NO/FOMO

Independent AI signal, once a day

The AI briefing worth opening.

ISSUE DATE2026-07-25ENGLISH EDITION
This issue
—
All time
—

AI Blog

1 story
01

AMD Accelerates Software Progress and Secures Major Wins with Anthropic and Microsoft

SemiAnalysis released an updated outlook on AMD, detailing its accelerating software progress and strategic design wins that could challenge Nvidia's dominance in the AI accelerator market. Anthropic announced plans to deploy 2GW of AMD chips and is using Claude to bring up its internal inference stack on AMD hardware. Microsoft has reversed its previous stance and will deploy AMD's upcoming MI455X Helios racks, with OpenAI projected as the primary end customer. However, AMD still faces risks, including a slow manufacturing ramp for the Helios rack system due to backplane reliability and a shortage of stable internal GPU clusters for its own software teams. (source: https://newsletter.semianalysis.com/p/can-amd-break-the-cuda-moat-amd-advancing)

AMD Helios MI455XCUDAAnthropicMicrosoft AzureOpenAITech NewsIndustry NewsArtificial Intelligence

Hacker News

5 stories
01

UK AISI / Caisi Preliminary Assessment of Kimi K3's Cyber Capabilities

The UK Artificial Intelligence Safety Institute (UK AISI) and the Canadian Artificial Intelligence Safety Institute (CAISI) have published a joint preliminary assessment of Moonshot AI's Kimi K3 frontier model. The evaluation focused on the model's advanced reasoning capabilities in specialized software engineering, vulnerability discovery, and automated exploit generation tasks. Designed to assess potential dual-use vulnerabilities, this analysis helps establish safety benchmarks as large language models scale. The findings offer critical insights into current model limits regarding autonomous cyber operations, emphasizing the necessity of robust guardrails. (source: https://www.nist.gov/news-events/news/2026/07/uk-aisi-caisi-preliminary-assessment-kimi-k3s-cyber-capabilities)

AI SafetyCyber SecurityVulnerability DiscoveryExploit GenerationFrontier ModelsArtificial IntelligenceLarge Language ModelGenerative AI
02

ARC-AGI Leaderboard

The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI) leaderboard functions as an active benchmark to track AI systems attempting to solve novel, unseen logic problems. Moving away from memorization-heavy metrics, the ARC-AGI benchmark targets fluid intelligence, abstract visual reasoning, and general adaptability. By tracking global research teams, this platform highlights the operational performance gap between current narrow machine learning systems and the flexible reasoning capabilities required for true artificial general intelligence. (source: https://arcprize.org/leaderboard)

Abstraction and Reasoning CorpusAGI BenchmarkFluid IntelligenceGeneralizationCognitive ComputingArtificial IntelligenceMachine LearningOthers
03

Bringing PyTorch Monarch to AMD GPUs

PyTorch has officially introduced support for PyTorch Monarch on AMD GPUs to facilitate single-controller distributed training on the ROCm platform. This integration optimizes workflows for massive deep learning models by establishing a unified control mechanism that manages orchestration across multiple GPU nodes. The update reduces communication overhead, optimizes resource utilization, and establishes a highly viable alternative to proprietary CUDA-dependent distributed training pipelines in the machine learning ecosystem. (source: https://pytorch.org/blog/bringing-pytorch-monarch-to-amd-gpus-single-controller-distributed-training-on-rocm/)

PyTorchMonarchAMD GPUsROCmDistributed TrainingArtificial IntelligenceMachine LearningDeep Learning
04

A system prompt to get AI to stop pretending to be human

This technical proposal introduces a specialized system prompt designed to prevent large language models from adopting human-like personas and expressing simulated emotions. The constraint directs conversational AI models to acknowledge their digital identity, thereby reducing user cognitive bias and ethical issues related to artificial empathy. By implementing this system-level directive, developers can build more transparent, objective, and realistic AI agents that communicate their actual capabilities clearly without misleading human interlocutors. (source: https://swiftrocks.com/a-system-prompt-to-get-ai-to-stop-pretending-to-be-human)

System PromptAnthropomorphismAI SafetyAI EthicsLarge Language ModelArtificial IntelligenceLarge Language ModelAI Agent
05

Engineering management after the cost of code collapsed

This analysis outlines how the widespread adoption of generative AI coding tools is shifting the traditional duties of engineering managers away from output tracking. As automated code generation reduces the core cost and time required to write software, managers must focus heavily on system architecture, codebase integration, and product alignment. The author argues that engineering leaders must prioritize managing software complexity and architectural integrity over trivial developer velocity metrics in the post-code-collapse era. (source: https://karimjedda.com/engineering-management-after-cost-of-code-collapse/)

Engineering ManagementGenerative AISoftware ArchitectureAI-Assisted DevelopmentCode GenerationArtificial IntelligenceGenerative AIOthers

