Nvidia to Acquire HuggingFace for $13B; OpenAI Publishes Incident RetrospectiveHot Chips 2025: OpenAI's Jalapeño, Cerebras CS-5, Groq 3 LPX, Apple M6 DebutExFold: Training-Free MoE Inference Acceleration via Expert FoldingGoogle DeepMind Pilots World's First Double-Blind AI EvaluationsOpenAI Study: ChatGPT and Critical-Thinking Training Boost Assignment OriginalityClaude Managed Agents Now Work with Chat SDK for Persistent Slack BotsNew Study: Evidence Frontloading and Pressure-Adaptive Budgeting Relieve RAG BottlenecksMolEmb: Multimodal LLMs Prove Effective as Molecular Embedding ModelsFirst Real-Hardware Study of Masked Diffusion LLMs Offers Key Design PrinciplesDataKernelBench: Benchmarking LLMs for Optimizing GPU Database QueriesSelfGraphRAG: Bridging Supervision Gap in Graph RAG with Synthetic QAGoogle Search AI Mode Adds Hotel Booking, Airfare Tracking, and Miles ManagementGoogle Releases Gemini Omni 1.1 Flash with Enhanced Granular ControlsOpenClaw (formerly Open Interpreter) Becomes Fastest-Growing GitHub Project, Maintainers Share Security InsightsVercel CEO: The Best Workflow Engine Is a Programming LanguageThe Teaser Period: Why the AI Boom Is Hitting a Reset WallAnthropic Makes Claude Code Auto Mode Default; Security Expert Raises ConcernsDeveloper Says Heavy AI Coding Tool Use Is 'Killing My Brain'Low-Code and No-Code Tools Make AI Agents Accessible to EveryoneNvidia Projects $673B in Sales as AI Demand SurgesNvidia to Acquire HuggingFace for $13B; OpenAI Publishes Incident Retrospective

At the Hot Chips conference, OpenAI unveiled its Jalapeño chip, Cerebras announced the CS-5, Groq introduced the 3 LPX, and Apple showed off its new M6. Each targets different performance and efficiency niches, signaling intense competition in AI hardware.

A new arXiv paper introduces ExFold, a training-free acceleration method for Mixture-of-Experts (MoE) models. By unifying expert folding, it optimizes both prefill and decode phases without altering model weights, offering a direct way to reduce serving latency for existing MoE models.

Google DeepMind has initiated a pilot for the world’s first double-blind AI evaluations. By concealing model identity, this approach aims to reduce bias and measure AI capabilities more objectively, potentially setting a new standard for AI benchmarking.

OpenAI released a randomized study involving over 1,000 students, examining ChatGPT, critical-thinking training, originality, and student performance on a real-world university assignment. The findings suggest that using ChatGPT in conjunction with critical-thinking training can improve the breadth of thinking and originality in student work. This provides empirical evidence for effective AI tool application in educational settings.

Vercel Blog announced that Claude Managed Agents are now integrated with Chat SDK. Managed Agents handle the complete agent loop server-side, including model calls, tool execution, session state, and sandboxed web search. This allows developers to build a Slack research bot with a persistent session without managing complex underlying infrastructure.

A new study addresses efficiency bottlenecks in Retrieval-Augmented Generation (RAG) systems. Unlike existing methods that optimize downstream LLM generation, this approach introduces ‘Evidence Frontloading’ to load key evidence early and ‘Pressure-Adaptive Budgeting’ to dynamically allocate resources, effectively relieving the bottleneck between retrieval and generation.

A new arXiv paper introduces MolEmb, demonstrating that multimodal large language models can serve as powerful molecular embedding models. This approach bypasses traditional specialized training, offering a new foundation for drug discovery, property prediction, and virtual screening.

An arXiv paper characterizes masked diffusion language models (dLLMs) on real hardware, in a first-of-its-kind study. It identifies key serving bottlenecks and offers concrete design principles for building efficient, real-world dLLM serving systems.

A new benchmark, DataKernelBench, evaluates LLMs’ ability to optimize GPU database query kernels. Existing LLM benchmarks focus on ML operators, ignoring the irregular, heterogeneous, data-movement-heavy kernels in database scenarios. This benchmark fills that gap, providing a new standard for assessing LLM potential in database performance optimization.

SelfGraphRAG is a new approach addressing the supervision gap in graph-based RAG. Existing methods underuse relational structures in knowledge graphs. By automatically generating synthetic QA pairs, SelfGraphRAG trains graph RAG models without manual annotation, significantly improving entity-relationship capture and retrieval accuracy.

Google AI Blog introduced 3 new capabilities for AI Mode in Google Search: direct hotel booking, airfare price tracking, and viewing/managing frequent flyer miles and rewards. These features aim to position AI Mode as a one-stop travel planning tool, allowing users to complete the entire journey from search to booking without switching between multiple websites and apps.

Google DeepMind’s blog announced the release of Gemini Omni 1.1 Flash, a new model version emphasizing enhanced granular controls for developers. By adjusting model output parameters and logic, developers can more precisely guide model behavior to meet various application needs. This update aims to improve the customizability and reliability of Gemini models in complex tasks.

The GitHub Blog interviewed maintainers of the OpenClaw project, including Peter Steinberger. OpenClaw has become the fastest-growing project in GitHub history. In its first six months, maintainers shared lessons learned about building, scaling, and securing the project. The article details how the project handled the influx of users and potential security risks after going viral.

Vercel CEO Guillermo Rauch argues that traditional workflow engines (message queues, job runners, microservice choreographies) have poor developer experience for orchestrating long-running stateful logic. He advocates using a programming language itself as the workflow engine, leveraging native state and error handling to reduce abstraction overhead and build reliable systems more efficiently.

An analysis argues the AI industry is in a ‘Teaser Period,’ where marketing promises far exceed actual product delivery. Many AI projects remain at demo stage, struggling to scale, causing growth to hit a ‘reset wall.’ The next phase will favor real engineering, scalability, and user value over mere parameter-count competitions.

Simon Willison highlights that Anthropic has made Auto Mode in Claude Code the default and has made bold claims about its effectiveness against prompt injection attacks. However, renowned security researcher Johann Rehberger has raised doubts. The default enablement means all users will operate in this mode, and its security will need to withstand more real-world attack scrutiny.

A Hacker News user shared concerns that after a year of using AI coding tools like Claude Code under pressure, his independent thinking and code review skills are deteriorating. He admitted to accepting AI output more readily, fearing a loss of core programming abilities.

Ben’s Bites argues that with the launch of new tools and platforms, the barrier to creating and deploying AI agents is rapidly lowering. These tools simplify complex agent logic into visual configurations and simple interfaces, enabling users without deep programming backgrounds to build automated workflows and smart applications. This means AI agents are moving from a developer-only niche to the mass market.

Nvidia has projected sales of $673 billion, driven by surging demand for AI hardware. This forecast exceeding market expectations reinforces Nvidia’s dominant position in the AI chip market and sparks discussion on supply chains and industry landscape.