NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-15ENGLISH EDITION
This issue
—
All time
—

AI Blog

1 story
01

Launches Partner Network with $150 Million Investment

OpenAI has launched the OpenAI Partner Network alongside a committed $150 million investment to accelerate enterprise artificial intelligence adoption and digital transformation globally. The structured initiative is designed to assist global system integrators, technology partners, and advisory firms in deploying advanced AI models and solutions directly within enterprise environments. Through this program, OpenAI provides partner organizations with dedicated funding, technical expertise, and resources necessary to scale AI integrations. (source: https://openai.com/index/introducing-openai-partner-network)

OpenAI Partner NetworkEnterprise AIAI AdoptionPartner FundingDigital TransformationArtificial IntelligenceProduct LaunchIndustry News

Hacker News

8 stories
01

Salesforce to Acquire Fin (formerly Intercom) for $3.6B

Salesforce has signed a definitive agreement to acquire customer service platform Fin, formerly known as Intercom, in a transaction valued at $3.6 billion. This strategic acquisition is designed to accelerate Salesforce's autonomous agent roadmap, specifically enhancing its Service Cloud and Agentforce suites. By integrating Fin's customer service technology, which leverages advanced generative artificial intelligence and large language models, Salesforce aims to deliver more powerful conversational AI agents capable of resolving complex customer inquiries autonomously. This merger represents a major consolidation in the AI-driven customer experience market (source: https://www.salesforce.com/news/press-releases/2026/06/15/salesforce-signs-definitive-agreement-to-acquire-fin/?bc=HL).

Autonomous AgentsGenerative AICustomer Relationship ManagementEnterprise SoftwareAgentic AIArtificial IntelligenceAI AgentGenerative AI
02

Apple Foundation Models

Apple has released and integrated its proprietary foundation models within modern software development ecosystems, emphasizing on-device performance and computational optimization on Apple silicon. The integration allows developers to build highly responsive, privacy-focused applications using advanced machine learning capabilities natively. The documentation details the APIs, library structures, and software development kits available for deployment. This release enables developers to run sophisticated language and multimodal tasks locally, expanding generative AI applications without relying on cloud infrastructure (source: https://platform.claude.com/docs/en/cli-sdks-libraries/libraries/apple-foundation-models).

Apple Foundation ModelsOn-Device AIApple SiliconMachine Learning SDKLocal InferenceArtificial IntelligenceLarge Language ModelGenerative AI
03

Anthropic's Safety Superpower

Stratechery has published an analysis of Anthropic's unique position in the competitive generative AI landscape, focusing on its strategic emphasis on safety as a primary competitive differentiator. The article explores how Anthropic's commitment to alignment research, constitutional AI, and robust safety protocols serves as an enterprise-grade capability that attracts risk-averse corporate clients. The piece argues that establishing rigorous safety boundaries gives Anthropic the freedom to push large language model capabilities forward with greater confidence (source: https://stratechery.com/2026/anthropics-safety-superpower/).

Constitutional AIAI SafetyModel AlignmentLarge Language ModelsEnterprise AIArtificial IntelligenceLarge Language ModelGenerative AI
04

Openrouter Fusion API

OpenRouter has introduced the Fusion API, a novel routing and model aggregation service designed to optimize performance and reduce latency for developers using large language models. The platform acts as an intelligent intermediary, dynamically combining and routing prompts to the most efficient AI models based on cost, speed, and capability requirements. Serving as a unified gateway, the Fusion API simplifies the integration of diverse machine learning backends, allowing developers to build robust, multi-model AI agents and applications through a single streamlined endpoint (source: https://openrouter.ai/openrouter/fusion).

Fusion APIModel RoutingAPI GatewayMulti-Model IntegrationModel AggregationArtificial IntelligenceLarge Language ModelGenerative AI
05

Claude Corps

Anthropic has introduced Claude Corps, a new initiative designed to expand the capabilities and deployment of their Claude AI model in professional workflows. This initiative focuses on deploying highly capable, contextual, and safe AI systems to assist users in complex decision-making, programming, and writing tasks. By establishing this specialized corps, Anthropic aims to refine model interactions, optimize performance across diverse domains, and ensure robust safety standards for generative artificial intelligence in enterprise environments (source: https://www.anthropic.com/news/claude-corps).

