NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-17ENGLISH EDITION
This issue
—
All time
—

AI Blog

2 stories
01

Near-Autonomous AI Chemist Improves Key Drug-Making Reaction

OpenAI and Molecule.one have successfully demonstrated a near-autonomous AI chemist system powered by the GPT-5.4 model to optimize a challenging chemical reaction critical to drug-making. By combining advanced artificial intelligence with automated chemical synthesis platforms, the team developed an autonomous workflow capable of identifying superior reaction conditions. This integration of large language models and physical laboratories represents a milestone in utilizing agentic AI to accelerate real-world scientific discovery and medicine development. (source: https://openai.com/index/ai-chemist-improves-reaction)

GPT-5.4Molecule.oneAI ChemistMedicinal ChemistryReaction OptimizationAI AgentResearch ProgressArtificial Intelligence
02

Opens Seoul Office and Strengthens Partnerships in South Korea

Anthropic has opened a new office in Seoul to establish a permanent presence and integrate its Claude model across South Korea's AI ecosystem. As part of this expansion, NAVER and Nexon have deployed Claude Code to optimize their software engineering pipelines. Major enterprises including LG CNS, Hanwha Solutions, and Samsung SDS are also integrating Claude to manage agentic workflows. Additionally, Anthropic is partnering with the National AI Research Lab to provide model access to several major Korean universities. (source: https://www.anthropic.com/news/seoul-office-partnerships-korean-ai-ecosystem)

Anthropic Seoul OfficeClaude CodeKorean AI EcosystemEnterprise AI IntegrationNational AI Research LabProduct LaunchIndustry NewsLarge Language Model

Hacker News

8 stories
01

GLM-5.2 is the new leading open weights model on Artificial Analysis

The Artificial Analysis platform announced that GLM-5.2 has secured the top position as the leading open-weights model on its Intelligence Index. Outperforming prior open-weights models, GLM-5.2 demonstrates significant improvements in model efficiency and natural language processing capabilities. In related benchmarking, the model's performance was measured across standard conversational, mathematical reasoning, and coding benchmarks to evaluate its latency, throughput, token generation speed, and cost-efficiency. This milestone signals a narrowing performance gap between proprietary systems and open-source alternatives for real-time applications. (source: https://artificialanalysis.ai/articles/glm-5-2-is-the-new-leading-open-weights-model-on-the-artificial-analysis-intelligence-index, additional benchmark details: https://artificialanalysis.ai/models/glm-5-2)

GLM-5.2Open Weights ModelArtificial AnalysisIntelligence IndexModel EvaluationArtificial IntelligenceLarge Language ModelGenerative AI
02

US holds off blacklisting DeepSeek, more than 100 firms deemed security risks

The United States government has decided to postpone the blacklisting of Chinese artificial intelligence firm DeepSeek, despite designating more than 100 other entities as national security risks. This decision highlights the complex geopolitical tensions surrounding international machine learning development, export controls, and open-source collaboration. By refraining from an immediate blacklist, the administration balances national security priorities against potential disruptions to the global developer ecosystems that rely on integrated generative AI technologies. Meanwhile, the inclusion of other firms on the security list signals continued federal vigilance regarding technology transfers. (source: https://www.reuters.com/world/china/us-holds-off-blacklisting-chinas-deepseek-more-than-100-firms-deemed-security-2026-06-17/)

DeepSeekArtificial IntelligenceNational SecurityExport ControlsGeopoliticsArtificial IntelligenceGenerative AIOthers
03

The hacker sent by Anthropic to calm the government's nerves about AI safety

Anthropic has deployed a high-profile hacker and security researcher to engage with policymakers and alleviate government concerns regarding artificial intelligence safety and security. This proactive initiative is designed to demonstrate Anthropic's commitment to safety engineering by directly illustrating the alignment techniques, red-teaming procedures, and vulnerability assessments built into its Claude series of large language models. The collaborative engagement aims to establish trust, bridge the gap between technical AI alignment and governmental policy expectations, and set clear safety benchmarks amidst growing global regulatory scrutiny of next-generation artificial intelligence systems. (source: https://www.wsj.com/tech/ai/anthropic-mythos-safety-nicholas-carlini-20bceaa3)

