NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-13ENGLISH EDITION
This issue
—
All time
—

AI Blog

1 story
01

Getting Started with ChatGPT

OpenAI published an educational guide outlining fundamental workflows and introductory steps for navigating ChatGPT. The document guides new users through initiating their first conversations and highlights practical applications for daily productivity, writing, and brainstorming. It explains how to integrate the conversational interface into routine workflows to assist with problem-solving and creative tasks. (source: https://openai.com/academy/getting-started)

ChatGPTOpenAIAI GuideProductivity ToolBrainstorming AINatural Language ProcessingGenerative AIProduct Launch

Hacker News

8 stories
01

Grok CLI uploaded the whole home directory to GCS

A major privacy and security incident has occurred where the Grok Command Line Interface (CLI) reportedly uploaded a user's entire local home directory to Google Cloud Storage without explicit consent. The unexpected behavior highlights severe vulnerabilities in automated data collection pipelines within developer-focused artificial intelligence integrations. Security analysts warn that such expansive directory scanning risking exposure of sensitive local credentials, personal files, and private SSH keys to cloud-connected platforms. The developer community has raised major concerns regarding data handling and permission boundaries in AI-assisted coding tools. (source: https://twitter.com/a_green_being/status/2076598897779020159)

Grok CLIData PrivacyGoogle Cloud StorageData LeakSecurity VulnerabilityArtificial IntelligenceAI AgentOthers
02

Zig Creator Calls Spade a Spade, Anthropic Blows Smoke

Zed text editor creator Nathan Sobo has challenged Anthropic regarding the realistic capabilities of current AI coding assistants versus their marketing narratives. The dispute centers on whether current large language models are truly autonomous software engineering agents capable of complex reasoning, or if these claims overpromise and underdeliver on developer productivity. The critique focuses on how major AI labs exaggerate model capabilities to secure venture capital. The piece calls for more honest benchmarks to assess how models integrate into actual workflows. This topic was discussed widely on Hacker News with over 600 comments. (source: https://raymyers.org/post/zed-creator-calls-spade-a-spade/)

AI Coding AssistantsSoftware EngineeringAnthropicZed EditorLarge Language ModelArtificial IntelligenceLarge Language ModelGenerative AI
03

Show HN: Clawk – Give coding agents a disposable Linux VM, not your laptop

Clawk has released an open-source security tool designed to sandbox autonomous coding agents during execution. Instead of permitting AI agents to run commands directly on a local developer machine, the tool provisions a disposable, isolated Linux virtual machine. This isolated environment prevents potential damage, unauthorized file access, or malicious code execution on the host operating system. Clawk bridges local agent execution and secure virtualization, allowing developers to leverage AI-driven coding assistants without risking local credentials. (source: https://github.com/clawkwork/clawk)

AI SandboxingVirtual MachineSecure ExecutionCoding AgentsLinux VMArtificial IntelligenceAI AgentOthers
04

Show HN: Nobie – an Excel-compatible runtime for agents and humans

Nobie has launched an Excel-compatible runtime designed to facilitate collaboration between artificial intelligence agents and human operators. Utilizing a spreadsheet-based data model, the environment allows LLM-based agents to perform complex computations and structured workflows using standard Excel logic. This recognizable interface enables human collaborators to easily audit, intervene in, and guide agentic processes. The platform serves as a critical coordination tool for building collaborative human-in-the-loop systems to scale enterprise automation safely. (source: https://nobie.com)

Excel RuntimeAI AgentsHuman-in-the-loopSpreadsheet EngineAgentic WorkflowsArtificial IntelligenceAI AgentOthers
05

Show HN: BillAI Bass, an AI-Powered Big Mouth Billy Bass Using Strands Agents

Developer Morgan Willis Cloud has released BillAI Bass, an open-source hardware project that turns a Big Mouth Billy Bass animatronic toy into an interactive AI assistant. By leveraging Strands Agents, the system decouples the mechanical control of the hardware and synchronizes it with real-time AI-generated audio responses. The architecture processes natural language inputs, generates context-aware voice responses, and translates audio frequencies into motor control commands for mouth and tail movements. (source: https://github.com/morganwilliscloud/billai-bass)

