Tag · 21 items
AI
Writing and projects filed under this topic.
Writing
- 10月2日AI探索速递Five high-signal developments: Pi 1.0 and durable agents, Supabase acquiring Turso for per-agent databases, Cinema 4D native MCP, PixelLeak, and regulatory accountability for autonomous agents.
- 10月3日AI探索速递Five high-signal developments for October 3, 2026: macOS tightens Full Disk Access for the agent era, OpenAI Dots meets real-world workflows, TypeSafe AI's Jev challenges generative-model defaults, DevDay tooling reaches the frontend stack, and AI companies turn to character design as an interaction layer.
- Codex_Harness_1003A first-principles reverse engineering of the Codex harness: how context, iterative tool use, execution environments, verification, recovery, safety boundaries, skills, and multi-agent infrastructure turn a language model into a practical engineering agent.
- Concept_Test-Time_Compute_1003A first-principles study of test-time compute: how reasoning systems trade inference-time computation for better answers, why search, verification and adaptive compute matter, and why future AI may be defined by how intelligently it allocates thinking.
- 10月1号AI探索速递Five high-signal developments spanning Gemini 4 Argon, adversarial model distillation, self-hosted coding agents, prompt-native Lightroom editing, and orbital AI compute.
- Concept_MoE_1002A first-principles study of Mixture of Experts: how conditional computation separates total capacity from per-token compute, why routing and load balancing become the real engineering problem, and how MoE points toward dynamically allocated intelligence.
- 9月29日AI探索速递Five high-signal developments spanning Anthropic's IPO risk disclosures, cross-model agent deception, NVIDIA agent containment, GitHub's post-chat Canvas interface, and GitHub's migration away from CSS-in-JS.
- 9月30日AI探索速递Five high-signal developments: OpenAI Dots, GPT-6.1 Sol and agent economics, DeepSeek-Huawei's Ascend software stack, AMD's World Labs acquisition, and a real agent privacy failure.
- AgentLightning_1001A systems-level study of Microsoft Research's Agent Lightning v1.0 and its Harnessed Agentic RL paradigm: why real agent harnesses should participate directly in post-training, how the framework decouples agent execution from RL infrastructure, and what this means for the future of continuously improving AI systems.
- Concept_Speculative Decoding_1001A first-principles study of speculative decoding: how cheap prediction plus exact verification converts part of autoregressive generation from a serial process into parallel work.
- MCP_1001A first-principles reverse engineering of the Model Context Protocol: why agents need a protocol layer, how hosts, clients, servers, tools, resources, transports, authorization, feedback loops, and asynchronous tasks fit together, and what MCP teaches us about designing agent infrastructure.
- 9月28号AI探索速递Five high-signal developments across agent containment, AI product interfaces, persistent-agent economics, creative AI finishing workflows, and hybrid local/cloud agent architecture.
- 9月29日每日Concept A deep research article explaining how PagedAttention solves KVCache memory management problems in large language model serving by transferring virtual memory ideas from operating systems.
- Browser Use and AgentA deep system analysis of Browser Use, exploring how AI agents connect language models with real digital environments through tools, feedback loops, and workflow architectures.
- Codex_Agent_InsightA deep reverse engineering analysis of Codex and modern AI agents, exploring how models, context, tools, memory, execution loops, and evaluation systems combine into autonomous problem-solving architectures.
- 9月24日AI探索速递Five high-signal developments: Android Studio becomes agent-agnostic, Google turns REST APIs into MCP tools, Databricks adds a governance plane for coding agents, Google reproduces OLMo 3 across training stacks, and Microsoft extends Zero Trust to AI agents.
- 9月25日AI探索速递Five high-signal developments: Microsoft reorganizes Copilot around persistent agency; Stanford challenges AI benchmark validity; Docker makes long-running agent sandboxes portable; Lovable reveals a durable agent control plane; and Next.js prepares a coordinated nine-vulnerability security release.
- 9月26日AI探索速递Five high-signal developments across agent safety, local AI, ambient interfaces, realtime voice systems, and sandbox security.
- 9月27日AI探索速递Five high-signal developments across frontier-agent safety, AG-UI, Claude Opus 5.5, Google Vids, and human-agent research workflows.
- 9月24日AI探索速递Five high-signal developments: Cursor closes the loop from code generation to production verification; Microsoft challenges onboard-only robot inference; Google Vids pushes generative video into timeline workflows; HP and Google revisit spatial telepresence; and Palo Alto Networks turns frontier models into continuous security infrastructure.
- 9月21~23日AI探索速递整理 2026 年 9 月 21–23 日最值得保留的 12 条 AI 与技术新闻：从 Agent Harness、Prompt Caching、AGENTS.md，到 Robotics Skills、Local/Sovereign AI 与 Generative Workflow。重点不是新闻本身，而是它们共同揭示的 AI Systems Engineering 趋势。