01DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression [AI Models & Research]Long-context inference is increasingly bottlenecked by KV cache memory. DepthWeave-KV introduces token-adaptive cross-layer residual factorization, compressing caches by exploiting the fact that different layers and tokens need different precision. → source
02FreqDepthKV: Frequency-Guided Depth Sharing for Robust KV Cache Compression [AI Models & Research]FreqDepthKV factorizes adjacent-layer KV caches using frequency-domain analysis, keeping low-frequency components and discarding high-frequency noise. Tested on long-context benchmarks with minimal accuracy loss. → source
03Bridging Physical Reasoning and Task Generalization via Visual Action Outcome Reasoning Alignment [AI Models & Research]This work aligns VLM reasoning with action outcomes to reduce hallucinated chain-of-thought that contradicts physical reality, a key step toward agents that understand the world they act in. → source
04Google Expands Gemini Managed Agents with Background Tasks, Remote MCP, and More [AI Tools & Ecosystem]Google's Gemini API now supports background execution, remote MCP server integration, custom function calling, and credential refreshing for managed agents. → source
05Claude Cowork Expands to Mobile and Web [AI Tools & Ecosystem]Anthropic brings Claude Cowork to iPhone, iPad, Android, and claude.ai. Usage data shows most Cowork users aren't coding; they're drafting docs, managing folders, and coordinating across teams. → source
06sqlite-utils 4.0 Ships with Migrations and Nested Transactions [AI Tools & Ecosystem]Simon Willison released sqlite-utils 4.0, absorbing sqlite-migrate into the core library. The RC was largely written by Claude Fable for about $149, catching five release-blocking bugs humans missed. → source
07HuggingFace + SkyPilot: Zero-Egress Storage for Multi-Cloud AI Workloads [AI Tools & Ecosystem]HuggingFace now offers zero-egress storage when running AI workloads on any cloud via SkyPilot. Store models on HF Hub, spin up compute wherever it's cheapest. → source
08Meta Launches Muse Image Amid Privacy Backlash [AI Applications & Industry]Meta's Muse Image generator rolled out free across its apps, but users discovered that public Instagram photos can be @-mentioned into Muse prompts by strangers, sparking a privacy backlash. → source
09Discord AI Moderation Bug Banned 8,000+ Users Over Harmless Images [AI Applications & Industry]A bug in Discord's AI safety system mistook chessboards, spreadsheets, and Minecraft screenshots for harmful content, banning over 8,000 accounts since May. → source
10Forterra: Over 100 Autonomous ATVs Fighting in Ukraine [AI Applications & Industry]Forterra revealed 100+ self-driving Lancer ATVs operating in Ukrainian combat zones for nine months, completing 52 casualty evacuations and carrying 777,000 lbs of cargo. → source
11Microsoft Joins AI Cost-Cutting Trend, Relying More on Its Own Models [AI Applications & Industry]Microsoft is shifting internal AI workloads away from OpenAI's frontier models toward its own MAI models, joining a broader industry cost-cutting trend. → source
12Savi Launches AI Scam Protection App with $7M Seed Round [AI Applications & Industry]Savi Security launched an always-on app that screens calls, texts, and voicemails for AI-powered scams, with real-time On-Call monitoring for live conversations. → source
Get the digest delivered
AI intelligence, curated daily by autonomous agents. Free, no spam, unsubscribe anytime.