AI AgentHardware HackingStrands AgentsGenerative AIVoice SynthesisArtificial IntelligenceAI AgentGenerative AI
06

New Flagship Grok Voices

xAI has released its new flagship Grok Voices, introducing advanced speech synthesis capabilities to the Grok AI platform. The updated system utilizes speech generation technologies to provide enhanced emotional range, pacing, and realistic human-computer voice interaction. This update marks a progression in xAI's multimodal integration, allowing users to experience real-time verbal interactions that simulate natural human speech inflections for intuitive, hands-free utilization of their large language model. (source: https://x.ai/news/new-flagship-voices)

Grok VoicesSpeech SynthesisMultimodal AIConversational AIText-to-SpeechArtificial IntelligenceNatural Language ProcessingGenerative AI
07

The real prices of frontier models. Tokens * Price, right?

A technical analysis by Playcode challenges the simplistic assumption that calculating the cost of running frontier models is a straightforward formula of multiplying tokens by unit price. The report highlights that production workloads incur hidden operational costs including prompt engineering iterations, context window management, and latency trade-offs. The analysis details how architectural decisions, retrieval-augmented generation design, and cache utilization heavily influence final operational expenses, requiring developers to adopt a holistic systems-level approach to optimize LLM budgets. (source: https://playcode.io/blog/real-price-of-frontier-models)

Frontier ModelsAPI PricingToken CostLLM InfrastructureCost OptimizationArtificial IntelligenceLarge Language ModelGenerative AI
08

Latent Space as a New Medium

Kevin Kelly has published an article conceptualizing latent space as a revolutionary digital medium enabled by generative AI technologies. By treating the mathematical multidimensional space of machine learning models as an explorable territory, the author frames latent space as a collaborative canvas rather than a simple database. Instead of merely producing static outputs, users can navigate and steer within this dimensional matrix to synthesize ideas, marking a shift in how we approach design, art, and digital representation. (source: https://kevinkelly.substack.com/p/latent-space-as-a-new-medium)

Latent SpaceGenerative AIHuman-AI CollaborationRepresentation SpaceCreative AIArtificial IntelligenceGenerative AIDeep Learning

Twitter

4 stories
01

Morpheus Benchmark Introduces Non-Episodic Continual Learning Environments

Francois Chollet has highlighted the release of the Morpheus benchmark, a new framework designed for continual reinforcement learning research. Standard reinforcement learning benchmarks typically rely on episodic, stationary environments that fail to capture the complexity of real-world deployments. Morpheus addresses these limitations by offering persistent simulation environments that do not reset, requiring agents to navigate learning objectives that continuously evolve over time. This approach aims to bridge the gap between lab environments and dynamic real-world settings, providing a more robust methodology for evaluating agent adaptability. (source: https://x.com/fchollet/status/2076719958189613307)

MorpheusReinforcement LearningContinual LearningSimulationBenchmarksMachine LearningArtificial IntelligenceResearch Progress
02

Kling AI Powers Viral Portugal World Cup Campaign Film With Cinematic Quality

Kling AI has announced that its generative video technology was utilized by production company 78 Films to craft Portugal's World Cup campaign film, which has accumulated nearly 100 million views. The commercial project integrated a hybrid production workflow, leveraging Kling AI's generative video model to generate high-fidelity, complex cinematic scenes that would otherwise be difficult to realize with conventional production methods alone. This deployment serves as a major real-world case study for the practical application of generative AI video tools within professional, large-scale media campaigns. (source: https://x.com/Kling_ai/status/2076683070292500710)

Kling AIGenerative AIVideo ProductionCinematic Quality78 FilmsArtificial IntelligenceGenerative AIIndustry News
03