ClaudeAnthropicAI SafetyGenerative AIEnterprise AIArtificial IntelligenceLarge Language ModelAI Agent
06

Can Europe train a frontier AI model on the compute it owns?

A new investigation explores the feasibility of training frontier artificial intelligence models using Europe's existing, distributed high-performance computing (HPC) infrastructure. The project, named 'EuroMesh', proposes a decentralized approach to pooling public and sovereign supercomputing clusters across various European nations. By addressing critical bottlenecks in high-latency, low-bandwidth inter-datacenter communication, the repository models how distributed training algorithms and specialized network topologies can synchronize gradients across geographically separated clusters to compete with centralized AI labs (source: https://github.com/sammysltd/euromesh).

Distributed TrainingHigh-Performance ComputingFrontier AI ModelsModel ParallelismDecentralized ComputeArtificial IntelligenceLarge Language ModelGenerative AI
07

Launch HN: Drafted (YC P26) – Models for residential architecture

Drafted, a Y Combinator-backed startup (YC P26), has launched a new platform that trains machine learning models to generate detailed residential architecture based on structured design constraints. Traditional custom home design ranges from $10,000 to $50,000 and takes months, preventing most homes from using professional architectural input. Drafted addresses this by training generative AI models to understand physical space and the built environment, enabling users to easily visualize, explore, and customize home designs (source: https://news.ycombinator.com/item?id=48543908).

Generative DesignSpatial ComputingArchitectural AIDesign ConstraintsPhysical Space ModelingArtificial IntelligenceGenerative AIOthers
08

Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?

A popular Hacker News discussion explores whether software developers have successfully transitioned from commercial AI systems like Claude and GPT to local, open-weight Large Language Models for daily programming. The inquiry gathers feedback on hardware configurations, model architectures, orchestration frameworks, and quantitative performance metrics such as tokens per second. This thread highlights growing developer interest in achieving data privacy, offline capabilities, and cost savings by hosting powerful open-source alternatives directly on local workstations (source: https://news.ycombinator.com/item?id=48542100).

Local LLMAI Coding AssistantsOpen-Weight ModelsInference PerformanceHardware AccelerationArtificial IntelligenceLarge Language ModelGenerative AI

Twitter

4 stories
01

Sakana AI Launches Marlin: An Autonomous Ultra Deep Research Agent

Sakana AI has officially launched Sakana Marlin, its inaugural commercial product designed to function as an autonomous Ultra Deep Research agent and virtual Chief Scientific Officer (CSO). Moving beyond traditional research assistants, Marlin is engineered to automate complex scientific workflows and data-intensive analytical tasks for enterprise clients. The release marks Sakana AI's transition from theoretical studies to commercializing specialized agentic solutions in the professional market. This entry consolidates multiple concurrent announcements from the company. (source: https://x.com/hardmaru/status/2066529282588094713)

Sakana AIMarlinAI AgentDeep ResearchVirtual CSOArtificial IntelligenceAI AgentProduct Launch
02

Runway Video Generation Tools Integrated Directly Into ChatGPT

Runway has officially integrated its advanced video generation capabilities directly into the ChatGPT interface. This partnership allows users to generate high-quality video content using conversational natural language prompts without leaving the ChatGPT environment. The collaboration reflects an industry trend toward integrating specialized multimodal generative AI models into unified conversational platforms, dramatically expanding the accessibility of professional-grade video tools for creators and visual storytellers. (source: https://x.com/c_valenzuelab/status/2066599959437160754)

RunwayChatGPTGenerative AIVideo GenerationIntegrationGenerative AIMultimodalProduct Launch
03

