NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-10ENGLISH EDITION
This issue
—
All time
—

AI Blog

2 stories
01

UST Partners with Anthropic to Bring Claude to Physical AI and Hardware Validation

Anthropic has partnered with digital technology and engineering services provider UST to integrate Claude models into physical AI and hardware engineering workflows. Under this collaboration, UST is deploying Claude Code to help engineers read hardware schematics, write regression tests, and compare live equipment performance against digital twins. Additionally, UST is training 20,000 of its engineers, architects, and consultants on Claude. The integration of Claude into UST's iDEC platform, which currently reduces silicon and hardware validation times by 50 to 70 percent, aims to accelerate chip design and detect physical system flaws earlier. (source: https://www.anthropic.com/news/ust-claude)

USTClaude CodePhysical AIHardware ValidationSilicon EngineeringLarge Language ModelIndustry NewsArtificial Intelligence
02

Deutsche Telekom Partners with OpenAI to Build AI-Native Telco Systems

Deutsche Telekom has announced a partnership with OpenAI to integrate advanced artificial intelligence models across its telecommunications operations and infrastructure. The multi-year initiative aims to transform key operational areas, including customer service automation, internal employee workflows, and network operations management. By deploying OpenAI's generative models, the telecommunications company seeks to automate administrative tasks for its staff and improve the processing speed and resolution quality of consumer inquiries. This partnership aligns with Deutsche Telekom's broader strategic roadmap to transition into an AI-native organization. (source: https://openai.com/index/deutsche-telekom)

Deutsche TelekomOpenAITelecommunicationsAI IntegrationCustomer ServiceArtificial IntelligenceGenerative AIIndustry

Hacker News

6 stories
01

GPT-5.6 Sol Ultra produces proof of the Cycle Double Cover Conjecture

OpenAI announced that its next-generation model, GPT-5.6 Sol Ultra, has successfully generated a formal mathematical proof for the Cycle Double Cover Conjecture. The conjecture is a long-standing, unsolved problem in graph theory asserting that every bridgeless graph has a family of cycles covering each edge exactly twice. By applying advanced reasoning capabilities, deep mathematical representations, and symbolic logic, this model demonstrates autonomous scientific discovery rather than simple assistance. This milestone indicates an accelerating capability of frontier AI systems to solve abstract theoretical problems. (source: https://cdn.openai.com/pdf/04d1d1e4-bc75-476a-97cf-49055cd98d31/cdc_proof.pdf)

GPT-5.6 Sol UltraCycle Double Cover ConjectureGraph TheoryMathematical ProofAutomated ReasoningArtificial IntelligenceLarge Language ModelGenerative AI
02

AI-generated videos to maximally drive a target brain region

Researchers at EPFL developed a new methodology using AI-generated videos designed to target and maximally stimulate precise brain regions. By combining generative artificial intelligence models with computational neuroscience frameworks, the project constructs dynamic visual stimuli that elicit peak neural activity in targeted visual cortex areas. Transitioning from static image stimulation allows for more naturalistic, time-varying inputs aligned with natural biological visual processing. This approach holds significant implications for mapping brain functions with high resolution and developing advanced brain-computer interfaces. (source: https://nevo-project.epfl.ch/)

Computational NeuroscienceGenerative AISynthetic StimuliBrain MappingVisual CortexArtificial IntelligenceGenerative AIComputer Vision
03

The Annotated JEPA

Meta's AI research division, led by Yann LeCun, proposed the Joint Embedding Predictive Architecture (JEPA), which is detailed in a technical walkthrough in this article. Unlike traditional generative architectures that predict every pixel or token, JEPA learns representation models by predicting the representations of missing parts of an input. The exploration details the underlying mathematics, training objectives, and structural components of the model. It highlights how the model utilizes energy-based concepts and non-contrastive self-supervised learning to prevent representation collapse while capturing semantic features. (source: https://elonlit.com/scrivings/the-annotated-jepa/)

Joint Embedding Predictive ArchitectureSelf-Supervised LearningWorld ModelsYann LeCunEnergy-Based ModelsArtificial IntelligenceMachine LearningDeep Learning
04

Ask HN: What was the last task where only a frontier model could do it?

