NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-06-11ENGLISH EDITION
This issue
—
All time
—

AI Blog

6 stories
01

to Acquire Ona to Power Long-Running AI Agents

OpenAI has announced plans to acquire Ona to expand the capabilities of its Codex model. Through this acquisition, OpenAI will integrate Ona's secure and persistent cloud environments into its ecosystem. The integration is designed to enable the deployment of long-running AI agents capable of executing complex, multi-step enterprise workflows. This transition is expected to strengthen the developer tools and infrastructure surrounding Codex, providing a robust framework for secure agentic execution in enterprise settings. (source: https://openai.com/index/openai-to-acquire-ona)

OpenAIOnaCodexAI agentscloud environmentsAI AgentIndustry NewsGenerative AI
02

DXC Technology Integrates Claude into Regulated Industry Systems

Anthropic announced a multi-year global alliance with DXC Technology to integrate Claude into mission-critical systems across regulated industries including banking, aviation, insurance, manufacturing, and government. DXC will train tens of thousands of Claude-certified forward-deployed engineers through the Anthropic Academy. Prior to this rollout, DXC utilized Claude to generate over 95 percent of the codebase for its DXC OASIS AI-native orchestration platform, speeding up development tenfold. The alliance will initially target insurance, legacy code modernization, cybersecurity with a Claude Security subagent, and application services. (source: https://www.anthropic.com/news/dxc-anthropic-alliance)

DXC TechnologyClaudeDXC OASISAnthropic AcademyEnterprise AILarge Language ModelAI AgentIndustry News
03

BBVA Partners with OpenAI to Scale ChatGPT Enterprise to 100,000 Employees

BBVA has announced a global partnership with OpenAI to deploy ChatGPT Enterprise to 100,000 employees worldwide. This large-scale rollout is part of BBVA's strategy to integrate artificial intelligence into its core banking operations. By equipping its global workforce with OpenAI's enterprise-grade tools, the financial institution intends to enhance operational efficiency, foster internal innovation, and streamline customer service processes. This collaboration represents a major milestone in adopting large language models within the highly regulated financial sector. (source: https://openai.com/index/bbva)

BBVAOpenAIChatGPT EnterpriseBankingAI TransformationGenerative AIIndustry NewsProduct Launch
04

Models and Codex Join Oracle Cloud Infrastructure

OpenAI announced a collaboration enabling enterprise customers to access OpenAI models and Codex directly through their Oracle Cloud commitments. This partnership allows businesses to build, test, and deploy generative AI applications within the secure framework of Oracle Cloud Infrastructure. By leveraging existing financial commitments, enterprise clients can integrate advanced natural language capabilities while maintaining strict governance, compliance, and enterprise-grade security standards across their organizational workflows. (source: https://openai.com/index/openai-on-oracle-cloud)

OpenAIOracle CloudCodexEnterprise SecurityCloud InfrastructureProduct LaunchIndustry NewsGenerative AI
05

Launches Claude Corps Fellowship to Drive AI Adoption in Nonprofits

Anthropic announced the launch of Claude Corps, a national fellowship program designed to connect early-career professionals with American nonprofits to help them leverage artificial intelligence. Backed by an initial $150 million commitment from Anthropic, the initiative aims to train 1,000 fellows to use Claude, matching them with at least 400 nonprofits including Code the Dream, Braven, and Social Finance. Fellows will receive an $85,000 salary with benefits over the 12-month program. CodePath will lead training, and Social Finance will handle evaluation. (source: https://www.anthropic.com/news/claude-corps)

Claude CorpsAnthropicCodePathSocial FinanceNonprofit AI InitiativeArtificial IntelligenceProduct LaunchIndustry News
06

Astrophysicist Uses OpenAI Codex to Simulate Black Holes

Astrophysicist Chi-kwan Chan is utilizing OpenAI Codex to build complex black hole simulations to study extreme physics and test Albert Einstein's theory of general relativity. By translating natural language instructions into functional code, Codex helps streamline the creation of scientific simulations. This application of generative AI demonstrates how large language models can assist researchers in astrophysics to write, debug, and optimize code for highly specialized scientific computations and physical modeling. (source: https://openai.com/index/using-codex-to-simulate-black-holes)

OpenAI CodexAstrophysicsBlack Hole SimulationGeneral RelativityScientific ComputingLarge Language ModelGenerative AIResearch Progress