Kling AI Teases Major Upcoming Updates With Cinematic Announcement

Kling AI has released a cinematic teaser titled 'Winter is Coming. Legends Will Rise' to signal an upcoming major update to its generative video platform. Although specific technical documentation has not yet been released, the update is expected to improve motion coherence, expand multimodal functionalities, and increase output fidelity. The strategic teaser targets professional creators who require high-quality video synthesis toolsets in an increasingly competitive AI video market. (source: https://x.com/Kling_ai/status/2066536216523391327)

Kling AIGenerative VideoVideo GenerationAI InnovationCinematic AIArtificial IntelligenceGenerative AIProduct Launch
04

AutoScientist Challenge Launches With Fifty Thousand Dollar Healthcare Prize

The AutoScientist Challenge has launched its healthcare track, inviting developers to build frontier artificial intelligence models optimized for medical applications. The competition offers a total prize pool of fifty thousand dollars to incentivize advancements in health sciences. Participants will leverage advanced modeling and machine learning architectures to address complex clinical research challenges, showcasing the practical utility of frontier generative AI in active scientific fields. (source: https://x.com/sarahookr/status/2066572989752390022)

AutoScientistHealthcare AIFrontier ModelsAI CompetitionMedical InnovationArtificial IntelligenceLarge Language ModelIndustry News

huggingface

8 stories
01

From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI

Researchers have conceptualized the structural transition of Large Language Models from episodic chatbots into integrated digital colleagues. The paper maps this evolution along two dimensions: the cognitive core level, where systems progress toward thinking LLMs leveraging inference-time computation, and the task execution level, which moves toward workstation systems equipped with persistent workspaces, reusable skills, and verification loops. The authors analyze how this shift updates data construction from simple instruction-response pairs to complex state-action-observation trajectories, requiring evaluations to transition from static benchmarks to sandboxed, auditable, and self-evolving environments. (source: https://huggingface.co/papers/2606.14502)

Thinking LLMsDigital ColleagueAI AgentsInference-time ComputationWorkspace paradigmArtificial IntelligenceLarge Language ModelAI Agent
02

APPO: Agentic Procedural Policy Optimization

Researchers have introduced Agentic Procedural Policy Optimization (APPO), a reinforcement learning framework that improves the multi-turn tool-use capabilities of large language model agents. While traditional agentic reinforcement learning assigns credit over coarse heuristic units, APPO targets fine-grained decision points. It selects branching locations using a Branching Score that combines token uncertainty with policy-induced likelihood gains of continuations, while introducing procedure-level advantage scaling to distribute credit. Across 13 benchmarks, APPO consistently improves strong agentic reinforcement learning baselines by nearly 4 points, maintaining efficient tool calls and behavioral interpretability. (source: https://huggingface.co/papers/2606.12384)

Agentic Reinforcement LearningCredit AssignmentBranching ScoreLarge Language ModelsPolicy OptimizationArtificial IntelligenceLarge Language ModelAI Agent
03

Orchestra-o1: Omnimodal Agent Orchestration

Researchers have proposed Orchestra-o1, an omnimodal agent orchestration framework designed for agent swarms collaborating across text, image, audio, and video modalities. Orchestra-o1 implements modality-aware task decomposition, online sub-agent specialization, and parallel sub-task execution. To optimize this system, the authors introduced decision-aligned group relative policy optimization (DA-GRPO), an efficient reinforcement learning approach for training Orchestra-o1-8B. The framework outperforms the second-best approach by 10.3% accuracy on the OmniGAIA benchmark and achieves state-of-the-art performance against existing open-source omnimodal agents. (source: https://huggingface.co/papers/2606.13707)

multi-agent systemsagent orchestrationomnimodal learningdecision-aligned group relative policy optimizationagentic reinforcement learningArtificial IntelligenceMultimodalAI Agent
04

Pythagoras-Prover: Advancing Efficient Formal Proving via Augmented Lean Formalisation