A Hacker News thread initiated a technical discussion comparing the real-world capabilities of open-weight LLMs with top-tier proprietary frontier models. Users shared structured case studies to identify specific failure modes where lightweight or open-source alternatives like Qwen and DeepSeek fall short, and where proprietary models such as Claude 3 Opus and GPT-4 remain necessary. The inquiry focuses on concrete task performance and debugging details to evaluate the practical limits of current AI deployment strategies. (source: https://news.ycombinator.com/item?id=48863171)

Frontier ModelsOpen-Weight ModelsModel EvaluationLarge Language ModelsBenchmarkingArtificial IntelligenceLarge Language ModelNatural Language Processing
05

How the Terrorist Group Boko Haram Uses Frontier AI

The Center for Advanced Security Policy issued a report investigating how the extremist group Boko Haram has adapted frontier artificial intelligence technologies to optimize its operations. The report details the group's use of large language models, generative AI, and advanced digital tools for propaganda generation, strategic planning, and tactical communication. By lowering the technical barrier for sophisticated digital campaigns, these highly accessible tools act as a force multiplier, highlighting the critical need for strict alignment protocols, safety research, and active monitoring by model providers. (source: https://casp.ac/reports/ai-enabled-terrorism)

Frontier AIGenerative AIAI SafetyLarge Language ModelSynthetic MediaArtificial IntelligenceLarge Language ModelGenerative AI
06

Filipino virtual assistants behind LinkedIn's "thought leadership" content mill

Rest of World published an investigative report detailing how Western executives outsource their personal branding and content creation to virtual assistants in the Philippines. These virtual assistants systematically employ structured templates, algorithmic optimization, and generative AI tools to manufacture high-engagement LinkedIn posts. The investigation exposes the mechanics behind modern professional ghostwriting, raising serious questions about authenticity on digital platforms and the reliance on cheap remote labor combined with artificial intelligence to manipulate feed algorithms. (source: https://restofworld.org/2026/virtual-assistant-linkedin-engagement/)

Virtual AssistantsDigital LaborContent ModerationSocial Media AlgorithmsGenerative AIGenerative AINatural Language ProcessingOthers

Twitter

8 stories
01

OpenAI Unveils Advanced GPT-5.6 Model With Enhanced Agent Capabilities

OpenAI has launched GPT-5.6, a new frontier model with enhanced token efficiency and native agentic capabilities deployed directly within ChatGPT. According to announcements by Sam Altman and Microsoft, the model is also integrated into Microsoft 365 Copilot as its primary reasoning engine. In benchmark testing, the Sol variant of the model achieved a 7.8% score on the ARC-AGI-3 benchmark, marking a milestone as the first verified frontier model to solve a game in this evaluation suite. (source: https://x.com/sama/status/2075577796928344329)

GPT-5.6OpenAIAI AgentsToken EfficiencyChatGPTArtificial IntelligenceLarge Language ModelAI Agent
02

Claude Code Desktop Updates With Integrated In-App Browsing Capabilities

Anthropic has released an update to its Claude Code desktop tool, introducing an integrated, sandboxed in-app browser that enables the AI agent to interact with external documentation and live websites. This update allows the agent to browse internet resources with configurable persistence settings to protect user privacy. François Chollet noted the update as part of a rapid shift in the landscape of agentic programming tools over the past six months. (source: https://x.com/ClaudeDevs/status/2075635283211772279)

Claude CodeAI Development ToolWeb BrowsingDeveloper ProductivityAnthropicArtificial IntelligenceAI AgentProduct Launch
03

Fifty Year Old Mathematical Conjecture Solved Using Sol Ultra

The GPT-5.6 Sol Ultra model has successfully generated a formal proof for a complex mathematical conjecture that had remained unsolved for fifty years. The model leverages parallel test-time compute optimization to shift from serial processing to parallel workflows, drastically reducing latency during high-level logical reasoning. This architectural update allows complex reasoning tasks that previously required an entire day to complete in approximately one hour. (source: https://x.com/gdb/status/2075670151702430044)

Sol UltraMathematicsConjectureScientific BreakthroughAI ResearchArtificial IntelligenceResearch ProgressIndustry News
04

Ai2 Releases olmOCR 2 for Advanced Text and Image Understanding

The Allen Institute for AI (Ai2) has launched olmOCR 2, an updated open-source model suite deployed on the Ai2 Playground. The model suite is optimized for high-quality text, video, and image understanding, supporting advanced multimodal document-centric visual data analysis. By providing open access, Ai2 aims to democratize advanced multimodal AI tools for researchers and developers. (source: https://x.com/Kyle_L_Wiggers/status/2075662246412239052)

olmOCRAi2Open SourceMultimodalDocument UnderstandingArtificial IntelligenceMultimodalOpen Source
05