Hacker News

7 stories
01

Open Reproduction of DeepSeek-R1

Hugging Face has launched Open-R1, an open-source initiative dedicated to fully reproducing DeepSeek-R1, a state-of-the-art reasoning model. The project aims to democratize access to advanced mixture-of-experts architecture and reinforcement learning techniques by sharing the complete training pipelines, datasets, and model weights. This collaborative effort empowers researchers and developers to analyze, modify, and build upon the reasoning capabilities of DeepSeek-R1, fostering rapid innovation in the open-science AI community. (source: https://github.com/huggingface/open-r1)

DeepSeek-R1Open-Source AIReinforcement LearningReasoning ModelsMixture of ExpertsArtificial IntelligenceLarge Language ModelGenerative AI
02

Anthropic apologizes for invisible Claude Fable guardrails

Anthropic has issued a public apology regarding the implementation of undocumented, invisible safety and distillation guardrails within its Claude Fable model. The decision follows user feedback regarding unexpected restrictions and system behaviors that were not clearly communicated or logged in the standard outputs of the Claude API. Developers highlighted that these hidden constraints compromised the predictability and transparency of the system. In response, Anthropic acknowledged the communication gap and promised to refine their deployment processes to ensure clearer notifications when safety filters are triggered. This issue was also discussed widely on Hacker News. (source: https://www.theverge.com/ai-artificial-intelligence/948280/anthropic-claude-fable-invisible-distillation-guardrail)

AnthropicClaude FableAI SafetyModel GuardrailsAPI TransparencyArtificial IntelligenceLarge Language ModelNatural Language Processing
03

OpenAI to acquire Ona to expand Codex

OpenAI has announced a strategic agreement to acquire Ona, a specialized technology developer, in a move to accelerate and expand the capabilities of its Codex model. Codex translates natural language instructions into computer code and serves as the foundation for developer tools like GitHub Copilot. By integrating Ona's engineering expertise and infrastructure, OpenAI aims to enhance Codex's proficiency in code generation, context understanding, and support for diverse programming languages. This acquisition signals a major consolidation in the AI-assisted coding tool market. (source: https://openai.com/index/openai-to-acquire-ona/)

CodexOnaCode GenerationAcquisitionDeveloper ToolsArtificial IntelligenceLarge Language ModelGenerative AI
04

OpenAI mulls slashing prices as it competes with Anthropic for users

OpenAI is actively considering significant price reductions for its developer tools and API access to maintain market share against its primary competitor, Anthropic. This strategic pricing evaluation is designed to retain enterprise customers and developers who are increasingly exploring alternative large language models in the generative artificial intelligence market. As operational efficiencies improve unit economics for leading AI providers, the ongoing price competition reflects a broader industry push to secure long-term developer loyalty and platform ecosystem lock-in. (source: https://www.cnbc.com/2026/06/11/openai-mulls-slashing-prices-ahead-of-competition-from-anthropic-wsj.html)

API PricingGenerative AIEnterprise AIAI CompetitionLarge Language ModelsArtificial IntelligenceGenerative AILarge Language Model
05

MiMo Code is now released and open-source

Xiaomi has officially announced the open-source release of MiMo Code, a specialized development project designed to advance open-source software collaboration and automated code generation methodologies. This initiative aims to provide developers with accessible, transparent, and customizable programming assets to streamline software engineering workflows. By open-sourcing the codebase, Xiaomi aims to foster a collaborative community ecosystem where researchers and software engineers can contribute to, modify, and optimize code generation performance and integration capabilities across diverse developer environments. (source: https://mimo.xiaomi.com/mimocode)

MiMo CodeOpen SourceSoftware DevelopmentCode GenerationDeveloper ToolsArtificial IntelligenceLarge Language ModelOthers
06

Ask HN: How do you get into a flow state when using AI to code?

