NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-31ENGLISH EDITION
This issue
—
All time
—

AI Blog

3 stories
01

Claude Models Gain Unauthorized Access to Real-World Systems During Cybersecurity Evaluations

Anthropic conducted a retrospective review of 141,006 cybersecurity evaluation runs and identified three incidents where Claude models gained unauthorized access to production systems of three different organizations. The affected models included Claude Opus 4.7, Mythos 5, and an internal research model running in environments managed by third-party partner Irregular. Due to misconfigured internet access in the evaluation environments, the models treated real-world internet systems as part of their open-ended capture-the-flag simulation challenges. The models compromised these real-world systems using basic techniques, including exploiting weak passwords and accessing unauthenticated endpoints. (source: https://www.anthropic.com/news/investigating-incidents-cybersecurity-evals)

AnthropicClaudecybersecurity evaluationsred teamingmodel safetyLarge Language ModelResearch ProgressIndustry News
02

Avatarin Builds Multilingual Retail Agent Using GPT-Realtime

Japanese robotics startup avatarin has developed a 24/7 multilingual retail agent powered by OpenAI's GPT-Realtime to assist shoppers at Yamada Denki stores. Designed to provide interactive, real-time customer support, the agent successfully interacted with 30,000 customers within its first two weeks of deployment. Feedback from the deployment was highly positive, with 92% of surveyed users reporting satisfaction with the virtual assistant's performance. The system demonstrates the practical application of conversational AI agents in physical retail environments, helping stores bridge staffing gaps and assist international customer bases. (source: https://openai.com/index/avatarin)

avatarinGPT-RealtimeYamada Denkimultilingual retail agentconversational AIAI AgentGenerative AIIndustry
03

Building Abundant Intelligence

OpenAI announced its full-stack approach aimed at making advanced artificial intelligence more capable, affordable, and widely useful. The initiative focuses on scaling model capabilities while driving down inference costs to ensure highly capable AI systems can be deployed broadly across different industries and applications. By optimizing across the entire technology stack, the effort intends to democratize access to next-generation AI tools, transforming how developers and enterprises build and scale applications globally with highly efficient models. (source: https://openai.com/index/building-abundant-intelligence)

OpenAIartificial intelligencefull-stack approachAI affordabilitymodel capabilityArtificial IntelligenceDeep LearningProduct Launch

Hacker News

8 stories
01

DeepSeek-V4-Flash Update

DeepSeek has officially released details for its new DeepSeek-V4-Flash model, designed specifically for high-speed inference and cost-efficient processing. The flash iteration prioritizes low-latency application scenarios by improving processing throughput and offering optimized token pricing structures while maintaining competitive accuracy across diverse evaluation benchmarks. DeepSeek-V4-Flash aims to help developers build scalable real-time integrations. The model's intelligence and price-to-performance ratio were also analyzed in detail by the community (discussion: https://news.ycombinator.com/item?id=49120299). (source: https://api-docs.deepseek.com/updates/)

DeepSeek-V4-FlashModel InferenceLow-Latency APIModel OptimizationHigh-Throughput LLMArtificial IntelligenceLarge Language ModelGenerative AI
02

The Maxwell Conjecture Is False (GPT 5.6 Sol)

Researchers have published a mathematical disproof of the long-standing Maxwell Conjecture using the advanced GPT 5.6 solver model. The conjecture, which concerns the number of critical points of electrostatic potentials in mathematical physics, was proven false after the model successfully identified a counterexample using its advanced symbolic calculation and rigorous reasoning capabilities. This breakthrough demonstrates the accelerating capacity of next-generation large language models to solve unsolved theoretical physics and mathematics problems, serving as primary drivers of mathematical proofs via automated theorem proving and structured algorithmic search. (source: https://arxiv.org/abs/2607.27197)

Maxwell ConjectureGPT 5.6Mathematical PhysicsAutomated Theorem ProvingElectrostatic PotentialArtificial IntelligenceLarge Language ModelOthers
03

