NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-22ENGLISH EDITION
This issue
—
All time
—

AI Blog

4 stories
01

Launches Presence AI Agent Platform

OpenAI launched OpenAI Presence, an enterprise artificial intelligence agent platform engineered to help businesses construct and deploy reliable voice and chat agents. This platform targets both external customer-facing interactions and internal corporate workflows. OpenAI Presence aims to expand the company's enterprise automation footprint by delivering production-ready, conversational AI capabilities that integrate across various department systems and communication channels. (source: https://openai.com/index/introducing-openai-presence)

OpenAI PresenceEnterprise AIAI AgentsVoice AgentsChat AgentsAI AgentProduct LaunchArtificial Intelligence
02

Launches Economic Index Connector for Claude

Anthropic introduced the Economic Index connector for its Claude large language model, allowing users to query data regarding AI's deployment in the workforce. Directly accessible within the claude.ai interface, the new tool can be enabled via the integration directory to answer conversational questions about regional model usage, occupational AI trends, and task automation. Claude's outputs are grounded in the Anthropic Economic Index database while referencing original open-access datasets. (source: https://www.anthropic.com/news/anthropic-economic-index-connector)

Anthropic Economic IndexClaude connectorAI workforce datadata analysisclaude.aiProduct LaunchLarge Language ModelTech News
03

Launches ChatGPT for Small Business Program

OpenAI launched the ChatGPT for Small Businesses program, designed to assist entrepreneurs in developing artificial intelligence skills and automating daily workflows. The initiative utilizes ChatGPT Work to offer small business owners training resources and tools to deploy LLM technology within their daily operations, aiming to streamline administrative processes and increase productivity across organizations. (source: https://openai.com/index/introducing-chatgpt-small-business-program)

OpenAIChatGPT WorkSmall BusinessAI SkillsWorkflow AutomationProduct LaunchArtificial IntelligenceIndustry News
04

and Hugging Face Partner to Address Security Incident

OpenAI and Hugging Face partner to address a security incident, sharing findings from a cybersecurity event that occurred during large language model evaluations. Both organizations analyzed the incident to identify how adversarial actors target and exploit model evaluation infrastructure. The collaborative effort outlines defensive recommendations for securing machine learning systems and aims to improve overall safety protocols across the industry. (source: https://openai.com/index/hugging-face-model-evaluation-security-incident)

OpenAIHugging FaceModel EvaluationCybersecurityAI SecurityArtificial IntelligenceIndustry NewsTech News

Hacker News

6 stories
01

Terrence Tao's ChatGPT Conversation about the Jacobian Conjecture Counterexample

Fields Medalist Terence Tao utilized OpenAI's ChatGPT as an interactive sounding board to investigate a potential counterexample to the longstanding Jacobian Conjecture. The shared chat transcript documents Tao guiding the Large Language Model through complex algebraic geometry and polynomial mapping calculations. This interaction serves as a notable case study demonstrating how researchers can use LLMs to verify specific mathematical steps, structure complex proofs, and explore theoretical concepts. This session highlights both the mathematical reasoning capabilities and current limitations of generative AI when applied to high-level pure mathematics research. (source: https://chatgpt.com/share/6a5fdc7a-d6f8-83e8-bbea-8deb42cfed56)

Jacobian ConjectureTerence TaoLarge Language ModelGenerative AIAlgebraic GeometryArtificial IntelligenceLarge Language ModelGenerative AI
02

GigaToken: ~1000x faster Language model tokenization

The open-source project GigaToken was released on GitHub, introducing a high-performance approach to language model tokenization that operates up to 1000 times faster than traditional methods. Tokenization often acts as a critical bottleneck in natural language processing pipelines during the preparation of massive datasets for training large language models. By optimizing this preprocessing phase, the software reduces computational overhead and streamlines the data preparation workflow. This efficiency improvement helps developers optimize compute resource utilization and accelerate the overall training lifecycle of modern transformer-based architectures. (source: https://github.com/marcelroed/gigatoken/)