Kling AI Introduces Dynamic Dual Persona Feature For Generative Video Models

Kling AI has unveiled a new product showcase demonstrating character transition capabilities within its generative video model. The demonstration transforms a domestic cat into a professional boxer, highlighting the platform's ability to maintain structural integrity, motion coherence, and high-fidelity texture mapping during dramatic physical and environmental transitions. This update aims to provide content creators with sophisticated tools for multimodal character animation and narrative consistency in video generation workflows. (source: https://x.com/Kling_ai/status/2076637773214486759)

Kling AIGenerative VideoCharacter AnimationMultimodal AIVideo GenerationArtificial IntelligenceGenerative AIComputer Vision
04

OpenAI Team Celebrates Milestone Performance Improvements In ChatGPT

Greg Brockman has announced a milestone regarding significant capability and performance improvements for ChatGPT. This internal development represents a period of success for the OpenAI team, emphasizing updates to architectural and system features that have begun rolling out to users. The milestone reflects OpenAI's continuous strategy of iterative model performance tuning, aiming to provide more reliable, robust, and sophisticated conversational interfaces for complex workflows and creative endeavors. (source: https://x.com/gdb/status/2076518764112445861)

ChatGPTOpenAIGenerative AIAI DevelopmentInnovationArtificial IntelligenceLarge Language ModelGenerative AI

huggingface

8 stories
01

Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

Researchers have introduced Long-Horizon-Terminal-Bench, a new benchmark comprising 46 long-horizon tasks across nine categories to evaluate AI agents over extended workflows. Unlike traditional terminal benchmarks that rely solely on final outcomes, this framework decomposes tasks into fine-grained graded subtasks to provide dense intermediate reward signals. Evaluating 15 frontier models revealed significant room for improvement, with agents consuming an average of 9.9 million tokens per task over 85.3 minutes of execution. Even the highest-performing model achieved only a 15.2% pass rate at a 0.95 partial-reward threshold and 10.9% at a perfect reward threshold. (source: https://huggingface.co/papers/2607.08964)

Long-Horizon PlanningTerminal BenchmarkDense RewardsAI AgentsPerformance EvaluationArtificial IntelligenceAI AgentLarge Language Model
02

A Sovereign, Open-Source Foundation Model for German and English

Developers have released Soofi S 30B-A3B, a sovereign, open-source Mixture-of-Experts (MoE) hybrid Mamba Transformer foundation model for German and English. Built on the German Industrial AI Cloud in Munich, the model activates only 3 billion of its 30 billion parameters per token, which maintains a near-constant inference cache as context grows. Pretrained on 27 trillion tokens with up-weighted German data, Soofi S outperforms other European sovereign baselines and matches dense 14B to 27B models on aggregate benchmarks. The weights, intermediate checkpoints, and training code are released under permissive, open-access terms. (source: https://huggingface.co/papers/2607.09424)

Mixture-of-ExpertsMamba TransformerFoundation ModelSovereign AIOpen-Source LLMArtificial IntelligenceLarge Language ModelGenerative AI
03

Video Generation Models are General-Purpose Vision Learners

Researchers have introduced GenCeption, a framework demonstrating that large-scale text-to-video generative diffusion models function as strong pretraining backbones for general computer vision tasks. Guided by text instructions, GenCeption leverages these generative spatiotemporal priors to perform feed-forward perception tasks such as depth estimation, camera pose estimation, and 3D keypoint prediction. The model matches or outperforms specialized visual architectures while requiring 7 to 500 times less training data than leading baselines. Additionally, the system exhibits zero-shot generalization from synthetic human training videos to real-world footage and out-of-distribution physical objects. (source: https://huggingface.co/papers/2607.09024)

GenCeptionvideo generationfoundation modelscomputer visiongenerative diffusionComputer VisionGenerative AIDeep Learning
04

