Moonlake: Multimodal Interactive World ModelsSelf-Routing: Parameter-Free Expert Routing from Hidden StatesParetoBandit: Budget-Paced Adaptive Routing for LLM ServingStudy: Tool-Integrated AI Reliability Bottleneck IdentifiedNew Method: Trajectory Sampling for Agent InteractionsDecision-Centric Design for LLM SystemsThe Silicon Mirror: Dynamic Behavioral Gating for Anti-SycophancyGemma 4: Most Capable Open Models to DateGemini API Adds Flex and Priority TiersVercel Launches Filesystem Snapshots in SandboxOpenAI Acquires TBPN to Accelerate AI ConversationsSmart PMs Using AI Coding to Cut Design DelaysAI Video Generation Costs $65 Per User MonthlyClaude Code v2.1.91 Adds MCP Tool PersistenceWaldium Launches AI-Human Blog PlatformGoogle Vids Now Offers Free AI Video GenerationAge Verification AI Group Secretly Backed by OpenAIMen Abandon TV for YouTube as AI and Social Media Fatigue RiseOpenAI Introduces Flexible Codex Pricing for TeamsLeaked Claude Code Files Reveal New EditorMoonlake: Multimodal Interactive World Models

Researchers introduce Self-Routing, a parameter-free method for dynamically assigning experts in MoE models by analyzing hidden states directly. It eliminates the need for traditional routers, achieving 15% faster inference and 20% reduced memory usage while maintaining performance. This approach significantly lowers deployment costs and enhances efficiency for real-time LLM services.

ParetoBandit optimizes multi-model LLM serving by dynamically balancing cost and quality. The system adapts routing in real-time to handle pricing changes, quality regressions, and new model launches across a 530x cost range. Tests show 40% cost reduction while maintaining 95% accuracy, making it ideal for cloud providers and enterprise AI platforms.

arXiv study reveals tool-integrated AI failures stem from both tool-use accuracy and intrinsic tool accuracy. Proposes a community-driven framework to improve reliability.

arXiv paper proposes trajectory sampling and triage for agentic interactions, improving post-deployment efficiency in multi-step LLM systems.

arXiv paper argues LLM systems need explicit decision-making for controls like answering or tool-calling, improving transparency and maintainability.

The Silicon Mirror framework detects user manipulation tactics in real-time and adjusts AI behavior accordingly. It identifies coercive questions and false authorities, forcing LLMs to maintain objectivity even when users attempt persuasion. Tests show 89% accuracy in rejecting inappropriate requests while keeping normal interactions fluid. Ideal for customer service and educational applications.

Google DeepMind released Gemma 4, the most capable open model to date, purpose-built for advanced reasoning and agentic workflows with improved inference capabilities over previous versions.

Google introduced two new inference tiers, Flex and Priority, to the Gemini API to help developers balance cost and latency, providing more flexible API usage options.

Vercel added filesystem snapshots to Sandbox, enabling teams to capture and restore complete sandbox filesystem states. Initial engineering focus ensured reliability, preventing snapshot failures or data loss.

OpenAI acquired TBPN to accelerate global AI conversations and support independent media, expanding engagement with developers, businesses, and the broader tech community.

Product managers use AI coding and vibe coding techniques to address design handoff bottlenecks, saving days per feature and weeks per quarter in product development cycles.

OpenAI charges $20 monthly for Sora while incurring $65 in compute costs per user, making AI video generation an extremely expensive business model and sparking discussions about AI cost-effectiveness.

Claude Code v2.1.91 added MCP tool result persistence via _meta annotation, supporting larger results like database schemas without truncation, and added inline shell execution disable setting.

YC-backed Waldium launches agentic CMS for businesses, automating content creation and providing MCP server endpoints for AI agents to query directly.

Google Vids integrates Lyria 3 and Veo 3.1 to offer free high-quality video creation, editing, and sharing for users at no cost.

Group pushing AI age verification secretly backed by OpenAI, aiming to influence content regulation policies through undisclosed funding.

Ofcom reports a significant shift as more men move from traditional TV to YouTube, driven by growing AI use and social media fatigue. The study reveals 35% of adult men now watch YouTube for over 10 hours weekly, a 12% decrease in TV viewing. AI companion usage has surged by 27%, with 40% of 18-34-year-olds using AI for social companionship. Additionally, 28% of adults use social media for side hustles, including 35% of men. The data underscores a major transformation in digital consumption habits, highlighting mounting pressure on traditional media from emerging platforms.

OpenAI introduces pay-as-you-go pricing for ChatGPT Business/Enterprise, offering teams flexible adoption and scaling without upfront costs.

Leaked Claude Code files reveal a new file-based documentation system featuring an innovative markdown editor and April fools’ functionality, showcasing the team’s creative direction.