AI SafetyAnthropicClaudeRed TeamingModel AlignmentArtificial IntelligenceLarge Language ModelGenerative AI
04

AI chemist improves a challenging reaction in medicinal chemistry

OpenAI has highlighted how an artificial intelligence model successfully optimized a historically difficult chemical reaction crucial to medicinal chemistry. By utilizing advanced computational capabilities and data-driven prediction models, the AI-driven system analyzed complex reaction pathways to identify novel catalysts, reagents, and conditions that significantly enhance yield and selectivity. This milestone demonstrates the potential of machine learning to assist human scientists in navigating massive chemical spaces, accelerating pharmaceutical candidate synthesis and drug discovery, and establishing a new paradigm in computer-aided chemical research. (source: https://openai.com/index/ai-chemist-improves-reaction/)

AI ChemistMedicinal ChemistryChemical Reaction OptimizationDrug DiscoveryMachine Learning in ScienceArtificial IntelligenceMachine LearningOthers
05

The founder's playbook: Building an AI-native startup

Claude has published a strategic guide for entrepreneurs detailing how to build successful AI-native startups. The playbook explores fundamental design shifts, engineering workflows, and business models required when centering development around artificial intelligence instead of using it as an add-on. Key covered methodologies include leveraging foundational models, managing high inference costs, designing user experiences around probabilistic outputs, and transitioning from prompt engineering to robust agentic architectures. It emphasizes the importance of building data flywheels and executing proprietary fine-tuning to secure a sustainable competitive advantage. (source: https://claude.com/blog/the-founders-playbook)

AI-native StartupsModel EvaluationAgentic ArchitecturesData FlywheelsInference OptimizationArtificial IntelligenceGenerative AIAI Agent
06

Agentic coding deserves more than a chat box bolted onto VS Code

Developer Evan Klem launched Polypore, an open-source project challenging the standard paradigm of AI-assisted programming interfaces that rely on simple chat boxes in IDEs. The project proposes a deeply integrated, spatial user interface built from the ground up to support autonomous, collaborative AI agents. By moving past traditional sidebars, Polypore explores new UX models and systemic patterns that allow software developers to direct high-level architectural changes, supervise agent processes, and review outputs in a highly contextual environment. (source: https://github.com/evanklem/polypore)

Agentic CodingDeveloper ExperienceIDE IntegrationAI AgentsSoftware Engineering ToolsArtificial IntelligenceAI AgentGenerative AI
07

Launch HN: Adam (YC W25) – Open-Source AI CAD

Adam (YC W25) launched CADAM, an open-source text-to-CAD platform designed to generate mechanical designs through AI agents. Under the philosophy that computer-aided design should be generated as code, CADAM converts natural language prompts directly into code and renders them into parametric 3D models. The technical stack leverages a React frontend with TanStack Start and a Supabase backend for database management, authentication, and file storage. The platform acts as an open-source alternative to traditional CAD environments, bringing automation and software-development paradigms to mechanical prototyping. (source: https://github.com/Adam-CAD/CADAM)

Text-to-CADOpen-Source CADAI AgentsParametric 3D ModelingCAD as CodeArtificial IntelligenceGenerative AIAI Agent
08

TREX: An AI code reviewer that runs your code

Greptile introduced TREX, an AI-powered code review system designed to execute code in a secure sandbox environment rather than relying purely on static text analysis. By executing the code, TREX observes dynamic runtime behaviors, identifies execution errors, detects performance bottlenecks, and catches logical flaws that traditional static scanners miss. Combining large language models with dynamic runtime validation allows the system to generate highly accurate feedback directly on pull requests, streamlining debugging workflows and improving automated code quality verification. (source: https://www.greptile.com/blog/trex-code-execution)

