NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-23ENGLISH EDITION
This issue
—
All time
—

AI Blog

5 stories
01

Launches Claude Tag for Collaborative Teamwork in Slack

Anthropic launched the beta release of Claude Tag, a collaborative feature integrating Claude directly into Slack as an active team member. Users can tag @Claude in channels to trigger task execution, where the agentic model breaks down projects, codes, and utilizes connected tools. The feature supports multiplayer collaboration, contextual memory, and proactive updates. Currently, 65% of Anthropic's own product team's code is generated using an internal version of Claude Tag. System administrators can manage data and tool permissions for separate channel-scoped Claude identities. (source: https://www.anthropic.com/news/introducing-claude-tag)

Claude TagSlack IntegrationAI AgentAnthropicCollaborative AIAI AgentProduct LaunchTech News
02

GPT-5 Pro Helps Solve Three-Year-Old Immunology Mystery

OpenAI announced that its GPT-5 Pro model assisted immunologist Derya Unutmaz in resolving a three-year-old mystery regarding T cell behavior. By analyzing complex immunology datasets, the advanced large language model successfully decoded cellular patterns that had previously eluded researchers. This milestone demonstrates how frontier models can be leveraged to accelerate scientific discovery, specifically in modeling cellular biology to support future medical research into autoimmune diseases and cancer treatments. (source: https://openai.com/index/gpt-5-immunology-mystery)

GPT-5 ProOpenAIImmunologyT CellsMedical AILarge Language ModelResearch ProgressArtificial Intelligence
03

The Rise of Harness Level Loops in Agentic Code Generation

Software developer Armin Ronacher detailed a shift in software engineering where developers write external harness loops to run coding agents autonomously. In analyzing autonomous tools like Claude Code, Ronacher notes that unsupervised iterations by large language models often degrade overall code quality. Rather than establishing robust system design, these models tend to produce overly defensive, complex, and localized workarounds that mask underlying architectural issues, contrasting with higher-quality human-in-the-loop developer workflows. (source: https://lucumr.pocoo.org/2026/6/23/the-coming-loop/)

Coding AgentsAgentic LoopsClaude CodeSoftware EngineeringLLM Code QualityAI AgentLarge Language ModelTech News
04

Omio Builds Conversational Travel Experiences Using OpenAI

Omio has partnered with OpenAI to develop conversational travel experiences, integrating generative AI into its primary booking systems. The partnership enables users to plan complex travel itineraries, compare multi-modal routes, and book transportation using conversational interfaces. The implementation is part of Omio's broader organizational transition to an AI-native company, aimed at streamlining internal software engineering workflows and generating highly personalized travel recommendations based on real-time consumer queries. (source: https://openai.com/index/omio)

OmioOpenAIConversational TravelProduct DevelopmentAI-nativeGenerative AIIndustry NewsNatural Language Processing
05

Josh Elman Joins Andreessen Horowitz to Back Consumer AI Startups

Venture capital firm Andreessen Horowitz announced that Josh Elman has joined as a partner to identify and support early-stage consumer AI companies. Leveraging his previous product engineering and leadership experience at Facebook, LinkedIn, Twitter, and Robinhood, Elman will assist founders with product development, growth loops, and consumer patterns. The move highlights a venture thesis focused on generative AI interfaces alongside a rising generation of digital native users accustomed to interactive platforms. (source: https://a16z.com/the-world-building-doors-are-open-again/)

Josh ElmanAndreessen HorowitzConsumer AIVenture CapitalProduct DevelopmentArtificial IntelligenceGenerative AIIndustry News

Hacker News

5 stories
01

OpenAI DayBreak – GPT-5.5-Cyber

OpenAI has introduced GPT-5.5-Cyber, a specialized reasoning model designed for defensive cybersecurity operations, as part of its DayBreak initiative. The model is engineered for vulnerability analysis, automated patching, real-time threat hunting, and reverse engineering. It leverages advanced reasoning to assist security professionals and automate defense procedures at scale. This milestone reflects a broader strategic shift toward deploying domain-specialized autonomous agents to safeguard digital infrastructure against increasingly complex cyber threats. (source: https://openai.com/index/daybreak-securing-the-world/)

GPT-5.5-CyberCybersecurity AIVulnerability AnalysisThreat HuntingDayBreak InitiativeArtificial IntelligenceLarge Language ModelAI Agent
02

VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO

Researchers have released VibeThinker, a 3-billion-parameter language model that outperforms larger systems like Claude Opus 4.5 on complex reasoning tasks. The model utilizes a novel training methodology combining Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) to optimize internal search paths and reasoning trajectories. This reinforcement learning breakthrough demonstrates that compact parameter configurations can achieve high-tier logical and scientific reasoning performance, offering a highly efficient alternative for localized edge deployments. (source: https://arxiv.org/abs/2606.16140)

VibeThinkerGroup Relative Policy OptimizationSupervised Fine-TuningModel ReasoningParameter EfficiencyArtificial IntelligenceLarge Language ModelDeep Learning
03

The Coming Loop

An analysis explores the "Coming Loop," a transition where artificial intelligence systems autonomously design, train, and refine subsequent generations of AI without human intervention. The article details the shift from human-in-the-loop engineering to fully autonomous, recursive self-improvement cycles. This self-reinforcing loop accelerates capabilities exponentially, introducing complex alignment and control challenges. The piece concludes that managing this feedback loop is the most critical challenge currently facing AI safety and computer science. (source: https://lucumr.pocoo.org/2026/6/23/the-coming-loop/)

Feedback LoopRecursive Self-ImprovementAI SafetyAutonomous SystemsArtificial General IntelligenceArtificial IntelligenceGenerative AIAI Agent
04

Anthropic updates their terms to verify age or identity

Anthropic has updated its service terms and privacy policies to introduce new age and identity verification requirements. The verification mechanisms aim to ensure regulatory compliance and enhance user safety, specifically regarding underage access to sophisticated artificial intelligence models. This change highlights an industry-wide trend toward stricter governance and compliance frameworks for commercial large language models, balancing broad application access with responsible development and risk mitigation. (source: https://www.anthropic.com/legal/privacy)

AnthropicIdentity VerificationPrivacy PolicyComplianceAI SafetyArtificial IntelligenceLarge Language ModelGenerative AI
05

Introducing Claude Tag

Anthropic has launched Claude Tag, a new productivity feature designed to let users classify and organize their conversation threads. This update aims to improve productivity and workflow efficiency within generative AI platforms by enabling custom labels for grouping, retrieving, and managing past AI interactions. It reflects a broader industry trend of expanding simple conversational assistants into structured productivity suites that support complex, multi-session tasks. (source: https://www.anthropic.com/news/introducing-claude-tag)

Claude TagGenerative AIInformation RetrievalUser InterfaceProductivity ToolArtificial IntelligenceGenerative AINatural Language Processing

Twitter

7 stories
01

Google Introduces Gemini Spark Personal AI Agent For Workflow Automation

Google has officially introduced Gemini Spark, a new personal AI agent designed to automate complex, multi-step digital workflows. Operating continuously as a 24/7 digital assistant, the system performs professional and personal tasks from start to finish under direct user guidance. This product launch highlights Google's strategic shift toward autonomous agentic behaviors that move beyond conversational chat interfaces to execute integrated actions across various digital environments. (source: https://x.com/google/status/2069478010403594612)

Gemini SparkAI AgentWorkflow AutomationProductivityGoogleArtificial IntelligenceAI AgentProduct Launch
02

Anthropic Introduces Claude Tag For Enhanced Team Collaboration In Slack

Anthropic has announced the launch of Claude Tag, a new enterprise feature that integrates the Claude AI model directly into Slack workspaces. Operating as an active participant in team channels, the tool accesses shared information to assist with real-time collaboration, analytical tasks, and group project management. This product release expands Claude's utility by moving the model closer to an integrated corporate workspace assistant. (source: https://x.com/AnthropicAI/status/2069469669929386160)

ClaudeAnthropicSlackAI IntegrationEnterprise AIArtificial IntelligenceLarge Language ModelProduct Launch
03

OpenAI Announces Strategic Partnership and Integration with Samsung

OpenAI has announced a strategic partnership with Samsung to integrate its generative AI technology and large language models directly into Samsung's consumer electronics ecosystem. The collaboration aims to deploy advanced AI capabilities across mobile phones, personal computers, and household appliances. This deployment focuses on enabling localized on-device processing, improving productivity workflows, and enabling intuitive user interactions. (source: https://x.com/gdb/status/2069232102483378395)

OpenAISamsungGenerative AIConsumer ElectronicsStrategic PartnershipArtificial IntelligenceLarge Language ModelProduct Launch
04

Kling 3.0 Turbo Launches With Enhanced Video Generation Performance

Kling AI has released Kling 3.0 Turbo, an optimized iteration of its generative video platform, now available on SocialSight. The update modifies the underlying architecture to deliver faster processing times and reduce generation latency without compromising overall output fidelity. This release maintains standard stylistic controls and motion consistency, targeting creators who require efficient, high-fidelity automated video generation. (source: https://x.com/Kling_ai/status/2069331784903397773)