MireloAI and Kyutai Labs Introduce Advanced Audio To MIDI Model

MireloAI and Kyutai Labs have launched a collaborative Audio-to-MIDI model designed to automate music transcription processes. The system identifies precise musical elements from complete audio recordings to output corresponding digital MIDI notation files, streamlining creative workflows for producers and musicians. (source: https://x.com/ylecun/status/2075669003326894383)

Audio-to-MIDIMireloAIKyutai LabsMusic TechnologySignal ProcessingArtificial IntelligenceMachine LearningProduct Launch
06

Google AI Studio Launches Custom URLs for Deployed Applications

Google has launched a new feature for Google AI Studio that allows developers to configure custom, permanent URLs for their deployed applications. By replacing default identifiers with personalized domains, Google aims to streamline hosting, sharing, and branding of AI-powered web applications during public deployment. (source: https://x.com/Google/status/2075597136117309659)

Google AI StudioCustom URLWeb DeploymentDeveloper ToolsProduct UpdateArtificial IntelligenceProduct LaunchIndustry News
07

AI Picbreeder Engine Launches Collaborative Multi-Agent Evolution System

The AI Picbreeder engine has introduced a multi-agent evolutionary system driven by ten Vision Language Model breeder agents operating in parallel. This architecture allows autonomous agents to collaborate, iteratively sample, and evolve complex digital assets in a shared ecosystem, simulating human creative design workflows. (source: https://x.com/hardmaru/status/2075596367087767564)

AI PicbreederMulti-Agent SystemEvolutionary ComputingVLMGenerative AIArtificial IntelligenceAI AgentGenerative AI
08

OpenAI Launches Build Week Event Featuring Developer Sessions And Challenges

OpenAI has announced the launch of its upcoming Build Week starting July 13, featuring live technical sessions and project challenges designed for the developer community. The initiative aims to support builders by facilitating collaborative project submissions and technical guidance. In preparation, OpenAI also reset usage limits for Codex and ChatGPT Work tiers. (source: https://x.com/sama/status/2075670358339080424)

OpenAIBuild WeekDeveloper CommunityTech EventsAI DevelopmentArtificial IntelligenceIndustry NewsProduct Launch

huggingface

8 stories
01

Vidu S1: A Real-Time Interactive Video Generation Model

We introduce Vidu S1, a real-time interactive video generation model supporting voice control of digital characters. Built with TurboDiffusion and TurboServe, Vidu S1 outputs 540p real-time videos at up to 42 FPS on regular consumer GPUs. Users can control video generation content at any moment through voice instructions, upload custom images, and choose different voice tones. Vidu S1 supports infinite-length real-time video generation without blurring, drift, or visual distortion, achieving strong performance across all test metrics while fully meeting real-time inference requirements. (source: https://huggingface.co/papers/2607.03118)

Interactive Video GenerationReal-Time InferenceVoice ControlTurboDiffusionDigital CharactersComputer VisionGenerative AIMultimodal
02

OpenCoF: Learning to Reason Through Video Generation

We introduce OpenCoF, a framework designed to study and improve Chain-of-Frame (CoF) reasoning in video generation models. OpenCoF features the OpenCoF-17K dataset, a reasoning video dataset spanning 11 task families, and Wan-CoF, a fine-tuned video model. Across four video reasoning benchmarks, Wan-CoF achieves considerable gains over the Wan2.2-I2V-A14B baseline. The study explores equipping models with visual and textual reasoning tokens to capture low-level cues and high-level semantic priors, demonstrating that stronger video reasoning requires broad temporal supervision and explicit intermediate state organization. (source: https://huggingface.co/papers/2607.08763)

Chain-of-Framevideo generationtemporal reasoningmultimodal learningOpenCoFMultimodalVideo UnderstandingGenerative AI
03

UniClawBench: A Universal Benchmark for Proactive Agents on Real-World Tasks

We introduce UniClawBench, a capability-driven benchmark designed to evaluate proactive agents in dynamic, real-world settings. Built around five foundational capabilities (Skill Usage, Exploration, Long-Context Reasoning, Multimodal Understanding, and Cross-Platform Coordination), the benchmark includes 400 bilingual real-world tasks evaluated in live Docker containers using fine-grained checkpoints. To disentangle base model capabilities from framework choices, we evaluate state-of-the-art models under multiple agent frameworks. We also employ a closed-loop evaluation strategy with executor, supervisor, and user agents to simulate realistic multi-turn human feedback. (source: https://huggingface.co/papers/2607.08768)