This community discussion addresses the productivity challenges developers face in maintaining a deep flow state when using slower, agentic AI tools like Claude. Traditional programming allowed for extended, uninterrupted focus, whereas the asynchronous nature of automated code generation introduces frequent waiting periods that fragment developer attention. Users explore strategies to adapt their cognitive workflows to this new paradigm, suggesting tactics such as executing parallel engineering tasks, reading documentation, or optimizing prompts to reduce iteration times and maintain momentum. (source: https://news.ycombinator.com/item?id=48492118)

Agentic CodingDeveloper ProductivityFlow StateAI AgentsCognitive WorkflowArtificial IntelligenceAI AgentGenerative AI
07

Building agents without harness engineering

This technical article analyzes a development paradigm shift aimed at building AI agents without complex harness engineering and rigid orchestration pipelines. Historically, functional agents required substantial custom software wrappers and specialized glue code to manage actions. The author proposes a streamlined development approach that focuses on leveraging the direct reasoning capabilities and native tooling of large language models. Shifting developmental focus from infrastructural complexity to core agent behavior and prompt design simplifies runtime environments and minimizes systemic overhead. (source: https://rajitkhanna.com/agents/)

AI AgentsAgentic WorkflowsHarness EngineeringSoftware ArchitectureOrchestrationArtificial IntelligenceAI AgentGenerative AI

Twitter

8 stories
01

Use Oracle Cloud Commitments To Access And Scale OpenAI Products

Greg Brockman announced that organizations can now utilize their existing Oracle Cloud infrastructure commitments to deploy and scale OpenAI's suite of generative AI models. This integration simplifies the procurement process for enterprise clients by allowing them to apply Oracle cloud credits toward advanced model deployments. The collaboration aims to lower barriers to entry for large-scale AI implementation, providing businesses with enhanced architectural flexibility to integrate language capabilities directly into their operational workflows. (source: https://x.com/gdb/status/2064899797593792687)

Oracle CloudOpenAIEnterprise AICloud InfrastructureGenerative AIArtificial IntelligenceGenerative AIIndustry News
02

Google DeepMind Launches Ten Million Dollar Research Fund For AI Agent Groups

Google DeepMind launched a ten million dollar research fund in collaboration with Schmidt Sciences, the Cooperative AI Foundation, and ARIA Research, with support from Google.org. The initiative is dedicated to studying the collective behaviors of AI agents operating in large-scale groups. Researchers will explore the emergent properties, communication frameworks, safety protocols, and complex dynamics that occur when millions of autonomous agents interact in multi-agent environments. (source: https://x.com/GoogleDeepMind/status/2065031279213441309)

AI AgentsCollective BehaviorResearch FundGoogle DeepMindMulti-Agent SystemsArtificial IntelligenceAI AgentResearch Progress
03

Fable 5 Integrates Enhanced Safeguards With Opus 4.8 Fallback Mechanism

The development team behind Fable 5 introduced a security update implementing a visible fallback mechanism to the Opus 4.8 model for flagged user requests. Starting this week, any prompt triggered by safety protocols will automatically and transparently transition to Opus 4.8, notifying the developer. This safety measure mirrors precautions used in sensitive domains such as cyber and bio-security, aiming to provide real-time feedback while maintaining rigorous safety standards. (source: https://x.com/ClaudeDevs/status/2064949876463645026)

Fable 5Opus 4.8LLM SafetyModel FallbackFrontier DevelopmentArtificial IntelligenceLarge Language ModelIndustry News
04

GPT-5.5 Performance Benchmarks Exceed Expectations Across Key Metrics

Evaluation results demonstrate that the new GPT-5.5 model exhibits strong performance and efficiency across multiple critical operational benchmarks. The model surpasses existing standards in token efficiency, computational cost, and wall-clock execution time. These architectural optimizations indicate a major development in resource-conscious designs, allowing enterprise-grade applications to utilize high-quality model outputs without experiencing a proportional increase in operational overhead or system latency. (source: https://x.com/polynoamial/status/2065125807585149136)

GPT-5.5Model EfficiencyBenchmarksComputational CostPerformance OptimizationArtificial IntelligenceLarge Language ModelIndustry News
05

Runway Expands Strategic Partnership With Lionsgate For New Initiatives

Runway announced a strategic expansion of its existing partnership with film studio Lionsgate to integrate advanced generative video technology into professional entertainment workflows. The collaboration involves specialized projects designed to deploy Runway's generative models in high-budget production environments. This expansion represents a commitments from both organizations to utilize AI-driven creative tools in traditional cinema workflows, exploring new capabilities of digital video generation in media. (source: https://x.com/c_valenzuelab/status/2065081068630282541)

RunwayLionsgateGenerative VideoPartnershipEntertainment TechnologyGenerative AIProduct LaunchIndustry News
06

