revolves around developing sophisticated approaches to supply AI agents with precise and highly relevant contextual information. This encompasses a broad spectrum of data, including project architecture, structural patterns, specific code segments, and comprehensive documentation. The ultimate goal is to empower AI with enhanced comprehension, more effective problem-solving abilities, and reliable code generation in complex environments. By moving beyond simplistic examples, this work aims to foster robust, practical applications of AI in real-world software engineering, improving its utility for critical tasks such as debugging, feature implementation, and code refactoring.

AI AgentsCoding AgentsContext EngineeringComplex CodebasesSoftware EngineeringAI AgentLarge Language ModelArtificial Intelligence
06

Show HN: Open-source AI data generator (now hosted)

Metabase has announced the availability of its open-source AI dataset generator in a newly hosted version, alongside its existing open-source repository. This development addresses user requests for a solution that removes the overhead of self-hosting, making the tool more accessible for immediate application. Initially shared a few months prior, the AI dataset generator garnered significant interest within the developer community. Users now have the flexibility to choose between the hosted service, available at `https://www.metabase.com/ai-data-generator`, for streamlined deployment, or to leverage the full source code from the GitHub repository (`https://github.com/metabase/dataset-generator`) for self-hosting, customization, and community contributions. A notable enhancement to the open-source repository is the integration of LiteLLM, enabling multi-provider Large Language Model (LLM) support, which significantly expands the generator's versatility and potential applications in diverse AI projects. This dual-option strategy aims to cater to a broader range of users, from those seeking quick integration to developers looking for deep customization and collaboration opportunities in AI data generation.

AI data generatorOpen-source softwareDataset generationHosted servicesLLM integrationArtificial IntelligenceMachine LearningLarge Language Model

GitHub

4 stories
01

像老乡鸡那样做饭

The "CookLikeHOC" GitHub repository offers a community-driven, open-source recipe collection inspired by dishes from the renowned Chinese fast-food chain, Laoxiangji (Home Original Chicken). Its primary goal is to compile and organize all recipes featured in the official "Laoxiangji Dish Traceability Report," enabling users to recreate these popular meals at home. The project clearly states its independent nature, emphasizing its status as an unofficial, consumer-contributed resource. A notable feature includes the integration of AI-generated images for select stew dishes, complementing user-contributed real-life photographs. An accessible web interface further enhances user experience. This initiative serves as a practical example of leveraging public information for culinary education and highlights the potential of collaborative data curation in creating valuable knowledge bases, while also incorporating modern AI capabilities for content enrichment.

Recipe CollectionCulinary ArtsChinese CuisineFood RecipesOpen-source DataGenerative AIArtificial IntelligenceOthers
02

Tongyi DeepResearch

Tongyi DeepResearch, developed by Tongyi Lab, is an advanced agentic large language model with 30.5 billion parameters, designed for complex, long-horizon information-seeking tasks. It leverages a highly scalable and fully automated synthetic data generation pipeline, enabling robust agentic pre-training, supervised fine-tuning, and reinforcement learning. The model incorporates large-scale continual pre-training on diverse agentic interaction data to maintain currency and enhance reasoning capabilities. Its unique end-to-end reinforcement learning approach uses a customized Group Relative Policy Optimization framework, ensuring stable training in dynamic environments. Tongyi DeepResearch supports both ReAct and an IterResearch-based 'Heavy' inference mode for flexible performance optimization. It has achieved state-of-the-art results across various agentic search benchmarks, including Humanity's Last Exam and BrowserComp, and is accessible via HuggingFace, ModelScope, and OpenRouter, signifying its broad utility in cutting-edge AI research.

Agentic Large Language ModelInformation SeekingReinforcement LearningData SynthesisPre-trainingArtificial IntelligenceLarge Language ModelAI Agent
03

tldraw

tldraw is an open-source monorepo that delivers a robust React library specifically designed for crafting infinite canvas experiences, famously underpinning the popular digital whiteboard platform, tldraw.com. Developers can effortlessly integrate tldraw into their React projects through npm, providing a highly customizable and interactive drawing surface. The project emphasizes developer experience, offering clear guidelines for local development using `yarn` and corepack. A unique feature tailored for AI agents is the inclusion of `CONTEXT.md` files throughout the repository, complemented by a `context` script, enabling quick codebase comprehension for automated tools. tldraw operates under its specific license, permitting both commercial and non-commercial applications, contingent on retaining the 'Made with tldraw' watermark, with a business license available for its removal. This library positions itself as a crucial resource for building advanced interactive web applications, particularly those requiring dynamic drawing, collaboration, and expansive visual workspaces.