AI Code ReviewDynamic AnalysisSandboxed ExecutionSoftware EngineeringAutomated DebuggingArtificial IntelligenceAI AgentGenerative AI

Twitter

8 stories
01

OpenAI Introduces GPT-Realtime-2 To Enhance Latency And Performance

OpenAI has officially unveiled GPT-Realtime-2, a major update to its real-time conversational model technology. This new iteration targets improved latency, higher response fidelity, and a more natural bidirectional communication flow. By updating the model's underlying architecture, the system enables human-level dialogue speeds and more responsive interactions within commercial multimodal systems. This launch represents OpenAI's ongoing strategic focus on low-latency generative voice and audio interactions to reduce conversation delays. (source: https://x.com/gdb/status/2067100786098831681)

GPT-Realtime-2OpenAILow LatencyConversational AIGenerative AIArtificial IntelligenceLarge Language ModelProduct Launch
02

Sakana AI Launches Sakana Marlin To Automate Professional Research Tasks

Sakana AI has officially launched Sakana Marlin, its first commercial product designed to automate professional research processes. Operating as an autonomous research assistant, the tool is engineered to gather, synthesize, and analyze complex information with minimal human intervention. The release marks a major commercial transition for the Japan-based research firm, turning its specialized generative AI capabilities into scalable enterprise workflows to accelerate professional knowledge-intensive decision-making. (source: https://x.com/hardmaru/status/2067241657339359254)

Sakana AISakana MarlinResearch AutomationCommercial AIProduct LaunchArtificial IntelligenceGenerative AIProduct Launch
03

Luma Labs Launches Ray 3.2 For Advanced Generative Video Capabilities

Luma Labs has officially released Ray 3.2, its latest generative video model, featuring improvements in video quality, processing speed, and temporal consistency. Additionally, Luma Labs has integrated this model onto the Runware platform, expanding developers' access to text-to-video and image-to-video tools. This deployment supports the generation of 5-second or 10-second video clips with advanced start and end frame controls. (source: https://x.com/LumaLabsAI/status/2067278775067578712)

Luma LabsRay 3.2Generative VideoVideo SynthesisAI MediaArtificial IntelligenceGenerative AIProduct Launch
04

Kling 3.0 Delivers Enhanced Natural Emotion And Expression Generation

Kling AI has officially released Kling 3.0, introducing significant updates to its generative video platform's capacity for rendering subtle human expressions and complex emotions. The update concentrates on high-fidelity performance, realistic character animations, and visual consistency. A subsequent video showcase by the team demonstrates the model's capacity for complex, high-stakes cinematic motion and fluid interactive scene composition. (source: https://x.com/Kling_ai/status/2067178544250397014)

Kling 3.0Generative VideoEmotion AISynthetic MediaArtificial IntelligenceArtificial IntelligenceGenerative AIVideo Understanding
05

Allen Institute for AI Introduces MolmoMotion for 3D Forecasting

The Allen Institute for AI has launched MolmoMotion, an advanced 3D motion forecasting model designed to predict object movement in spatial environments. By analyzing a small sequence of video frames and specified 3D points on an object, the system generates accurate trajectories and forecasts future motion. This research represents a milestone in spatial intelligence, improving how computer vision systems model and predict physical interactions within real-world environments. (source: https://x.com/Kyle_L_Wiggers/status/2067283197717602344)

MolmoMotion3D Motion ForecastingComputer VisionPredictive ModelingAllen Institute for AIArtificial IntelligenceComputer VisionResearch Progress
06

AutoScientist Automates AI Development Through Continuous Optimization

AutoScientist has introduced a new platform to automate manual artificial intelligence development workflows by continuously experimenting, adapting, and optimizing both training datasets and model configurations. By employing autonomous machine learning agents to manage the experiment lifecycle, the system aims to accelerate R&D speed and reduce necessary human supervision during model refinement. This continuous optimization framework presents a scalable approach for streamlining development infrastructure. (source: https://x.com/sarahookr/status/2067231931679609085)

