NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-08-01ENGLISH EDITION
This issue
—
All time
—

AI Blog

1 story
01

Disrupts Malicious Cambodia-Based Criminal Scam Operation

OpenAI has disrupted a Cambodia-based criminal scam operation that utilized ChatGPT to support fraudulent schemes. The threat actors used the artificial intelligence platform to facilitate investment fraud, romance scams, gambling schemes, and impersonation activities. By detecting and terminating these accounts, OpenAI blocked the group's ability to generate persuasive content for their operations. This action represents a major effort to prevent the abuse of LLMs by malicious actors and to share threat intelligence with the broader cybersecurity community. (source: https://openai.com/index/disrupting-malicious-uses-of-ai-criminal-scam-operation)

ChatGPTOpenAICybersecurityMalicious UseThreat IntelligenceArtificial IntelligenceIndustry NewsLarge Language Model

Hacker News

5 stories
01

Show HN: Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO)

A developer has released nano-llm-posttraining, an open-source framework designed for running LLM post-training experiments on consumer-grade hardware with as little as 8GB of VRAM. The repository provides minimal, simplified implementations of key alignment and reinforcement learning techniques, including Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and Group Relative Policy Optimization (GRPO). By utilizing optimization strategies like Low-Rank Adaptation (LoRA) and 4-bit quantization, the library allows researchers and developers to run advanced fine-tuning and reasoning-oriented post-training processes locally on a tight budget. (source: https://github.com/pochenai/nano-llm-posttraining)

Supervised Fine-TuningDirect Preference OptimizationGroup Relative Policy OptimizationLoRAQuantizationLarge Language ModelMachine LearningGenerative AI
02

Flint: A Visualization Language for the AI Era

Microsoft has introduced Flint, a new declarative visualization language designed specifically to streamline data visualization tasks in the artificial intelligence era. Flint is engineered to be natively compatible with the reasoning systems of large language models, allowing AI agents and human developers to collaborate on generating and modifying interactive charts and dashboards. This new language reduces the cognitive load of data visualization by translating natural language prompts into technical execution. The project aims to improve productivity for workflows in data science and business intelligence. (source: https://microsoft.github.io/flint-chart/)

FlintData VisualizationVisualization GrammarAI CollaborationDeclarative LanguageArtificial IntelligenceGenerative AIOthers
03

Explorative modeling: Train on the best of K guesses

An article introduces explorative modeling, a novel training methodology designed to improve predictive accuracy and model adaptability by training on the best of K guesses. Instead of traditional single-path optimization, this approach generates multiple hypotheses during training and selectively updates model weights based on the top-performing candidate. This methodology helps the model navigate non-convex loss landscapes and avoid local minima. The technique is particularly beneficial for reinforcement learning, generative modeling, and complex sequence prediction tasks where uncertainty is high. (source: https://alexiglad.github.io/blog/2026/explorative_modeling/)

Explorative ModelingLoss OptimizationHypothesis GenerationModel GeneralizationReinforcement LearningArtificial IntelligenceMachine LearningDeep Learning
04

GenAI Added to Google Earth

Google has integrated generative artificial intelligence capabilities directly into Google Earth to alter how users interact with geospatial data and satellite imagery. This integration uses advanced generative models to simulate, visualize, and predict complex environmental changes, historical geographic transformations, and urban developments. The technology enables researchers and urban planners to generate realistic predictive visual models of landscapes under various scenarios, such as rising sea levels. This update represents a major step in bringing multimodal generative computer vision to spatial computing frameworks. (source: https://www.theatlantic.com/technology/2026/07/google-earth-ai-images/688145/)

Google EarthGenerative AIGeospatial AIComputer VisionSpatial ComputingArtificial IntelligenceGenerative AIComputer Vision
05

AI #179 Part 2: Hearing the Fire Alarm

An analysis published by Zvi Mowshowitz examines critical perspectives on the progression of artificial intelligence, centering on safety protocols, alignment challenges, and existential risks. The piece details the ongoing debate within the technology sector regarding the speed of AI development, regulatory action, and the implications of achieving artificial general intelligence (AGI) without control frameworks. It focuses on the tension between commercial deployment and safety, calling for a coordinated global approach to monitor capabilities and mitigate systemic risks. (source: https://thezvi.substack.com/p/ai-179-part-2-hearing-the-fire-alarm)