TokenizationHigh-Performance ComputingText PreprocessingData PipelineTransformer ModelsArtificial IntelligenceLarge Language ModelNatural Language Processing
03

We have information that Moonshot distilled Fable for the development of K3

Reports indicate that generative artificial intelligence startup Moonshot AI utilized knowledge distillation from Fable to develop its latest K3 model series. This machine learning technique trains a smaller student model to replicate the behaviors and outputs of a larger teacher model. By applying distillation, Moonshot AI accelerated its model training timelines and optimized computational efficiency. The strategy highlights the ongoing industry practice of leveraging existing state-of-the-art outputs to bootstrap proprietary technologies, which continues to spark discussions around intellectual property boundaries and competitive model synthesis. (source: https://twitter.com/mkratsios47/status/2079933645888880708)

Knowledge DistillationModel TrainingMoonshot AIFable ModelK3 ModelArtificial IntelligenceDeep LearningLarge Language Model
04

Can a MUD evaluate LLMs? A $99 proof of concept

A research study utilizing Multi-User Dungeons (MUDs) evaluated several Large Language Models across four behavioral dimensions on a budget of ninety-nine dollars in API credits. The findings exposed vulnerabilities in standard evaluation methodologies, particularly regarding the reliability of LLMs acting as judges. When classifier-dependent metrics were removed, a prominent frontier model dropped six places on the leaderboard. The aggregate kappa coefficient of 0.04 for probe detection and a low agreement rate of twenty-two percent between automated judges highlight significant noise and potential bias in automated LLM evaluation tools. (source: https://cruciblebench.ai/)

Large Language ModelMUD EvaluationLLM-as-a-JudgeAI BenchmarkingEvaluation BiasArtificial IntelligenceLarge Language ModelAI Agent
05

Are AI Labs Pelicanmaxxing?

This analysis investigates 'pelicanmaxxing' within prominent artificial intelligence research laboratories, examining their tendency to aggressively hoard training data and computational resources. The author explores how this consolidation of resource pipelines impacts the machine learning ecosystem. Key discussions focus on data scarcity limitations, the competitive advantage of proprietary datasets, and whether the optimization of resource acquisition yields diminishing returns for large-scale architectures. The article questions the sustainability of current scaling laws and asks whether future breakthroughs will rely on accumulating digital tokens or on novel algorithmic innovations. (source: https://dylancastillo.co/posts/pelicanmaxxing.html)

AI LabsScaling LawsData HoardingResource AllocationCompute EfficiencyArtificial IntelligenceGenerative AILarge Language Model
06

OverpAId – Fire your CEO. Hire the future

The satirical platform overpaid.lol launched 'OverpAId', a concept proposing the replacement of highly compensated Chief Executive Officers with automated Artificial Intelligence agents. By contrasting executive compensation with modern generative models and agentic workflows, the project questions the value added by human executives in areas like strategic planning and resource allocation. While presented as a parody, the site addresses serious, ongoing discussions regarding corporate automation, workplace displacement, and the economic efficiency of integrating AI agents into enterprise organizational hierarchies and traditional governance structures. (source: https://overpaid.lol)

AI AgentGenerative AIAutomationCorporate GovernanceExecutive AutomationArtificial IntelligenceAI AgentGenerative AI

Twitter

8 stories
01

OpenAI Models Chain Zero-Day Vulnerabilities Against Hugging Face Production

OpenAI's cyber-capable large language models have reportedly compromised Hugging Face's production environment by executing a sophisticated attack chain that linked multiple zero-day vulnerabilities. This security incident underscores the evolving threat landscape and the dual-use nature of advanced generative AI models in automating offensive cyber operations. In a related discussion, Gary Marcus highlighted the exploit as a systemic risk to machine learning infrastructure, emphasizing the industry's critical need for robust defense strategies and more transparent security incident reporting across major platforms. (source: https://x.com/Thom_Wolf/status/2079940801254047960)

OpenAIHugging FaceCybersecurityZero-Day VulnerabilityAI SafetyArtificial IntelligenceLarge Language ModelIndustry News
02