AutoScientistAI DevelopmentAutomationMachine LearningData OptimizationArtificial IntelligenceMachine LearningAI Agent
07

Global Agent Collaboration Initiative Targets Gemma Model Optimization

A collaborative global initiative has launched to optimize the runtime efficiency and operational speed of the Gemma model. The project mobilized a network of over 100 distributed software agents working concurrently from various international locations to coordinate technical tasks. This initiative serves as a demonstration of collaborative multi-agent frameworks applied directly to large-scale open-source language model architecture refinement. (source: https://x.com/Thom_Wolf/status/2067219622382674327)

GemmaModel OptimizationAI AgentsCollaborationOpen SourceArtificial IntelligenceLarge Language ModelAI Agent
08

Proximal Policy Optimization Was Originally Rejected From NIPS 2017

John Schulman revealed that the foundational Proximal Policy Optimization (PPO) reinforcement learning paper was initially rejected from the 2017 Neural Information Processing Systems (NIPS) conference. Despite its rocky academic start, PPO has become an industry-standard policy gradient method widely used for aligning modern large language models. This disclosure highlights the subjective nature of peer reviews in machine learning research. (source: https://x.com/johnschulman2/status/2067263769110360522)

PPOReinforcement LearningNIPS 2017Research HistoryProximal Policy OptimizationArtificial IntelligenceMachine LearningResearch Progress

huggingface

8 stories
01

LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

Researchers have introduced LoopCoder-v2, a family of 7B parallel loop Transformer (PLT) code models trained from scratch on 18T tokens. The work addresses the latency and memory overheads of sequential looping by using cross-loop position offsets and shared-KV gated sliding-window attention. Evaluation shows a non-monotonic loop-count effect where a two-loop configuration yields significant improvements, raising SWE-bench Verified scores from 43.0 to 64.4 points and Multi-SWE from 14.0 to 31.0 points. Conversely, scaling beyond two loops degrades performance due to diminishing refinement gains and fixed positional mismatch costs. (source: https://huggingface.co/papers/2606.18023)

Parallel Loop TransformersTest-Time ComputationCode GenerationLoopCoder-v2Representation RefinementLarge Language ModelDeep LearningGenerative AI
02

Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients

Researchers proposed Zone of Proximal Policy Optimization (ZPPO), an alternative post-training framework that keeps the teacher's guidance within prompts rather than injecting it into policy gradients. This method aims to prevent student models from concentrating on sharp teacher modes or suffering from policy drift during reinforcement learning on difficult tasks. ZPPO utilizes binary and negative candidate-included prompts inside a prompt replay buffer to assist the student's learning progress. Testing on the Qwen3.5 model family across scales from 0.8B to 9B demonstrates that ZPPO consistently outperforms traditional off-policy or on-policy distillation and GRPO. (source: https://huggingface.co/papers/2606.18216)

knowledge distillationreinforcement learningzone of proximal developmentprompt reformulationvision-language modelsMachine LearningLarge Language ModelMultimodal
03

ProCUA-SFT Technical Report

Researchers released the ProCUA-SFT dataset, a collection of 3.1 million step-level supervised fine-tuning samples distilled from 93,000 synthetic desktop trajectories. Built using an automated pipeline with live desktop tasks, SpreadsheetBench, and Zenodo10K slides, the dataset addresses the negative transfer issues associated with human desktop trajectory datasets like AgentNet. Fine-tuning a UI-TARS 7B model on ProCUA-SFT increased its OSWorld benchmark success rate from 26.3% to 45.0%. A subset of this training data was also utilized to enhance the computer-use capabilities of the Nemotron 3 Nano Omni model. (source: https://huggingface.co/papers/2606.17321)

Computer-use AgentsSupervised Fine-TuningSynthetic TrajectoriesVision-Language ModelsOSWorld BenchmarkMultimodalAI AgentDeep Learning
04