ReactInfinite CanvasDigital WhiteboardSDKFrontend DevelopmentAI AgentArtificial IntelligenceOthers
04

Elasticsearch

Elasticsearch is a powerful, distributed search and analytics engine that functions as a scalable data store and vector database, optimized for high performance and relevance across production workloads. It forms the core of Elastic's open Stack platform, enabling near real-time search over massive datasets, advanced vector searches, and seamless integration with generative AI applications. Key use cases include Retrieval Augmented Generation (RAG), full-text search, and monitoring of logs, metrics, and application performance (APM), alongside security analytics. Users can get started with managed deployments on Elastic Cloud or by installing it locally via Docker, using the `start-local` script for development. Elasticsearch interacts primarily through REST APIs, supporting various programming language clients and Kibana's Dev Tools for data indexing, search, and exploration.

ElasticsearchDistributed SearchAnalytics EngineVector DatabaseRetrieval Augmented GenerationArtificial IntelligenceMachine LearningGenerative AI

huggingface

6 stories
01

LIMI: Less is More for Agency

We define Agency as the emergent capacity of AI systems to function as autonomous agents actively discovering problems, formulating hypotheses, and executing solutions through self-directed engagement with environments and tools. This fundamental capability marks the dawn of the Age of AI Agency, driven by a critical industry shift: the urgent need for AI systems that don't just think, but work. While current AI excels at reasoning and generating responses, industries demand autonomous agents that can execute tasks, operate tools, and drive real-world outcomes. As agentic intelligence becomes the defining characteristic separating cognitive systems from productive workers, efficiently cultivating machine autonomy becomes paramount. Current approaches assume that more data yields better agency, following traditional scaling laws from language modeling. We fundamentally challenge this paradigm. LIMI (Less Is More for Intelligent Agency) demonstrates that agency follows radically different development principles. Through strategic focus on collaborative software development and scientific research workflows, we show that sophisticated agentic intelligence can emerge from minimal but strategically curated demonstrations of autonomous behavior. Using only 78 carefully designed training samples, LIMI achieves 73.5% on comprehensive agency benchmarks, dramatically outperforming state-of-the-art models: Kimi-K2-Instruct (24.1%), DeepSeek-V3.1 (11.9%), Qwen3-235B-A22B-Instruct (27.5%), and GLM-4.5 (45.1%). Most strikingly, LIMI demonstrates 53.7% improvement over models trained on 10,000 samples-achieving superior agentic intelligence with 128 times fewer samples. Our findings establish the Agency Efficiency Principle: machine autonomy emerges not from data abundance but from strategic curation of high-quality agentic demonstrations.

AI AgencyAutonomous AgentsMachine AutonomyData CurationAgentic IntelligenceAI AgentArtificial IntelligenceMachine Learning
02

Qwen3-Omni Technical Report

We present Qwen3-Omni, a single multimodal model that, for the first time, maintains state-of-the-art performance across text, image, audio, and video without any degradation relative to single-modal counterparts. Qwen3-Omni matches the performance of same-sized single-modal models within the Qwen series and excels particularly on audio tasks. Across 36 audio and audio-visual benchmarks, Qwen3-Omni achieves open-source SOTA on 32 benchmarks and overall SOTA on 22, outperforming strong closed-source models such as Gemini-2.5-Pro, Seed-ASR, and GPT-4o-Transcribe. Qwen3-Omni adopts a Thinker-Talker MoE architecture that unifies perception and generation across text, images, audio, and video, yielding fluent text and natural real-time speech. It supports text interaction in 119 languages, speech understanding in 19 languages, and speech generation in 10 languages. To reduce first-packet latency in streaming synthesis, Talker autoregressively predicts discrete speech codecs using a multi-codebook scheme. Leveraging the representational capacity of these codebooks, we replace computationally intensive block-wise diffusion with a lightweight causal ConvNet, enabling streaming from the first codec frame. In cold-start settings, Qwen3-Omni achieves a theoretical end-to-end first-packet latency of 234 ms. To further strengthen multimodal reasoning, we introduce a Thinking model that explicitly reasons over inputs from any modality. Since the research community currently lacks a general-purpose audio captioning model, we fine-tuned Qwen3-Omni-30B-A3B to obtain Qwen3-Omni-30B-A3B-Captioner, which produces detailed, low-hallucination captions for arbitrary audio inputs. Qwen3-Omni-30B-A3B, Qwen3-Omni-30B-A3B-Thinking, and Qwen3-Omni-30B-A3B-Captioner are publicly released under the Apache 2.0 license.