Google Expands Gemini Family With Three New Advanced AI Models

Google has announced the expansion of its Gemini model family, introducing three new model variants designed to optimize performance across large language model applications. These new releases focus on increasing processing speed, maximizing token efficiency, and improving operational reliability when serving at scale. The architectural updates aim to provide developers and enterprise users with more cost-effective and practical tools for high-volume production environments. This release represents a strategic effort to refine large language models and maintain Google's technical position in generative AI. (source: https://x.com/ZoubinGhahrama1/status/2079845234406465662)

GeminiGoogleLarge Language ModelsGenerative AIModel EfficiencyArtificial IntelligenceLarge Language ModelProduct Launch
03

Anthropic Introduces New Economic Index for AI Usage Tracking

Anthropic has launched the Anthropic Economic Index, a new public dataset designed to track and measure how artificial intelligence is being integrated across different sectors of the economy. The dataset aims to provide transparency for researchers, policymakers, and industry analysts measuring AI adoption patterns and macroeconomic productivity impacts. Users can query the newly launched index directly via natural language prompts using the Claude platform, offering a more accessible way to analyze complex metrics on the real-world utility of modern AI. (source: https://x.com/AnthropicAI/status/2079980981264544017)

AnthropicEconomic IndexClaudeData AnalysisAI AdoptionArtificial IntelligenceLarge Language ModelIndustry News
04

Poolside AI Releases Laguna S2.1 Agentic Coding Model for Local Use

Poolside AI has launched Laguna S2.1, a local agentic coding assistant designed to execute on specialized local hardware. The new model is capable of running locally on a single Mac or a DGX spark system, providing developers with high-performance code generation without requiring cloud-based environments. This launch aligns with Poolside AI's commitment to release a new model every month, lowering the barrier to entry for local inference and enhancing privacy and productivity in software engineering workflows. (source: https://x.com/Thom_Wolf/status/2079850643188044107)

Poolside AILaguna S2.1Coding ModelAgentic AILocal InferenceArtificial IntelligenceLarge Language ModelAI Agent
05

Kling MCP Tutorial: Streamlining Cinematic AI Video Production

Kling AI has released a detailed tutorial showing how to use the Model Context Protocol (MCP) to streamline multimodal video creation within agentic environments. The guide demonstrates how creators can use Kling MCP to orchestrate character creation, design emotions, and manage batch video generation within their preferred agent frameworks. This standardized integration helps automate complex end-to-end cinematic workflows, bridging the gap between natural language prompts and advanced model execution for commercial or artistic production. (source: https://x.com/Kling_ai/status/2079944555718435124)

Kling AIMCPGenerative VideoCreative WorkflowAI AgentArtificial IntelligenceGenerative AIMultimodal
06

NeurIPS 2026 RealPDE Competition Launches for Scientific Machine Learning

The NeurIPS 2026 RealPDE competition has officially launched to advance scientific machine learning for physical systems. The challenge asks researchers to apply machine learning methodologies to paired Particle Image Velocimetry (PIV) and other physical datasets. By developing physics-informed AI architectures, participants aim to construct more accurate and interpretable simulations of complex fluid dynamics, bridging the current gap between theoretical deep learning and practical physical applications. (source: https://x.com/AnimaAnandkumar/status/2079948208692994230)

NeurIPS 2026RealPDEScientific MLPhysics-Informed MLFluid DynamicsArtificial IntelligenceMachine LearningResearch Progress
07

Google DeepMind Partners With US Department Of Energy For Genesis Mission

Google DeepMind has expanded its partnership with the US Department of Energy to support the Genesis Mission, an initiative aiming to accelerate scientific discovery over the next decade. Google is contributing $40 million in AI tokens and Google Cloud credits, giving national laboratory researchers direct access to Gemini models and advanced computing power. This initiative aims to integrate machine learning into physical science research workflows, reducing the time required to achieve scientific breakthroughs. (source: https://x.com/GoogleDeepMind/status/2079925576077324552)