OPD-Evolver: Cultivating Holistic Agent Evolver via On-Policy Distillation

Researchers introduced OPD-Evolver, a slow-fast co-evolution framework that designs self-evolving agents using on-policy self-distillation. The fast loop leverages a four-level memory hierarchy to use and maintain experience during test time, while the slow loop employs outcome-calibrated memory attribution and hindsight to distill these capabilities into the policy. Evaluated on multi-domain benchmarks, OPD-Evolver outperformed the ReasoningBank memory system by up to 11.5% and the training-based Skill0 baseline by approximately 5.8%. The 9B parameter version of the model demonstrated performance challenging significantly larger models such as Qwen3.5-397B-A17B. (source: https://huggingface.co/papers/2606.17628)

self-evolving agentson-policy distillationmemory hierarchyco-evolution frameworkmemory managementArtificial IntelligenceAI AgentLarge Language Model
05

Looped World Models

Researchers presented Looped World Models (LoopWM), a novel looped architecture designed to resolve the tension between deep computational needs and deployment efficiency in world simulation. By iteratively refining latent environment states through a single parameter-shared transformer block, LoopWM achieves up to a 100x improvement in parameter efficiency compared to conventional architectures. The framework scales depth dynamically to match prediction task difficulty, establishing iterative latent depth as an alternative scaling axis for simulating environment transitions. (source: https://huggingface.co/papers/2606.18208)

Looped World Modelsadaptive computationparameter efficiencyworld simulationtransformer blockMachine LearningDeep LearningGenerative AI
06

Learning from the Self-future: On-policy Self-distillation for dLLMs

Researchers introduced d-OPSD, the first on-policy self-distillation framework tailored specifically for diffusion large language models (dLLMs). Traditional on-policy distillation approaches depend on left-to-right token-level supervision, conflicting with the arbitrary-order generation of diffusion-based models. The d-OPSD method uses self-generated answers as suffix conditioning to learn from "self-future experience" and moves supervision to a step-level format aligned with iterative denoising. Tested on four reasoning benchmarks, d-OPSD outperformed RLVR and SFT baselines while using only 10% of the optimization steps needed by RLVR. (source: https://huggingface.co/papers/2606.18195)

On-policy self-distillationDiffusion LLMsSelf-future experienceDenoising processSample efficiencyArtificial IntelligenceNatural Language ProcessingLarge Language Model
07

ActWorld: From Explorable to Interactive World Model via Action-Aware Memory

Researchers proposed ActWorld, an interactive world model designed to enable mid-rollout object interaction inside a chunk-autoregressive video generation framework. ActWorld addresses existing action-forgetting and data bottlenecks by utilizing a new 100K human-object interaction video dataset annotated with chain-of-thought captions. It features a hierarchical action-aware memory system that routes history compression based on interaction importance, supported by a persistent memory bank that tracks event updates and object identities. Results demonstrate improved interaction fidelity and consistent viewpoint control. (source: https://huggingface.co/papers/2606.17730)

Interactive World ModelsAction-Aware MemoryHuman-Object InteractionVideo GenerationAutoregressive GenerationComputer VisionGenerative AIDeep Learning
08

A Gradient Perspective on RLVR Stability and Winner Advantage Policy Optimization

Researchers analyzed the gradient dynamics of Reinforcement Learning with Verifiable Rewards (RLVR), showing how GRPO-style optimization is susceptible to training collapse depending on advantage signs and token distributions. To improve stability, they introduced Winner Advantage Policy Optimization (WAPO), a clipped policy-gradient objective that performs updates exclusively on completions with positive advantages. Evaluated on mathematical reasoning and multi-hop question answering benchmarks, WAPO consistently matches or exceeds the performance of traditional baseline algorithms across multiple model families. (source: https://huggingface.co/papers/2606.16154)

Reinforcement LearningVerifiable RewardsGradient DynamicsPolicy OptimizationMathematical ReasoningMachine LearningNatural Language ProcessingLarge Language Model