Researchers have introduced Pythagoras-Prover, a compute-efficient open-source family of Lean theorem provers containing 4B and 32B autoregressive models alongside a proof-of-concept diffusion-based prover. To train these models, the team constructed a verified Lean corpus stratified for curriculum training and developed Augmented Lean Formalisation (ALF), which expands training data through statement mutation and self-distillation. Pythagoras-Prover-4B surpasses DeepSeek-Prover-V2-671B at pass@32 on MiniF2F-Test (86.1% vs 82.4%) with approximately 167 times fewer parameters. The 32B model sets an open-source state-of-the-art at 93.0% on MiniF2F-Test and solves 93 PutnamBench problems. (source: https://huggingface.co/papers/2606.12594)

Lean theorem provingAugmented Lean Formalisationcurriculum SFTdiffusion-based proverformal reasoningArtificial IntelligenceMachine LearningLarge Language Model
05

Smaller Models are Natural Explorers for Policy-Level Diversity in GRPO

Researchers have proposed Small-to-Large Policy Optimization (S2L-PO), a training framework designed to enhance rollout diversity in Group Relative Policy Optimization (GRPO) for language models. Recognizing that smaller models within the same family inherently exhibit higher, temporally correlated policy-level diversity, the framework leverages a fixed small model to generate exploratory rollouts for training a larger model. A progressive annealing strategy transitions sampling from the small model to the large model to prevent performance drops. S2L-PO improves mathematical reasoning accuracy on AIME 24 by 8.8% using a 1.7B explorer to guide an 8B model. (source: https://huggingface.co/papers/2605.30789)

Group Relative Policy OptimizationSmall-to-Large Policy Optimizationpolicy-level diversitymathematical reasoningrollout explorationMachine LearningNatural Language ProcessingLarge Language Model
06

Avatar V: Scaling Video-Reference Avatar Video Generation

Researchers have introduced Avatar V, a production-scale framework for generating avatar videos based on video-reference conditioning. Moving beyond static image inputs, Avatar V utilizes the full token sequence of a reference video to preserve both visual identity and behavioral patterns like talking style and expressions. The architecture incorporates Sparse Reference Attention for linear-complexity conditioning on long references, alongside talking style transfer and an identity-aware refiner. Trained on over 100 million video clips using flow matching and reinforcement learning from human feedback, Avatar V consistently outperforms leading systems including Seedance 2.0 and Kling O3 Pro in identity preservation. (source: https://huggingface.co/papers/2606.13872)

Avatar Video GenerationVideo-Reference ConditioningSparse Reference AttentionTalking Style TransferGenerative AIGenerative AIComputer VisionMultimodal
07

HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry

Researchers have introduced HarnessX, a foundry designed to compose, adapt, and evolve runtime interfaces for AI agents. HarnessX assembles typed harness primitives using a substitution algebra and adapts them through AEGIS, a trace-driven multi-agent evolution engine. The system closes the harness-model loop by converting execution trajectories into both harness updates and model training signals. Across five agent benchmarks (ALFWorld, GAIA, WebShop, tau^3-Bench, and SWE-bench Verified), HarnessX yields an average improvement of 14.5% (up to 44.0%), demonstrating that systematically adapting execution interfaces provides a major lever for improving agent performance. (source: https://huggingface.co/papers/2606.14249)

agent harnesssubstitution algebraevolution engineexecution feedbackruntime interfaceArtificial IntelligenceLarge Language ModelAI Agent
08

AdaSR: Adaptive Streaming Reasoning with Hierarchical Relative Policy Optimization

Researchers have proposed AdaSR, an adaptive streaming reasoning framework designed for dynamic, continuous inputs like audio and video streams. Unlike static read-then-think architectures, AdaSR enables models to reason progressively during input streaming and perform final deliberation once the stream concludes. To optimize this multi-stage execution, the authors developed Hierarchical Relative Policy Optimization (HRPO), which splits policy optimization into streaming and deep reasoning phases. By combining format, accuracy, and adaptive thinking rewards, HRPO helps the model learn when to reason and how much computation to allocate, showing a superior latency-efficiency trade-off compared to supervised baselines. (source: https://huggingface.co/papers/2606.14694)

streaming reasoningpolicy optimizationcomputation allocationadaptive thinkinghierarchical reasoningArtificial IntelligenceMachine LearningLarge Language Model