AI SafetyAI AlignmentArtificial General IntelligenceAI GovernanceExistential RiskArtificial IntelligenceGenerative AIOthers

Twitter

5 stories
01

Astra Model Demonstrates Breakthrough Results In Mathematical Research

Sam Altman announced that an internal version of OpenAI's upcoming Astra model family has successfully resolved ten major open problems in mathematics, quantum complexity, and theoretical computer science. These achievements include proving the existence of nonsofic groups at an operational cost of approximately $2000 under current Sol API pricing. The milestone represents a major step forward for LLM logical deduction, showing potential transitions from simple linguistic generation to advanced scientific problem-solving. Critics like Gary Marcus have urged skepticism, questioning the formal validity of these theoretical proofs and calling for peer-reviewed academic validation. (source: https://x.com/sama/status/2083552732225417342)

AstraNonsofic GroupsMathematical ResearchModel DevelopmentAI-Driven DiscoveryArtificial IntelligenceResearch ProgressLarge Language Model
02

MiniMax Releases Advanced Video Generation Model For Creative MV Production

MiniMax has officially launched its new video generation model, H3, featured on the Hailuo AI platform. The model supports specialized text-to-video synthesis and creative music video production tools, featuring Chinese text prompts, single reference images, and stylistic character references. Demonstrations show the H3 architecture executing advanced motion graphics, fluid interface transitions, combat visual generation, lip-sync capabilities, and atmospheric cinematic sequences such as a stylized jazz-noir 'Midnight Line' video. The releases aim to lower the technical barrier for complex video workflows and high-fidelity audiovisual storytelling. (source: https://x.com/Hailuo_AI/status/2083496264742928448)

MiniMaxH3Generative AIVideo GenerationCreative ToolsGenerative AIMultimodalProduct Launch
03

ChatGPT Work Introduces Integrated Cloud Browser For AI Interaction

ChatGPT Work has unveiled an integrated cloud browser feature designed to enhance transparency and human control over autonomous AI agents. The tool allows users to monitor ongoing multi-step agent workflows in real-time as actions unfold on the screen. Additionally, the software introduces an intervention mechanism that lets developers or business users step in and directly adjust or influence the live application processes when required. This feature represents a strategic focus on making autonomous workflows more manageable, robust, and reliable for everyday practical operations. (source: https://x.com/gdb/status/2083641828415602989)

ChatGPT WorkCloud BrowserAI AgentsHuman-in-the-loopWorkflow MonitoringArtificial IntelligenceAI AgentProduct Launch
04

Gemini Robotics 2 Demonstrates Precision And 3D Reorientation Tasks

Google has showcased the advanced physical capabilities of Gemini Robotics 2, its latest robotic intelligence foundation model. The demonstration highlights significant technical progress in physical automation, specifically high-precision insertion maneuvers and complex 3D object reorientation. By utilizing the multimodal reasoning architectures of Gemini models, the physical robotic hardware can perceive changes, plan trajectories, and execute precise motor movements within varied environments. This integration represents a major step in bridging the gap between abstract AI reasoning and physical dexterity in industrial and domestic automation settings. (source: https://x.com/ZoubinGhahrama1/status/2083451198405784044)

Gemini Robotics 2RoboticsPrecision ManipulationMultimodal AIAutomationArtificial IntelligenceRoboticsResearch Progress
05

Advancing Open Source AI Model Safety and Responsible Development

John Schulman discussed his organization's strategic commitment to open-weight models and the active development of supporting fine-tuning tooling. The announcement outlines a technical roadmap aimed at balancing open-source scientific contributions with the mitigation of AI misuse risks. Schulman highlighted the critical research needed to navigate model distribution safely and issued an open invitation for security and alignment researchers to join the team. This initiative emphasizes a collaborative approach to solving model evaluation and alignment challenges while maintaining public access to powerful neural network weights. (source: https://x.com/johnschulman2/status/2083615606784573669)

Open WeightsAI SafetyFine-tuningModel ReleaseResponsible AIArtificial IntelligenceLarge Language ModelOpen Source