今天最明确的信号是 **AI 基础设施从“建模型”向“管代理”的范式转移**。Microsoft Build 2026 的 Scout、Meta 的 WhatsApp AI 代理、Coralogix 的 2 亿美元融资，都在指向同一个方向：AI 代理正在从 demo 走向生产，而管理、监控、编排这些代理将成为下一个基础设施级机会。与此同时，token 成本管控成为企业级痛点——Uber 的预算超支和 headroom、rtk 等压缩工具的火爆，说明“AI 用得起”比“AI 有多强”更迫切。最后，Nvidia RTX Spark 将 AI 推理推向笔记本端，边缘 AI 的硬件瓶颈正在被打破，开发者应开始关注本地部署的可行性。
AllNewsPapersProjects★ Top picks (4+)
📰 Industry News
Anthropic Files for IPO, Aiming for Q4 Listing
Anthropic has officially filed its IPO prospectus, planning to go public as early as Q4 2026, becoming one of the most anticipated AI IPOs.
Microsoft Build 2026 Unveils 7 Major Updates, Including In-House Reasoning Model and AI Assistant
Microsoft announced flagship reasoning model MAI-Thinking-1, AI assistant Microsoft Scout (based on OpenClaw), and a suite of AI agent tools at Build, signaling a shift away from OpenAI dependency toward in-house development.
Google Launches Gemini Omni: Clone Your AI Avatar in 15 Minutes
Google released Gemini Omni, allowing users to create their own AI digital avatar—including face cloning and voice synthesis—by scanning a QR code in under 15 minutes.
Google Launches Dreambeans: Turns Your Life Data into AI Comic Stories
Google introduced Dreambeans, a tool that extracts data from users' Google accounts to generate personalized AI-illustrated "stories."
Meta's WhatsApp Business AI Agent Goes Global, Charged by Token Usage
Meta announced the global availability of its WhatsApp Business AI agent, with businesses charged based on token usage, marking a new phase in AI customer service commercialization.
UK Regulator Mandates Google Allow Publishers to Opt Out of AI Search
The UK's Competition and Markets Authority (CMA) ruled that Google must provide tools for website publishers to opt out of AI Search features (e.g., AI Overviews), with the option to be tested in the UK before global rollout.
Alphabet's Record $85B Raise Fuels AI Business
Alphabet completed its largest-ever $85 billion stock sale, dedicated to supporting Google's AI business, signaling strong investor confidence in the AI sector.
Lovable Signs Multi-Year Renewal with Google Cloud, 5x Usage Increase
AI app-building platform Lovable signed a multi-year expansion deal with Google Cloud, growing its footprint 5x and gaining expanded access to Anthropic Claude.
Coralogix Raises $200M, Betting on AI Agent Monitoring
Infrastructure company Coralogix completed a $200M funding round, focusing on providing behavior monitoring, troubleshooting, and operational data platforms for AI agents.
Nvidia RTX Spark Laptop Chips Debut, AI PC May Hit Tipping Point
Nvidia released RTX Spark chips, poised to turn "AI PC" from concept into reality, delivering powerful local AI inference on laptops.
OpenAI Poaches Harvard's Youngest Tenured Professor, USTC Prodigy Alum
OpenAI continues to strengthen its research team, hiring Harvard's youngest-ever tenured professor (who entered university at age 12) and renowned scholar Su Weijie.
Uber Caps Employee AI Spending After Budget Exhausted in 4 Months
After encouraging staff to use AI as much as possible, Uber was forced to cap AI usage due to budget overruns within four months.
Cross-dimensional Intelligence Tops WorldArena World Model Leaderboard
Cross-dimensional Intelligence (Kuawei Zhineng) claimed the top spot on the WorldArena world model leaderboard, demonstrating progress in embodied AI and world understanding.
📄 Papers
WALL-WM: Event-Grounded World Action Model Pretraining
Proposes WALL-WM, shifting video-action learning from fixed-length chunk optimization to semantically coherent action events as the atomic unit for Vision-Language-Action pretraining, addressing granularity mismatch in existing world action models.