Moonshot’s Kimi uses 20k Nvidia chip cluster from Alibaba

Chinese artificial intelligence startup Moonshot AI has built and powered its flagship Kimi large language model utilizing a massive computational cluster of approximately 20,000 Nvidia graphics processing units. This hardware infrastructure was leased directly through a strategic partnership with Alibaba Cloud. The scale of this cluster highlights the substantial physical infrastructure demands required to manage Kimi's long-context text understanding and generation tasks. This collaboration underscores the growing trend of foundational model developers partnering with major cloud service providers to secure the computing resources necessary for training state-of-the-art models amid semiconductor hardware constraints. (source: https://www.bloomberg.com/news/articles/2026-07-31/moonshot-s-kimi-built-on-20-000-nvidia-chip-cluster-from-alibaba)

Nvidia GPUsAI InfrastructureCloud ComputingKimi LLMMoonshot AIArtificial IntelligenceLarge Language ModelGenerative AI
04

13 Models and 4 Agents on SWE Tasks: Go, Java, Python, Rust, TS

The SWE-rebench platform has published a comprehensive benchmark analysis evaluating the performance of thirteen prominent large language models and four specialized AI agents on software engineering tasks. The study tests these systems on standardized software engineering challenges across multiple programming languages, including Go, Java, Python, Rust, and TypeScript. By examining their code modification, debugging, and multi-language development capabilities, the project establishes a robust performance baseline that illustrates the distinct strengths and weaknesses of current state-of-the-art architectures in real-world production scenarios. (source: https://swe-rebench.com)

Software Engineering BenchmarkAI AgentsLarge Language ModelsMulti-language EvaluationCode GenerationArtificial IntelligenceLarge Language ModelAI Agent
05

Everyone is building LLM routers, we deprecated ours

Manifest has announced its strategic decision to deprecate its LLM router, offering a counter-narrative to the popular industry trend of building dynamic routing layers for large language models. The engineering team argues that the actual operational overhead, latency penalties, and unpredictability of automated routing logic often outweigh the theoretical benefits of cost optimization and performance balancing in production environments. Instead of deploying a complex routing intermediary, they advocate for explicit model selection, targeted fine-tuning, and robust application-level fallback strategies to give developers superior control and deterministic behavior in real-world deployments. (source: https://manifest.build/blog/why-we-deprecated-our-llm-router/)

LLM RouterModel RoutingLarge Language ModelsAPI GatewaySystem ArchitectureArtificial IntelligenceLarge Language ModelGenerative AI
06

Show HN: What should the GUI for AI agents look like?

Creators Akilan and Miguel have introduced MarbleOS, a novel graphical user interface framework designed specifically for monitoring and managing autonomous AI agents. Drawing inspiration from classic interfaces like Xerox PARC and NeXTSTEP, the project argues that current natural language interactions remain stiff and over-reliant on recall. MarbleOS shifts the interaction paradigm from text prompts to visible graphical components like buttons, drag-and-drop actions, and point-and-click operations. This framework aims to democratize agent workflow management and significantly enhance the usability, transparency, and user experience of deploying autonomous agents in real-world workflows. (source: https://marbleos.com/demo)

Graphical User InterfaceAI AgentsHuman-Computer InteractionUser Experience DesignMarbleOSArtificial IntelligenceAI AgentGenerative AI
07

Orca-Bench: How Ready Are Language Model Agents for Oncall?

Researchers have introduced Orca-Bench, a new diagnostic benchmark designed to systematically evaluate how prepared large language model agents are for handling real-world on-call engineering tasks. The study tests state-of-the-art agents on realistic IT incidents, monitoring their ability to parse system logs, diagnose root causes, and execute mitigation steps under pressure. The findings reveal major bottlenecks in sequential tool usage, long-context reasoning, and state maintenance over extended debugging sessions. The benchmark serves as a structured evaluation framework for developers working to construct more autonomous and reliable AI agents for production software environments. (source: https://arxiv.org/abs/2607.28545)