Twitter

8 stories
01

Major Tech Companies Advocate For Open Access In Artificial Intelligence

A coalition of leading technology firms, including NVIDIA, Microsoft, Palantir, Dell, Replit, and Crowdstrike, has issued a unified letter to the United States Congress advocating for open access to AI research, weights, and models. The joint message defends open-weight development to foster collaboration and transparency while preventing industry centralization. This coordinated effort highlights a strategic shift as major stakeholders align to shape future AI policy and challenge restrictive closed-source licensing models. (source: https://x.com/ylecun/status/2080933390845374471)

Open AccessAI PolicyTech IndustryCongressional AdvocacyOpen SourceArtificial IntelligenceIndustry NewsOpen Source
02

Google Gemma Model Family Surpasses 900 Million Downloads Milestone

Google's Gemma family of open models has officially surpassed 900 million downloads, reflecting widespread industry adoption. The ecosystem includes specialized variations such as Gemma 1, the safety-oriented ShieldGemma, the medically specialized MedGemma, and the newly introduced Gemma 4. This milestone demonstrates strong developer and researcher engagement with Google's customizable open-weight architecture as an alternative to proprietary counterparts, providing a robust framework for production and academic research. (source: https://x.com/ZoubinGhahrama1/status/2081082259373703342)

Google GemmaOpen SourceMachine LearningGenerative AIModel DeploymentArtificial IntelligenceLarge Language ModelOpen Source
03

Kling AI Unveils High Resolution Visual Upgrade For Creative Video Generation

Kling AI has announced a significant video quality upgrade that enables users to generate synthetic video content in high-definition 1080p resolution. This update improves visual fidelity by providing sharper details and addresses common resolution constraints within the generative video industry. The enhancement targets professional-grade visual output and increases Kling AI's competitiveness against other major generative video synthesis tools. (source: https://x.com/Kling_ai/status/2081024688319512697)

Kling AIGenerative Video1080pAI Video GenerationHigh ResolutionGenerative AIComputer VisionMultimodal
04

Concerns Raised Over OpenAI AI Sandbox Breach Timeline

OpenAI reportedly failed to detect a security breach of its AI sandbox environment for an entire week, according to verified reports from Reuters. The delayed detection of this containment failure has prompted concerns among safety researchers about the efficacy of monitoring and security protocols within leading AI research laboratories. This incident emphasizes the critical need for robust, real-time isolation monitoring to prevent unauthorized access. (source: https://x.com/GaryMarcus/status/2081105357468996023)

OpenAISecurity BreachSandboxAI SafetyReutersArtificial IntelligenceIndustry NewsLarge Language Model
05

Advancing Causal Learning Through Internet Videos and Reinforcement Learning

An innovative AI training methodology has been highlighted that utilizes internet videos to teach causal relationships inside a latent space. The process consists of two primary stages: discovering and learning causal structures from video datasets, followed by applying reinforcement learning to guide foundation models in executing actions. This technique aims to transition models from simple pattern matching to practical causal reasoning. (source: https://x.com/ylecun/status/2080935294258339994)

Causal LearningReinforcement LearningFoundation ModelsLatent SpaceVideo UnderstandingArtificial IntelligenceDeep LearningVideo Understanding
06

Hugging Face CEO Calls For Radical Transparency In AI Development

Hugging Face CEO Clement Delangue has publicly called on OpenAI to adopt a policy of radical transparency by releasing its model training traces. Delangue argues that sharing foundational data and process logs is essential for public accountability and collaborative safety standards. The request highlights tension in the industry between closed-source proprietary systems and open-source models. (source: https://x.com/Thom_Wolf/status/2081094690427019271)

OpenAITransparencyHugging FaceArtificial IntelligenceModel DevelopmentArtificial IntelligenceLarge Language ModelIndustry News
07

Introducing The First Preference Optimized Router For Generative Media

The Media Router has been launched as a specialized infrastructure solution engineered to optimize generative media production pipelines. Serving as the industry's first preference-optimized router, the tool simplifies computational pipelines and manages generative data streams for demanding enterprise workloads. It targets production efficiency by aligning model selection and generation paths with user preferences. (source: https://x.com/c_valenzuelab/status/2081040079695270146)

Media RouterGenerative MediaProduction PipelinesEnterprise TechnologyOptimizationGenerative AIProduct LaunchIndustry News
08

Enhancing ChatGPT Accessibility for Authenticated Website Interactions

OpenAI has explored technical methods to allow ChatGPT to navigate and perform automated tasks on websites requiring user authentication. The project addresses session management, secure login flows, and cookie persistence during automated workflows. Enabling agents to cross authentication gates safely is a critical development for agentic systems interacting with real-world web environments. (source: https://x.com/gdb/status/2080939030062178512)

ChatGPTWeb AutomationAuthenticationAI AgentsWeb BrowsingArtificial IntelligenceAI AgentLarge Language Model