Kling 3.0 TurboGenerative AIVideo GenerationSocialSightAI ToolsArtificial IntelligenceGenerative AIProduct Launch
05

Recovering Agent World Models Through Inverted Bellman Equation Analysis

Google DeepMind researchers have introduced a methodology for recovering an agent's internal world model from its value function by mathematically inverting the Bellman equation. This approach decodes the latent environmental understanding developed during reinforcement learning. By connecting value outputs to hidden environmental perceptions, this technique aims to improve transparency, model alignment, and architectural interpretability in reinforcement learning research. (source: https://x.com/GoogleDeepMind/status/2069433539116912739)

Bellman EquationReinforcement LearningWorld ModelAI InterpretabilityAgentic SystemsArtificial IntelligenceMachine LearningResearch Progress
06

Sakana AI Releases Fugu Technical Report and Detailed Documentation

Sakana AI has officially published the Fugu Technical Report, alongside detailed model release notes. The documentation outlines the model's design philosophy, performance benchmarks, and underlying system architecture. Designed as an open resource for AI developers and researchers, this publication provides insights into the experimental findings and structural efficiencies achieved during the Fugu project development cycle. (source: https://x.com/hardmaru/status/2069209921393144318)

Sakana AIFuguTechnical ReportRelease NotesModel ArchitectureArtificial IntelligenceResearch ProgressLarge Language Model
07

Insight Into Zhipu AI And The Development Of The GLM Model Family

This technical overview details the corporate trajectory and technical milestones of Zhipu AI, the developer of the General Language Model (GLM) series. The analysis explores Zhipu AI's role in engineering regional foundation models and its impact on the commercial landscape of large language models. The report details the core design aspects of GLM linguistic architectures, tracking their deployment in complex natural language applications. (source: https://x.com/natolambert/status/2069437601439003070)

Zhipu AIGLMLarge Language ModelsArtificial IntelligenceFoundation ModelsArtificial IntelligenceLarge Language ModelIndustry News

huggingface

8 stories
01

Self-Compacting Language Model Agents

Researchers have introduced SelfCompact, a training-free scaffolding framework designed to mitigate context window explosion in LLM agents. Instead of relying on rigid, token-threshold-based context pruning, SelfCompact equips models with a compaction tool to summarize their history alongside a lightweight decision rubric to trigger the compaction. Evaluated across six mathematical reasoning and agentic search benchmarks using seven open-weight models, the method improved math accuracy by up to 18.1 points and search task performance by 5 to 9 points, while achieving a 30% to 70% reduction in per-question token costs. (source: https://huggingface.co/papers/2606.23525)

context compactionLLM agentsinference-time scaffoldingadaptive summarizationtoken efficiencyArtificial IntelligenceLarge Language ModelAI Agent
02

PolicyTrim: Boosting Intrinsic Policy Efficiency of Vision-Language-Action Models

Researchers proposed PolicyTrim, a reinforcement learning-based post-training framework designed to boost the deployment efficiency of Vision-Language-Action (VLA) models in robotic tasks. PolicyTrim addresses unreliability and physical action redundancy by employing a dynamic exploration strategy that expands the reliable execution lengths of predicted action chunks, combined with a redundancy-aware reward that minimizes unnecessary physical steps. Across three benchmarks and three VLA models, the framework improved action chunk utilization by three times and reduced physical execution steps by 51.4%, achieving an end-to-end robotic deployment speedup of up to 5.83 times without degrading success rates. (source: https://huggingface.co/papers/2606.22540)

Vision-Language-Action ModelsPolicy EfficiencyReinforcement LearningRobotic ManipulationAction ChunkingRoboticsAI AgentMultimodal
03

Training Open Models for Agentic Phone Use

To address the challenges of training phone-use agents in dynamic real-world environments, researchers developed PhoneBuddy, a specialized training recipe and open-weight model series. PhoneBuddy leverages a mock-app environment called PhoneWorld alongside real Android applications. The training process sequentially applies supervised fine-tuning and a mixed reinforcement learning approach across both environment types. Evaluated on a 150-task human evaluation on physical phones, the model increased task success rate from 36.67% after supervised fine-tuning to 45.33% after mixed reinforcement learning, and achieved 83.2% success on the AndroidWorld benchmark. (source: https://huggingface.co/papers/2606.23049)