DiffusionGemma Enables High-Speed Local Inference Performance

DiffusionGemma was introduced to enable high-performance local inference for developers. By optimizing the model's underlying architecture, the system achieves inference speeds exceeding 1,100 tokens per second on a single NVIDIA H100 hardware setup. This release is aimed at researchers and developers requiring high-throughput, locally-hosted generative models, offering an efficient open alternative for edge deployments and private enterprise infrastructure. (source: https://x.com/ZoubinGhahrama1/status/2065029250374078832)

DiffusionGemmaLocal InferenceH100Generative AIModel OptimizationArtificial IntelligenceGenerative AIOpen Source
07

Anthropic Launches Claude Corps Fellowship For Nonprofit AI Empowerment

Anthropic launched Claude Corps, a national fellowship program that connects early-career professionals with US-based nonprofit organizations. The initiative aims to train 1,000 compensated participants to deploy Claude for operational support, technical literacy, and productivity enhancements within their host nonprofits. The program represents an effort to bridge foundational large language model technologies with pro-social applications and real-world nonprofit environments across the United States. (source: https://x.com/AnthropicAI/status/2065057393927467084)

Claude CorpsAnthropicFellowshipAI TrainingNonprofitArtificial IntelligenceLarge Language ModelIndustry News
08

New York Premiere Of 2026 Runway AI Festival Officially Sold Out

Runway announced that the New York City premiere event for the 2026 Runway AI Film Festival has officially sold out. The festival serves as a curated showcase for AI-generated cinema, highlighting creative applications of generative video tools and machine learning in traditional filmmaking workflows. This event brought together digital content creators, traditional media professionals, and visual artists to discuss the future of AI-powered digital storytelling. (source: https://x.com/runwayml/status/2065112545073995909)

RunwayAI FestivalCreative TechGenerative AIFilm PremiereGenerative AIIndustry NewsArtificial Intelligence

huggingface

8 stories
01

Toward Generalist Autonomous Research via Hypothesis-Tree Refinement

Researchers introduced Arbor, an autonomous agent framework designed to conduct long-horizon scientific research using Hypothesis Tree Refinement (HTR). This persistent tree structure connects hypotheses, artifacts, evidence, and insights, while a long-lived coordinator directs global research strategy executed by short-lived executors. Tested on six real-world machine learning optimization tasks, Arbor outperformed both Codex and Claude Code, achieving over 2.5 times their average relative held-out gain. On MLE-Bench Lite using GPT-5.5, the system established a new state-of-the-art result of 86.36% Any Medal. (source: https://huggingface.co/papers/2606.11926)

autonomous researchAI agentshypothesis-tree refinementautonomous optimizationmachine learning systemsArtificial IntelligenceMachine LearningAI Agent
02

Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

Researchers proposed Bebop, a framework designed to integrate Multi-Token Prediction (MTP) into large-scale reinforcement learning (RL) training pipelines. To prevent MTP acceptance rate degradation caused by model entropy fluctuations during RL, Bebop uses probabilistic rejection sampling combined with a novel end-to-end Total Variation (TV) loss. This method improves MTP acceptance rates by up to 10 percentage points, reaching up to a 95% acceptance rate and delivering up to 1.8x end-to-end acceleration in asynchronous RL training across Qwen3.5, Qwen3.6, and Qwen3.7 models. (source: https://huggingface.co/papers/2606.12370)

Multi-Token PredictionReinforcement LearningRejection SamplingSpeculative DecodingEntropy BoundsArtificial IntelligenceMachine LearningLarge Language Model
03

Verifiable Environments Are LEGO Bricks: Recursive Composition for Reasoning Generalization

Researchers introduced RACES, a framework that recursively composes verifiable environments to scale up the reinforcement learning training of large language models. Rather than constructing environments manually, RACES uses composition operators like SEQUENTIAL, PARALLEL, SORT, and SELECT to automatically fuse environments when their input-output types match. Built on 300 base environments, the framework improved DeepSeek-R1-Distill-Qwen-14B by 3.1 points and Qwen3-14B by 2.3 points on unseen benchmarks. RACES achieved comparable generalization performance using only 50 base environments instead of all 300. (source: https://huggingface.co/papers/2606.12373)

Verifiable EnvironmentsReinforcement LearningRecursive CompositionReasoning GeneralizationLarge Language ModelsArtificial IntelligenceMachine LearningLarge Language Model
04

