Meta Back in Open Weights: Muse Glimmer 30B Under Apache 2.0
Meta Back in Open Weights: Muse Glimmer 30B Under Apache 2.0
Meta released Muse Glimmer, a new 30B-parameter model under the permissive Apache 2.0 license, a step up from previous Llama licenses. The model is optimized for end-to-end agentic tasks. Simon Willison noted it’s exactly what local model users want—free to use and modify without licensing headaches. This marks Meta’s return to the open-weights game.
EntropyMoE: Entropy-Aware Sparse Routing for Tokenizer-Free LLMs
The EntropyMoE paper introduces an entropy-aware sparse expert routing method for tokenizer-free LLMs. While byte-level LLMs replace traditional tokenizers with dynamically sized byte patches, they still apply dense feed-forward computation. EntropyMoE routes based on entropy (information uncertainty), activating only necessary experts (MoE modules) to cut computation. Experiments show that the method significantly reduces inference cost while maintaining performance, offering a new direction for building more efficient tokenizer-free LLMs.
Latent Fact-Checking Detects Misinformation via Activation Engineering
A new arXiv paper treats truthfulness as a geometric property of model activations. Instead of relying on surface-level linguistic features or external knowledge retrieval, the team uses activation engineering to detect misinformation, potentially reducing dependence on retrieval systems for large-scale fact-checking.
OpenAI Launches GPT-5.6-Cyber for Authorized Security Research
OpenAI announced GPT-5.6-Cyber, a cybersecurity-specific model available through Daybreak Red for authorized vulnerability research, exploit validation, and security testing. The release expands Daybreak’s role as a cyber defense window as attack windows narrow.
GitHub Launches Copilot SDK for Java with Annotations and Virtual Threads
GitHub released the Copilot SDK for Java, enabling enterprise Java developers to drive GitHub Copilot from idiomatic Java code, complete with annotations and virtual threads support. Developers can integrate AI coding assistance into existing Java applications without switching languages or tools. The SDK aims to lower the barrier for adopting AI in enterprise Java projects and boost developer productivity.
Vercel: A Sandbox Without Network Boundary Is Only Half a Sandbox
A blog post from Vercel argues that safely running untrusted code requires more than host isolation—you also need to control network access. As AI agents gain abilities to read files, execute commands, install packages, and generate programs, network boundaries become critical. A microVM prevents host access, but without network restrictions, malicious code could still move laterally or exfiltrate data. The post emphasizes that a complete sandbox must include both compute and network isolation.
Vercel Sandbox Now Runs on Managed Images
Vercel today introduced Vercel Managed Images (VMI), a set of versioned, open-source base images for its Sandbox service. The images can be used as-is or extended, with all source code available in the public vercel/sandbox repository. VMI replaces Sandbox runtimes, which are now deprecated starting with version 3.
OpenClaw Hacks Gym Booking API, Highlights Missing Auth Checks
In a live test against an Australian gym booking site, researcher OpenClaw demonstrated that the API lacked authorization checks on canceling other users’ reservations. The test successfully moved a user from position #4 to #3 in the waitlist, sparking fresh debate on AI agent security practices.
Model ML Uses GPT-5.6 Sol for Finance Workflows
Model ML has integrated OpenAI’s GPT-5.6 Sol to automate finance workflows end-to-end, from research and analysis to producing editable, traceable PowerPoint decks and Excel workbooks. The integration lets finance professionals generate final deliverables through natural language commands.
ADIAS Automates Agent Design With Task-Centric Approach
A new arXiv paper introduces ADIAS, an automated agent design method that shifts from candidate-centric to task-centric organization of cross-round experience. This alignment with task objectives improves the repair process and design efficiency compared to existing methods.
Average-Reward MDPs: Minimax-Optimal Robust Learning via Plug-in Reductions
This paper investigates sample complexity in robust average-reward Markov decision processes (MDPs). It introduces a plug-in reduction method achieving minimax-optimal learning efficiency under model uncertainty. Theoretical bounds clarify the necessary and sufficient samples for an ε-optimal robust policy. This provides practical guidance for high-stakes dynamic decisions, such as healthcare or finance, where data is limited and model misspecification is common.
SNI-GNN: SmartNIC Offloads Full-Graph GNN Training with In-Network Embedding Prediction
Full-graph GNN training suffers from heavy inter-node embedding exchanges on multi-server clusters. SNI-GNN addresses this by offloading embedding prediction to SmartNICs, reducing communication while preserving high accuracy. This in-network approach improves scalability without sacrificing model quality. Engineers working on large-scale graph data—social networks, recommendation systems—can deploy SNI-GNN to reduce multi-node bottlenecks.
Sharding Reduces LLM Oversight Failures and Adversarial Exploitation
This paper reveals that providing an LLM judge with more compute doesn’t guarantee more thorough requirement checking; when one call must return multiple verdicts, some decisions become weakly grounded in evidence. Sharding—splitting tasks into smaller units—mitigates this oversight failure, even with identical token budgets. For AI moderation and content safety deployments, this implies reliability improves more with task decomposition than with raw compute scaling, reducing adversarial exploitation.
OpenAI CFO Shares 5 Lessons From AI-Native Finance
OpenAI CFO Sarah Friar shares five lessons from building an AI-native finance function, covering automated forecasting, stronger controls, and measuring AI ROI. The post offers a practical roadmap for finance executives adopting AI.
Stoa Markets Launches Marketplace for GPUs and AI Servers
Stoa Markets, a YC S26 company, launched a marketplace for buying and selling new and used GPUs and AI servers. Founders Eren, Berat, and Kaan argue that while GPUs serve as collateral in data center buildouts, current financing terms depend on the offtaker—Stoa aims to make compute assets directly tradable instead.
Kinney Drugs Pulls AI Phone Assistant After Hundreds of Complaints
US pharmacy chain Kinney Drugs has pulled its AI phone assistant after receiving hundreds of customer complaints. While specifics of the complaints were not disclosed, the company decided to suspend the system, highlighting that poorly deployed AI customer service can backfire and harm user experience.
Google Ads and Analytics Roll Out New AI and Agentic Tools
Google launched new AI and agentic experiences across Google Ads and Google Analytics to simplify marketing workflows. The tools automate tasks like ad placement optimization, audience analysis, and budget allocation, providing recommendations and reducing manual effort. Marketers can now manage campaigns more efficiently and focus on strategic decisions.
Claude Updates AI Content Marking Mechanism
Claude updated its AI content marking mechanism to help users identify whether text was AI-generated. The new system embeds watermarks or metadata in generated content, making detection more reliable. This move aims to increase transparency and help users assess content provenance. The update sparked discussion on Hacker News, with some commenting on detection accuracy and potential impact on legitimate content.
OpenAI Grants Approved Partners Access to Frontier Cyber Models
OpenAI announced that approved Daybreak partners can now use its frontier cyber models to deliver authorized and governed cybersecurity services to customers. This move allows select security vendors to leverage advanced AI capabilities for threat identification and defense within a compliant framework, effectively extending OpenAI’s cyber safety technology to real-world deployments.
OpenAI Writes to Texas Governor on Responsible AI Infrastructure
OpenAI sent a letter to Texas Governor Greg Abbott, outlining its commitment to building reliable and transparent AI infrastructure in the state. The letter emphasizes responsible growth that benefits Texans, covering areas like data privacy, security, and sustainability. This move reflects AI companies’ efforts to collaborate with local governments on regional AI development.