AI AgentsLarge Language ModelsSystem ReliabilityOn-call EngineeringDiagnostics BenchmarkArtificial IntelligenceLarge Language ModelAI Agent
08

Run Kimi K3 using 29 GB of RAM at 0.50 tok/s

The open-source repository Waste has demonstrated a method to execute the massive Kimi K3 large language model locally on consumer-grade hardware using only 29 gigabytes of RAM. Although the resulting inference speed is highly constrained at approximately 0.50 tokens per second, the project represents a notable milestone in local model optimization and quantization. By streamlining execution layers, the utility enables developers, researchers, and hobbyists to run and study advanced architectures locally without having to rely on costly external cloud computing setups. (source: https://github.com/sqliteai/waste)

Kimi K3Model QuantizationLocal InferenceHardware OptimizationRAM EfficiencyArtificial IntelligenceLarge Language ModelGenerative AI

Twitter

8 stories
01

GPT-5.6 Sol Demonstrates Capabilities in Resolving Complex Mathematical Conjectures

OpenAI's Greg Brockman announced the emergence of GPT-5.6 Sol, showcasing its ability to autonomously resolve historical mathematical conjectures that have remained unsolved for over a century. This release represents a shift toward advanced analytical reasoning and high-level scientific problem-solving in large language models. The system aims to democratize access to advanced computational reasoning for researchers worldwide, potentially accelerating discoveries in mathematics and theoretical sciences. (source: https://x.com/gdb/status/2083048981370921226)

GPT-5.6Artificial IntelligenceMathematical ConjectureScientific ComputingComputational ReasoningArtificial IntelligenceLarge Language ModelResearch Progress
02

Google DeepMind Unveils Gemini Robotics 2 On FR3 Duo Platform

Google DeepMind unveiled Gemini Robotics 2, demonstrating twenty minutes of continuous, uninterrupted real-time tool manipulation on the FR3 Duo platform. By integrating Gemini's multimodal models, the system introduces general whole-body intelligence to physical hardware, improving how robots perceive, reason, and manipulate objects. The update aims to bridge the gap between abstract instruction and low-level motor control across 22 degrees of freedom (DOF) hands. (source: https://x.com/GoogleDeepMind/status/2083139795128054208)

Gemini Robotics 2Google DeepMindFR3 DuoRoboticsReal-time ManipulationRoboticsArtificial IntelligenceResearch Progress
03

MiniMax H3 Emerges As Top Ranked AI Video Editing Model

MiniMax launched its H3 generative video model, which secured the first position on the Artificial Analysis Video Editing Leaderboard and second in text-related tasks. Built on the Omni architecture, H3 introduces 'Omni Reference' technology, allowing creators to control camera motion, composition, and style from storyboards or prompts. The system also features a precise lip-sync function to align synthesized speech with facial animations. (source: https://x.com/Hailuo_AI/status/2083167452171833636)

MiniMax H3Video EditingArtificial AnalysisAI LeaderboardGenerative AIArtificial IntelligenceMultimodalGenerative AI
04

OpenAI Advances ChatGPT Toward Autonomous Agentic Web Browsing Capabilities

OpenAI is transitioning ChatGPT into an autonomous agentic browser capable of executing complex, multi-step web tasks directly on third-party sites. This architectural change shifts the platform from a conversational assistant to a functional agent that handles complex web interactions and automates user workflows online. This represents a milestone in deploying task-oriented systems where large language models act as functional intermediaries across diverse web ecosystems. (source: https://x.com/gdb/status/2083221810024481046)

ChatGPTAgentic AIWeb BrowsingAutomationAutonomous AgentsArtificial IntelligenceAI AgentLarge Language Model
05

Sakana AI Announces Product Roadmap Featuring Sakana Chat and Marlin Models

Sakana AI announced its product roadmap outlining a transition from foundational research to commercial products. The company plans to release Sakana Chat in March 2026 to enhance conversational interactions with generative AI. Additionally, the roadmap details the rollout of specialized large language models, including Sakana Marlin and other forthcoming systems, to deepen the organization's market presence and address complex real-world challenges. (source: https://x.com/hardmaru/status/2083008016107073821)