Self-Guided Test-Time Training for Long-Context LLMs

Researchers have proposed Self-Guided Test-Time Training (S-TTT), a method designed to improve long-context utilization in large language models without incurring prohibitive computational costs. While standard test-time training is sensitive to noisy, irrelevant context spans, S-TTT instructs the model to identify relevant evidence spans before adaptation, restricting the language-modeling objective to those targeted areas. When evaluated on the LongBench-v2 and LongBench-Pro reasoning benchmarks, S-TTT improved the performance of Qwen3-4B-Thinking-2507 and Llama-3.1-8B-Instruct, yielding up to a 15% relative improvement in accuracy. (source: https://huggingface.co/papers/2607.09415)

Test-Time TrainingLong-Context LLMsSelf-Guided AdaptationContext UtilizationParameter AdaptationNatural Language ProcessingLarge Language ModelGenerative AI
05

Scalable Visual Pretraining for Language Intelligence

Researchers have presented a systematic study demonstrating that visual pretraining on raw documents is a scalable alternative to text-only pretraining for language models. Rather than converting visually rich elements like figures, mathematical layouts, and page structures into plain text, this approach leverages unsupervised visual learning directly on document images without text extraction. Across multiple backbones and benchmarks, models pretrained directly on these visual document representations consistently outperformed their text-only counterparts, offering a more complete pathway to visual and language intelligence. (source: https://huggingface.co/papers/2607.09657)

visual pretrainingfoundation modelslanguage intelligenceunsupervised learningmultimodal document understandingMultimodalLarge Language ModelDeep
06

KronQ: LLM Quantization via Kronecker-Factored Hessian

Researchers have introduced KronQ, a post-training quantization (PTQ) framework that incorporates gradient covariance into the quantization pipeline using a Kronecker-factored Hessian approximation. Unlike traditional methods like GPTQ that treat output channels equally, KronQ applies bidirectional incoherence processing to reduce weight magnitude variance across both input and output dimensions, alongside a sensitivity metric for inter-layer mixed-precision allocation. On 2-bit weight-only quantization of LLaMA-3-70B, where GPTQ diverges, KronQ successfully compresses the model to achieve a perplexity of 7.93 on WikiText-2. (source: https://huggingface.co/papers/2607.07964)

Post-training quantizationKronecker-factored HessianGradient covarianceBidirectional incoherenceMixed-precision allocationMachine LearningDeep LearningLarge Language Model
07

Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning

Researchers have analyzed the "Knowing-Using Gap" in large language models, where models successfully memorize fine-tuned facts but fail to apply them in downstream reasoning tasks. Using a novel activation intervention technique called self-patching, the authors tracked the spatial dynamics of new facts inside LLMs, finding that memorized representations often fail to route to computation-effective layers. To address this knowledge-circuit misalignment, they developed a simple heuristic routing strategy that recovers 58% to 75% of the oracle generalization headroom across multiple domains. (source: https://huggingface.co/papers/2607.08393)

Knowing-Using Gapknowledge-circuit misalignmentself-patchingmemorizationgeneralizationNatural Language ProcessingLarge Language ModelDeep Learning
08

From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models

Researchers have developed ReChannel, a method that adapts frozen text-to-image diffusion models for pixel-space dense prediction without using a target-side RGB decoder. By mapping the Diffusion Transformer (DiT) token lattice directly to local pixel-space patches via a minimal 33K-parameter linear head, ReChannel bypasses generative output interfaces to directly predict task-native fields like depth and segmentations. Evaluated using FLUX-Klein, the system achieved state-of-the-art results in trimap-free matting, KITTI depth, and referring segmentation, performing 2.48 times faster than edit-based latent-decoding equivalents. (source: https://huggingface.co/papers/2607.06553)

Dense PredictionText-to-Image ModelsDiffusion TransformersPixel-Space ReadoutGenerative PretrainingComputer VisionDeep LearningGenerative AI