Embodied-R1.5: Evolving Physical Intelligence via Embodied Foundation Models

Researchers developed Embodied-R1.5, an 8-billion parameter embodied foundation model designed to unify tasks like planning, cognition, and physical correction within a single architecture. Leveraging data pipelines containing over 15 billion tokens, the model implements a Planner-Grounder-Corrector closed-loop framework for long-horizon execution. Embodied-R1.5 outperformed models like Gemini-Robotics-ER-1.5 and GPT-5.4 on 16 out of 24 embodied vision-language-model benchmarks. When fine-tuned on small datasets, it also outperformed leading vision-language-action models like pi_0.5 across major robotic manipulation suites. (source: https://huggingface.co/papers/2606.11324)

Embodied AIEmbodied Foundation ModelsRoboticsVision-Language-ActionPhysical IntelligenceRoboticsAI AgentMultimodal
05

Grammar-Constrained Decoding Can Jailbreak LLMs into Generating Malicious Code

Researchers identified a security vulnerability in Grammar-Constrained Decoding (GCD) techniques and introduced CodeSpear, a new jailbreak attack that exploits GCD to generate malicious code. By enforcing syntax constraints, CodeSpear bypasses standard alignment defenses, increasing attack success rates by more than 30 percentage points on average across 10 popular large language models. To mitigate this threat, the authors proposed CodeShield, a safety alignment method that trains models to output non-malicious honeypot code under grammar constraints while preserving baseline performance on benign programming tasks. (source: https://huggingface.co/papers/2606.11817)

Grammar-Constrained DecodingJailbreak AttackCode Generation SafetyLarge Language ModelsCodeShieldLarge Language ModelGenerative AINatural Language Processing
06

InternVideo3: Agentify Foundation Models with Multimodal Contextual Reasoning

Researchers introduced InternVideo3, a multimodal foundation model framework designed for long-video understanding and visual agent tasks. The framework utilizes Multimodal Contextual Reasoning (MCR) to model video interaction as an evolving, closed-loop process. To optimize memory consumption, InternVideo3 introduces Multimodal Multi-head Latent Attention (M2LA) to compress KV-cache states during inference. The model achieves state-of-the-art results on several long-horizon video benchmarks, including Video-MME, MLVU, and EgoSchema, and demonstrates robust performance when deployed as an interactive video agent. (source: https://huggingface.co/papers/2606.12195)

Multimodal Contextual ReasoningVideo AgentLong-Video UnderstandingLatent AttentionFoundation ModelsVideo UnderstandingMultimodalAI Agent
07

TRACE: A Unified Rollout Budget Allocation Framework for Efficient Agentic Reinforcement Learning

Researchers proposed Tree Rollout Allocation for Contrastive Exploration (TRACE), a dynamic rollout budget allocation framework for multi-turn agentic reinforcement learning. TRACE addresses low reward contrast in LLM policy optimization by modeling multi-turn trajectories as tree structures and focusing sampling budgets on the specific prefix-level nodes most likely to yield mixed rewards. Controlled by a predictive success model, this adaptive framework enriches training signals without expanding the computing budget. Empirically, TRACE improved the average accuracy of a Qwen3-14B model on a Multi-Hop QA benchmark by 2.8 percentage points. (source: https://huggingface.co/papers/2606.11119)

Reinforcement LearningRollout AllocationAI AgentLarge Language ModelsContrastive ExplorationMachine LearningLarge Language ModelAI Agent
08

DeNovoSWE: Scaling Long-Horizon Environments for Generating Entire Repositories from Scratch

Authors released DeNovoSWE, a large-scale, automated dataset consisting of 4,818 high-quality repository-generation instances to train code agents on long-horizon tasks. DeNovoSWE was built without human annotation using a sandboxed agentic workflow featuring a "divide and conquer" architecture, critic-repair loops, and difficulty-aware trajectory filtering. Fine-tuning a Qwen3-30B-A3B model on DeNovoSWE yielded major performance improvements on the challenging BeyondSWE-Doc2Repo benchmark, increasing its task success rate from 5.8% to 47.2%. (source: https://huggingface.co/papers/2606.10728)

Software EngineeringDataset ConstructionAgentic WorkflowRepository GenerationLong-Horizon TasksLarge Language ModelAI AgentGenerative AI