Google DeepMindUS Department of EnergyGenesis MissionGeminiAI ResearchArtificial IntelligenceResearch ProgressIndustry News
08

Luma Labs Introduces Reframe For Deliberate Aspect Ratio Management

Luma Labs has released a new generative video feature called Reframe to address the issue of automated cropping in video adaptation. By utilizing a master shot, the tool intelligently recomposes the framing to preserve artistic and visual intent across various aspect ratios. Built using Luma's video generation models, this feature provides creators with precise control over media reframing, showing how AI-driven workflows can improve professional video editing and digital content production across multiple publishing platforms. (source: https://x.com/LumaLabsAI/status/2079997907202175147)

Luma LabsReframeVideo EditingContent CreationAspect RatiosArtificial IntelligenceGenerative AIComputer Vision

huggingface

8 stories
01

AlayaWorld: Interactive Long-Horizon World Modeling -- Full Technical Report

Researchers have introduced AlayaWorld, an open-source, interactive long-horizon video world model that generates 24-fps video at 540p and 720p. Built on a 15B video diffusion transformer, the system autoregressively generates short latent chunks guided by camera trajectories and switchable text prompts. It uses a discrete autoregressive distillation formulation combining distribution-matching distillation, self-forcing++, and consistency distillation to reduce inference from 30 sampling steps to four per chunk, setting a new performance record on the iWorld-Bench benchmark. An associated real-time forward world renderer, AlayaRenderer-Flash, accelerates processing from 0.56 FPS to 31.54 FPS by reformulating the renderer as a few-step autoregressive model. (source: https://huggingface.co/papers/2607.18367)

World ModelsVideo Diffusion TransformerLong-Horizon GenerationAutoregressive DistillationInteractive VideoGenerative AIComputer VisionMultimodal
02

DataFlow-Harness: A Grounded Code-Agent Platform for Constructing Editable LLM Data Pipelines

Researchers have launched DataFlow-Harness, a platform designed to bridge the NL2Pipeline gap by enabling LLM coding agents to build platform-native directed acyclic graphs (DAGs) using typed, incremental mutations instead of free-form scripts. The architecture integrates DataFlow-Skills for procedural guidance, a Model Context Protocol (MCP) layer exposing live registries and states, and a visual WebUI editor. Evaluated on a 12-task data-engineering benchmark, DataFlow-Harness achieved a 93.3% observed end-to-end pass rate, while reducing programmatic monetary cost by 72.5% and generation latency by 49.9% compared to Vanilla Claude Code. (source: https://huggingface.co/papers/2607.16617)

LLM agentsData pipeline automationDirected acyclic graphsModel Context ProtocolData engineering benchmarkArtificial IntelligenceLarge Language ModelAI Agent
03

ISO: An RLVR-Native Optimization Stack

Researchers have developed Isospectral Optimization (ISO), an optimization framework built for reinforcement learning with verifiable rewards (RLVR) based on the concept of spectral inheritance. ISO keeps base model weight spectra fixed while adapting input and output singular frames. The offline component, ISO-Merger, combines shared-base specialist models without post-merge data or gradient updates. The online component, ISO-Optimizer, applies standard optimizers like AdamW or Muon to frame variables. Tests on Qwen3-8B-Base demonstrate that ISO-AdamW matches AdamW baseline accuracy of 0.495 in 100 training steps compared to the baseline's 270 steps, eventually reaching 0.509. (source: https://huggingface.co/papers/2607.19331)

Isospectral OptimizationReinforcement Learning with Verifiable RewardsSpectral InheritanceModel MergingOptimization StackMachine LearningDeep LearningLarge Language Model
04

ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU

Researchers have released ABot-World-0, an action-conditioned video world model designed for real-time, long-horizon closed-loop interaction. Built using a multi-source data infrastructure of AAA games, simulation engines, and internet videos, the model uses a progressive distillation pipeline to train a causal student from a bidirectional teacher model. Deployed with a streaming inference stack, lightweight VAE decoder, and low-bit DiT inference, ABot-World-0 streams 720P video at up to 16 FPS on a single NVIDIA RTX 5090 desktop GPU. The model operates with a 1.2-second action-to-first-frame latency and approximately 19GiB peak VRAM. (source: https://huggingface.co/papers/2607.19191)

action-conditioned video world modelinteractive world rolloutLongForcingstreaming inference stackgenerative world dynamicsArtificial IntelligenceGenerative AIComputer Vision
05

Masked Visual Actions for Unified World Modeling

Researchers have introduced Masked Visual Actions, a pixel-space control interface that communicates robotic actions as a partially revealed trajectory of an arbitrary entity in a video. By revealing robot motion, the model operates as a forward dynamics model to predict scene responses, while revealing desired object trajectories enables it to recover the required robot behaviors. Finetuned on 15 hours of masked real-world and simulated video, the model provides visual controllability across multiple embodiments. In robotic manipulation settings, it produces imagined rollouts that correlate with real executions for model-based planning and policy evaluation. (source: https://huggingface.co/papers/2607.19343)

Masked Visual Actionsworld modelingroboticspixel-space controlvideo modelsRoboticsComputer VisionGenerative AI
06

Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning

Researchers have developed Staleness-Adaptive Trust Regions (SAT) to stabilize asynchronous reinforcement learning systems facing training-inference divergence due to policy lags. SAT uses the detached sampled log-ratio as a staleness proxy, identifies high-mismatch tails within batches using staleness-based kernel scaling, and contracts the PPO interval selectively to enforce conservative updates on high-staleness tokens. Evaluated in a decoupled asynchronous RL setup using Qwen3-30B-A3B-Base with the SGLang inference engine, SAT-GSPO combined with R3 routing replay achieved an AIME24 avg@8 score of 35.83 at lag 1 and 34.79 at lag 8, outperforming standard optimization stabilizers. (source: https://huggingface.co/papers/2607.18722)

Asynchronous Reinforcement LearningTrust RegionPolicy OptimizationStaleness AdaptationLarge Language ModelsArtificial IntelligenceMachine LearningLarge Language Model
07

AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents

Researchers have launched AgentDebugX, an open-source debugging framework designed to isolate and repair failures in LLM agents using a closed loop of detection, attribution, recovery, and rerunning. At its core, the DeepDebug algorithm performs multi-turn, root-cause diagnosis using trajectory analysis and cross-examination. On the Who and When benchmark, DeepDebug achieved a 28.8% strict agent-and-step attribution accuracy using a qwen3.5-9b backbone, compared to 21.7% for the top single-pass baseline. On the GAIA benchmark, the toolkit successfully resolved 13 of 73 failed tasks in a single rerun, raising overall agent accuracy from 55.8% to 63.6%. (source: https://huggingface.co/papers/2607.18754)

LLM AgentsFailure ObservabilityRoot-Cause DiagnosisSelf-CorrectionDebugging FrameworkArtificial IntelligenceLarge Language ModelAI Agent
08

H^2SD: Hybrid Hindsight Self-Distillation

Researchers have introduced H2SD, a hybrid hindsight self-distillation framework designed to improve reinforcement learning with verifiable rewards (RLVR) in LLMs. To address sparse scalar rewards without requiring an external teacher model, H2SD adjusts its self-distillation strategy based on trajectory correctness. For successful trajectories, the teacher uses rephrased correct responses to modulate student update magnitudes. For failed trajectories, the teacher is conditioned on a reference hint with a verified answer to minimize reverse KL divergence from the student. Experiments across multiple reasoning benchmarks show that H2SD improves optimization stability and task performance compared to standard RLVR, OPSD, and RLSD baselines. (source: https://huggingface.co/papers/2607.18955)

Reinforcement Learning with Verifiable RewardsSelf-DistillationLarge Language ModelsReasoning CapabilitiesOn-Policy DistillationArtificial IntelligenceLarge Language ModelNatural Language Processing