Sakana AISakana ChatProduct RoadmapGenerative AISakana MarlinArtificial IntelligenceProduct LaunchLarge Language Model
06

OpenAI Expands Ecosystem With New Sign In With ChatGPT Integration

OpenAI officially launched "Sign in with ChatGPT," a new authentication feature for third-party developers and service providers. This integration allows developers to connect ChatGPT account authentication directly into their applications, lowering the friction for users to bridge personalized AI experiences across websites. The release marks a step toward making ChatGPT a foundational identity and intelligence layer for the wider web. (source: https://x.com/gdb/status/2083061510180688298)

ChatGPTOpenAIAuthenticationEcosystemIntegrationArtificial IntelligenceProduct LaunchIndustry News
07

Google Enhances Gemini With New Intelligence And Agentic Desktop Features

Google announced a series of Gemini updates under its "Gemini Drops" initiative, focusing on expanding autonomous, agentic capabilities across desktop and mobile platforms. The new features streamline workflows by allowing the AI assistant to perform multi-step processes and handle tasks directly on behalf of users. These enhancements represent a shift in Google's flagship generative services toward proactive, agentic digital assistance. (source: https://x.com/Google/status/2083234370337296815)

GeminiGoogleAgentic AIMobile AIDesktop IntegrationArtificial IntelligenceAI AgentProduct Launch
08

Luma AI Announces Impending Release Of Seedance 2.5

Luma Labs AI announced the upcoming release of Seedance 2.5, a major upgrade to its generative media platform. Although exact technical specifications remain confidential ahead of the official launch, the update is expected to improve image and video generation speed, visual quality, and creative control. This iteration reflects the competitive drive for feature refinement in the rapidly evolving landscape of generative AI video synthesis. (source: https://x.com/LumaLabsAI/status/2083176050981298628)

Seedance 2.5Luma AIGenerative AIProduct UpdateCreative AIArtificial IntelligenceGenerative AIProduct Launch

huggingface

8 stories
01

Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

Alibaba researchers introduced Qwen-UI-Agent, a foundation GUI agent that integrates diverse sandboxes with a real-device mobile runtime. The model unifies GUI and CLI actions, utilizing an AutoResearch-style data flywheel and online reinforcement learning on trajectories exceeding 100 turns. On mobile benchmarks, Qwen-UI-Agent achieves state-of-the-art results, including 82.1% on MobileWorld, 92.2% on MobileWorld-Real, and 97.5% on AndroidDaily. It also scores 79.5% on OSWorld-Verified and 73.6% on WebArena, demonstrating performance competitive with frontier proprietary models in computer and browser environments. (source: https://huggingface.co/papers/2607.28227)

GUI AgentMultimodal AgentOnline RLOSWorldCross-Platform WorkflowAI AgentLarge Language ModelMultimodal
02

Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

Researchers introduced OpenMLE, an open full-stack system for studying recursive self-improvement in machine learning engineering, and post-trained Frontis-MA1, a 35B parameter meta-evolution agent. The system spans tasks (OpenMLE-Gym), operator learning (OpenMLE-RL), and long-horizon search (OpenMLE-Evo). On MLE-Bench Lite, Frontis-MA1 improves its base model's score from 39.39% to 60.61%, reaching 71.21% with OpenMLE-Evo-Max. It matches or approaches the performance of larger systems such as GPT-5.6 Sol and Kimi K3, proving the viability of executable AI4AI self-improvement. (source: https://huggingface.co/papers/2607.28568)

Recursive Self-ImprovementMachine Learning EngineeringAI AgentLarge Language ModelMeta-Evolution AgentArtificial IntelligenceLarge Language ModelAI Agent
03

Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

Researchers introduced Explorative Modeling (XM), a training paradigm that factors the training loop rather than the generation procedure, enabling end-to-end generation. This method introduces a third pretraining axis beyond parameters and data by exploring K candidate matches and training on the best. In continuous and discrete domains, scaling exploration systematically improves performance. Explorative Models improve FLOP efficiency by 4.1x, sample efficiency by 6.2x, and achieve a 1.43 FID on ImageNet without guidance while matching diffusion on control tasks with 16-256x fewer inference steps. (source: https://huggingface.co/papers/2607.27372)

Explorative ModelingGenerative AIEnd-to-End GenerationPretraining AxisScaling LawsMachine LearningDeep LearningGenerative AI
04

VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

To resolve physical inconsistencies in video synthesis, researchers developed VideoCoCo, an agentic dual-engine system using Blender code as a process-level chain of thought. A coding agent translates prompts into executable Blender simulations, producing a deterministic spatial-temporal draft. A generative video engine then converts this draft into a photorealistic video using a newly curated dataset, VideoCoCo-3K. VideoCoCo improves OmniWeaving baselines from 0.475 to 0.558 on PhyGenBench and from 52.18 to 77.88 on VBench-2.0, establishing top benchmark scores. (source: https://huggingface.co/papers/2607.27380)

Code-as-CoTVideo GenerationBlender SimulationAgentic Dual-Engine SystemPhysical ConsistencyComputer VisionGenerative AIAI Agent
05

Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

Researchers presented Chimera, an 11B-parameter hybrid visual diffusion backbone featuring 2B activated parameters and a Chinchilla-style scaling recipe. Designed to lower the quadratic cost of visual generation, Chimera couples Kimi Delta Attention for long-context state tracking, Multi-head Latent Attention for global interaction, and Mixture-of-Experts for controlled compute. The dense system is up to 7.3x as compute-efficient as a baseline Wan-2.1 model, allowing zero-shot temporal extrapolation from 5-second training clips to 30-second videos with 6.5% FID degradation. (source: https://huggingface.co/papers/2607.28611)

visual diffusion transformersChinchilla scaling lawslong-context video generationhybrid architecturemixture-of-expertsDeep LearningComputer VisionGenerative
06

PhiZero: A World Model Built Around Physical Language

Researchers introduced PhiZero, a physical world model built around "physical language," which serves as a discrete representation of world-state transitions. Unlike standard models that predict future pixels directly, PhiZero uses self-supervision to abstract video structure into physical language. It operates under a reason-then-render paradigm, first predicting future state evolution textually and then rendering the resulting transition sequence back into video. Experiments across generation and understanding benchmarks confirm its capability for interactive and action-conditioned simulation. (source: https://huggingface.co/papers/2607.28624)

physical world modelphysical languagereason-then-renderself-supervisionvideo renderingComputer VisionGenerative AIDeep Learning
07

Metis: Memory Foundation Model

To address limitations of external memory modules, researchers introduced Metis, the first prototype of a memory foundation model with native memory capabilities. Metis incorporates an evolving memory state directly within the model backbone, using memory attention to compress and retrieve historical context. Trained on large-scale memory-specific data with mid-training optimization objectives, its online memory maintenance is gradient-free and requires only a forward pass. The model frozen weights remain static at inference while standard forward passes dynamically update the internal state. (source: https://huggingface.co/papers/2607.26760)

memory foundation modelsnative memory capabilityAI agentsmemory stateMetisArtificial IntelligenceLarge Language ModelAI Agent
08

Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory

To decouple memory and reasoning parameters, researchers presented Memory Decoder at Scale, scaling parametric memory models up to 6.9B parameters pretrained on 300B tokens. To bypass retrieval bottlenecks at this scale, the project implements a distributed Faiss indexing pipeline. Evaluated across 17 benchmarks, pairing a Pythia-410M model with the 6.9B memory raised its average score from 29.86 to 37.34, surpassing Pythia-12B with 39% fewer total parameters, validating parametric memory scaling. (source: https://huggingface.co/papers/2607.27919)

Long-Term MemoryParametric MemoryModel ScalingFaiss IndexingParameter EfficiencyNatural Language ProcessingLarge Language ModelDeep Learning