Agentic Phone UseReinforcement LearningGUI AgentsMock EnvironmentsSupervised Fine-TuningAI AgentMachine LearningLarge Language Model
04

A Verifiable Search Is Not a Learnable Chain-of-Thought

This paper investigates the computational limits of chain-of-thought (CoT) distillation in large language models. Through systematic experiments across model sizes ranging from 3B to 671B parameters, the author demonstrated that while forward-computable procedures like arithmetic transfer readily (achieving accuracy above 0.99), backtracking search tasks like cryptarithms fail to train reliably. Even under extensive reinforcement learning, self-training, and varied CoT architectures, distillation of backtracking search remained capped at 0.01 to 0.07 accuracy. The study isolates search-based reasoning as a fundamental bottleneck for left-to-right autoregressive sequence models. (source: https://huggingface.co/papers/2606.21884)

Chain-of-ThoughtBacktracking SearchModel Fine-TuningLarge Language ModelsReasoning LimitsArtificial IntelligenceMachine LearningLarge Language Model
05

Vera: A Layered Diffusion Model for Content-Preserving Video Editing

Researchers introduced Vera, a novel layered video diffusion model designed to solve the content preservation problem in generative video editing. Unlike conventional systems that regenerate entire pixel frames, Vera generates a separate edit layer and an alpha matte to composite edits cleanly over the original source footage. The architecture extends text-to-video Diffusion Transformers into a Mixture-of-Transformers (MoT) design to enable interaction via joint self-attention. Trained on a high-quality dataset of 486,000 layered frames, Vera demonstrated competitive edit quality and superior content preservation compared to previous open-source alternatives. (source: https://huggingface.co/papers/2606.23610)

Video EditingLayered DiffusionContent PreservationMixture-of-TransformersAlpha MattesComputer VisionGenerative AIDeep Learning
06

When Agents Commit Too Soon: Diagnosing Premature Commitment in LLM Agents

This work characterizes and diagnoses "premature commitment," a common failure mode where long-horizon agents settle early on a reasoning trajectory and fail to correct errors. The researchers define representational commitment as cross-run hidden-state similarity and show it acts as an early indicator of trajectory convergence across models like Llama-3.1-70B and Qwen-2.5-72B. Using activation states, a runtime monitor predicted trajectory consistency with an AUROC up to 0.97. A test-time prompting intervention successfully reduced agent behavioral variance by 28% without affecting downstream accuracy, proving helpful for tracking hidden process failures. (source: https://huggingface.co/papers/2606.22936)

Premature CommitmentLLM AgentsHidden-State ConvergenceBehavioral ConsistencyRepresentational CommitmentLarge Language ModelAI AgentArtificial Intelligence
07

Connect the Dots: Training LLMs for Long-Lifecycle Agents with Cross-Domain Generalization Via Reinforcement Learning

This paper presents Connect the Dots (CoD), a general framework for training large language models to support long-lifecycle agent deployment. Under the CoD framework, agents must solve sequential tasks, autonomously explore environments, learn from errors, and iteratively update their operating contexts. The authors designed a GRPO-style reinforcement learning algorithm with fine-grained credit assignment, optimized for rollouts that interleave task-solving and context-updating steps. Empirical results validate that CoD training successfully elicits the intended meta-capabilities and demonstrates strong out-of-distribution generalization performance across diverse operational domains. (source: https://huggingface.co/papers/2606.20002)

Reinforcement LearningLarge Language ModelsAI AgentsCross-Domain GeneralizationLong-Lifecycle AgentsArtificial IntelligenceLarge Language ModelAI Agent
08

Exploring the Design Space of Reward Backpropagation for Flow Matching

To address backpropagation difficulties in aligning text-to-image models, the authors introduced FlowBP, a unified surrogate-trajectory framework. FlowBP separates sampling from optimization by combining a no-gradient cached rollout for sampling with a lightweight, backward surrogate path constructed from cached and re-forwarded velocities. This formulation isolates key design parameters, including integration weights and bridge coupling. Three instantiated variants—FlowBP-Sparse, FlowBP-Bridge, and FlowBP-Lagrange—successfully bound training memory by active-set size and limit gradient chaining, yielding improvements across SD3.5-M, FLUX.1-dev, and FLUX.2-Klein-base models on quality and composition metrics. (source: https://huggingface.co/papers/2606.11075)

Flow MatchingReward BackpropagationSurrogate TrajectoryAlignmentText-to-ImageMachine LearningDeep LearningGenerative AI