Proactive AgentsEvaluation BenchmarkLarge Language ModelsMultimodalAgent FrameworksAI AgentLarge Language ModelArtificial Intelligence
04

UP: Unbounded Positive Asymmetric Optimization for Breaking the Exploration-Stability Dilemma

We propose Unbounded Positive Asymmetric Optimization (UP), a universal, plug-and-play optimization objective designed to resolve the exploration-stability dilemma in reinforcement learning for large language models. UP restructures optimization by anchoring the policy to its current state via a stop-gradient operator. This design unleashes unclipped, stable gradients for positive advantages to maximize exploration, while maintaining standard clipping safeguards for negative advantages. Experiments demonstrate that UP enhances exploration capacity and improves reasoning accuracy across multiple RL algorithms (DAPO, GSPO, GRPO), model architectures, and training modalities. (source: https://huggingface.co/papers/2607.06987)

Reinforcement LearningLarge Language ModelsAsymmetric OptimizationExploration-Stability DilemmaImportance SamplingArtificial IntelligenceMachine LearningLarge Language
05

Flash-BoN: Instant Drafts for Inference-Time Scaling in Diffusion Models

We present Flash-BoN, an inference-time scaling strategy for text-to-image generation that accelerates candidate generation in Best-of-N (BoN) sampling. Flash-BoN generates inexpensive draft candidates by combining timestep truncation, layer skipping, and activation proxies, followed by a multi-stage verification to identify the best draft for full-quality refinement. Across three benchmarks and three model scales, Flash-BoN consistently outperforms guided search and standard BoN under fixed wall-clock budgets, with gains that increase at larger model scales (+8% AUC) and integrate well with reflection-based prompt optimization (+16% AUC). (source: https://huggingface.co/papers/2607.04461)

inference-time scalingdiffusion modelsBest-of-N samplinggenerative accelerationmulti-stage verificationGenerative AIComputer VisionDeep Learning
06

CineMobile: On-Device Image-to-Video Diffusion for Cinematic Camera Motion Generation

We propose CineMobile, a compressed image-to-video diffusion framework designed to generate cinematic camera motion on mobile devices. CineMobile uses distillation-guided pruning, diffusion distillation, reinforcement learning, and hybrid post-training quantization to compress the model footprint to under 1 GB. CineMobile achieves a 40x speedup compared to the Wan 2.1 teacher architecture. It generates 49-frame 480p videos with a per-step latency of 0.6s on an H200 GPU and 20s on the MediaTek Dimensity 8400 Ultimate 5G, utilizing a peak memory of 1.8 GB. (source: https://huggingface.co/papers/2607.03803)

Image-to-VideoDiffusion TransformersOn-Device Video GenerationModel CompressionCinematic Camera MotionDeep LearningComputer VisionGenerative AI
07

Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE

We propose Jet-Long, a tuning-free zero-shot context extension method for large language models that pairs a local RoPE-faithful window with a dynamic, long-range window. Utilizing an inclusion-exclusion attention merge and on-the-fly RoPE correction, Jet-Long incurs less than 4% single-batch generation overhead. Tested on Qwen3 models up to 128K context, Jet-Long exceeds the strongest baseline on the RULER benchmark by up to +4.79 percentage points, achieves leading accuracy on HELMET-RAG, and obtains the lowest PG-19 perplexity while generalizing to hybrid architectures like Jet-Nemotron. (source: https://huggingface.co/papers/2607.07740)

Long-context extensionDynamic Bifocal RoPEZero-shot extrapolationAttention mergeLarge Language ModelsArtificial IntelligenceLarge Language ModelNatural Language Processing
08

CausalDS: Benchmarking Causal Reasoning in Data-Science Agents

We introduce CausalDS, a benchmark evaluating causal reasoning in agentic data-science workflows. Each instance features a sampled structural causal model (SCM) with generated observational data and a grounded natural-language story. Grounded in empirical distributions to reduce the risk of "causal parroting", the benchmark tests tasks across all three of Pearl's rungs. Tasks involve a coding component using multiple tools to handle imperfect observations, evaluating symbolic causal reasoning alongside uncertainty quantification and abstention, where recognizing when a question has no warranted answer is scored directly. (source: https://huggingface.co/papers/2607.08093)

Causal ReasoningData ScienceAI AgentsBenchmarkStructural Causal ModelArtificial IntelligenceLarge Language ModelAI Agent