OmniOPD: Logit-Free On-Policy Distillation via Speculative Verification
Introduces OmniOPD, enabling on-policy distillation without accessing teacher model logits via a speculative verification mechanism, allowing closed-source models to serve as teachers.
KVarN: Variance-Normalized KV-Cache Quantization
Proposes KVarN, normalizing variance in KV-cache quantization to mitigate error accumulation during long-sequence decoding in reasoning tasks.
AURA: Action-Gated Memory for Robot Policies at Constant VRAM
Introduces AURA-Mem, an action-gated memory architecture designed for edge robots, supporting long-horizon operation at constant VRAM, addressing KV-cache's unsuitability for robotics.
Small RL Controller, Large Language Model: RL-Guided Adaptive Sampling
Formulates adaptive sampling for test-time scaling as an MDP, training a lightweight RL controller to dynamically decide when to stop sampling, improving reasoning while reducing cost.
Releases YOLO26 series, achieving NMS-free end-to-end detection, lighter heads, shorter training schedules, and solving small-object positive assignment.
ByG: Unpaired Flow Matching for Image Editing
Proposes Bootstrap Your Generator (ByG), leveraging base generative model priors to train flow matching editing models without paired data, extensible to video.
PaddleOCR-VL-1.6: Region-Aware Optimized Document Parsing Model
Baidu releases PaddleOCR-VL-1.6, identifying "under-optimized regions" (unstable model behavior, sparse data coverage) and applying targeted data augmentation and progressive post-training, significantly boosting document parsing at 0.9B parameters.
🔧 Open Source
codegraph: Pre-Indexed Code Knowledge Graph to Cut AI Coding Agent Token Usage
Provides a pre-indexed code knowledge graph for AI coding agents like Claude Code, Codex, and Gemini, reducing token consumption and tool calls, running 100% locally.
oh-my-pi: Terminal AI Coding Agent
A terminal AI coding agent supporting hash-anchored edits, optimized tool harness, LSP, Python, browser, subagents, and more.
headroom: Compress Tool Outputs and Logs, Cutting 60-95% Token Usage
Compresses tool outputs, logs, files, and RAG chunks before they reach the LLM, reducing token consumption by 60-95% while maintaining answer quality. Available as library, proxy, and MCP server.
Understand-Anything: Turn Code into Interactive Knowledge Graphs
Converts any codebase into an explorable, searchable, and queryable interactive knowledge graph, supporting Claude Code, Codex, Cursor, Copilot, Gemini CLI, and more.
taste-skill: Gives AI "Good Taste," Avoiding Generic Output
A high-agency frontend tool that prevents AI from generating boring, generic, "slop" content, improving output aesthetics and uniqueness.
paseo: Orchestrate Coding Agents Remotely from Phone, Desktop, and CLI
Allows users to remotely orchestrate and manage coding agents from phone, desktop, or CLI, enabling cross-device AI coding workflows.
rtk: CLI Proxy Reducing Token Consumption by 60-90% on Common Dev Commands
A Rust-based CLI proxy that reduces LLM token consumption by 60-90% on common development commands. Single binary, zero dependencies.
ppt-master: AI Generates Natively Editable PPTX from Any Document
AI generates native PowerPoint files from any document using real shapes (not images), requiring no design skills.
No items match this filter.
💡 Today's Take
The clearest signal today is the **paradigm shift from "building models" to "managing agents" in AI infrastructure**. Microsoft Build 2026's Scout, Meta's WhatsApp AI agent, and Coralogix's $200M funding all point in the same direction: AI agents are moving from demos to production, and managing, monitoring, and orchestrating these agents will be the next infrastructure-level opportunity. Simultaneously, token cost control has become an enterprise pain point—Uber's budget overrun and the popularity of compression tools like headroom and rtk show that "affordable AI" is more urgent than "how powerful AI is." Finally, Nvidia's RTX Spark brings AI inference to laptops, breaking edge AI hardware bottlenecks; developers should start paying attention to local deployment feasibility.