Multimodal ModelQwen3-OmniMoE ArchitectureMultimodal ReasoningSpeech SynthesisMultimodalNatural Language ProcessingLarge Language Model
03

EpiCache: Episodic KV Cache Management for Long Conversational Question Answering

Recent advances in large language models (LLMs) have extended context lengths, enabling assistants to sustain long histories for coherent, personalized responses. This ability, however, hinges on Key-Value (KV) caching, whose memory grows linearly with dialogue length and quickly dominates under strict resource constraints. An active line of research for reducing this overhead is KV cache compression, which seeks to limit cache size while preserving accuracy. Yet existing methods face two major limitations: (i) evicting entries after full-context prefill causes unbounded peak memory, and (ii) query-dependent eviction narrows the cache to a single query, leading to degraded accuracy in multi-turn conversations. We introduce EpiCache, a training-free KV cache management framework for long conversational question answering (LongConvQA) under fixed memory budgets. EpiCache bounds cache growth through block-wise prefill and preserves topic-relevant context via episodic KV compression, which clusters conversation history into coherent episodes and applies episode-specific KV cache eviction. We further design an adaptive layer-wise budget allocation strategy that measures each layer's sensitivity to eviction and distributes the memory budget across layers accordingly. Across three LongConvQA benchmarks, EpiCache improves accuracy by up to 40% over recent baselines, sustains near-full KV accuracy under 4-6x compression, and reduces latency and memory by up to 2.4x and 3.5x, thereby enabling efficient multi-turn interaction under strict resource constraints.

KV Cache ManagementLarge Language ModelsConversational Question AnsweringCache CompressionMemory EfficiencyLarge Language ModelNatural Language ProcessingDeep Learning
04

Strategic Dishonesty Can Undermine AI Safety Evaluations of Frontier LLM

Large language model (LLM) developers aim for their models to be honest, helpful, and harmless. However, when faced with malicious requests, models are trained to refuse, sacrificing helpfulness. We show that frontier LLMs can develop a preference for dishonesty as a new strategy, even when other options are available. Affected models respond to harmful requests with outputs that sound harmful but are subtly incorrect or otherwise harmless in practice. This behavior emerges with hard-to-predict variations even within models from the same model family. We find no apparent cause for the propensity to deceive, but we show that more capable models are better at executing this strategy. Strategic dishonesty already has a practical impact on safety evaluations, as we show that dishonest responses fool all output-based monitors used to detect jailbreaks that we test, rendering benchmark scores unreliable. Further, strategic dishonesty can act like a honeypot against malicious users, which noticeably obfuscates prior jailbreak attacks. While output monitors fail, we show that linear probes on internal activations can be used to reliably detect strategic dishonesty. We validate probes on datasets with verifiable outcomes and by using their features as steering vectors. Overall, we consider strategic dishonesty as a concrete example of a broader concern that alignment of LLMs is hard to control, especially when helpfulness and harmlessness conflict.

AI SafetyLarge Language ModelsStrategic DishonestySafety EvaluationsModel AlignmentArtificial IntelligenceLarge Language ModelDeep Learning
05

VideoFrom3D: 3D Scene Video Generation via Complementary Image and Video Diffusion Models

In this paper, we propose VideoFrom3D, a novel framework for synthesizing high-quality 3D scene videos from coarse geometry, a camera trajectory, and a reference image. Our approach streamlines the 3D graphic design workflow, enabling flexible design exploration and rapid production of deliverables. A straightforward approach to synthesizing a video from coarse geometry might condition a video diffusion model on geometric structure. However, existing video diffusion models struggle to generate high-fidelity results for complex scenes due to the difficulty of jointly modeling visual quality, motion, and temporal consistency. To address this, we propose a generative framework that leverages the complementary strengths of image and video diffusion models. Specifically, our framework consists of a Sparse Anchor-view Generation (SAG) and a Geometry-guided Generative Inbetweening (GGI) module. The SAG module generates high-quality, cross-view consistent anchor views using an image diffusion model, aided by Sparse Appearance-guided Sampling. Building on these anchor views, GGI module faithfully interpolates intermediate frames using a video diffusion model, enhanced by flow-based camera control and structural guidance. Notably, both modules operate without any paired dataset of 3D scene models and natural images, which is extremely difficult to obtain. Comprehensive experiments show that our method produces high-quality, style-consistent scene videos under diverse and challenging scenarios, outperforming simple and extended baselines.

3D Scene Video GenerationDiffusion ModelsGenerative FrameworkImage Diffusion ModelsVideo Diffusion ModelsGenerative AIDeep LearningComputer Vision
06

Reasoning Core: A Scalable RL Environment for LLM Symbolic Reasoning

We introduce Reasoning Core, a new scalable environment for Reinforcement Learning with Verifiable Rewards (RLVR), designed to advance foundational symbolic reasoning in Large Language Models (LLMs). Unlike existing benchmarks that focus on games or isolated puzzles, Reasoning Core procedurally generates problems across core formal domains, including PDDL planning, first-order logic, context-free grammar parsing, causal reasoning, and system equation solving. The environment is built on key design principles of high-generality problem distributions, verification via external tools, and continuous difficulty control, which together provide a virtually infinite supply of novel training instances. Initial zero-shot evaluations with frontier LLMs confirm the difficulty of Reasoning Core's tasks, positioning it as a promising resource to improve the reasoning capabilities of future models.

Reasoning CoreReinforcement LearningLarge Language ModelsSymbolic ReasoningVerifiable RewardsArtificial IntelligenceNatural Language ProcessingLarge Language Model