Zhipu AI has open-sourced the GLM-4.7 large model, achieving state-of-the-art (SOTA) level in programming capabilities among open-source models. It scored 73.8% on SWE-bench, 84.9% on LiveCodeBench V6, and surpassed GPT-5.2 in Code Arena blind tests. The model introduces three new thinking modes to enhance stability for complex tasks, along with improvements in front-end generation and tool calling capabilities. The API and full code are now open.

The MiniMax M2.1 model excels in multi-programming language generation, mobile development, and office automation scenarios. Its test results surpass Claude Sonnet 4.5. The company has simultaneously open-sourced the VIBE full-stack application evaluation benchmark to promote the implementation of AI applications for real-world complex tasks.

Fun-Audio-Chat 8B is an end-to-end voice dialogue model that bypasses traditional ASR+LLM+TTS pipelines, resulting in lower latency and higher efficiency. It possesses emotion perception and Speech Function Call capabilities, allowing completion of complex tasks through natural speech. The model weights and code are fully open-sourced.

Based on Berkeley research and DeepMind experiments, 68% of Agents are limited to fewer than 10 steps. Multi-agent systems suffer from a coordination tax and error amplification; increasing budgets does not linearly improve performance. Breakthroughs require systematic evolution in tool management, verification capabilities, and communication protocols.

A retrospective of an enterprise-level front-end AI Agent project revealed technical breakthroughs but poor product adoption. The 80/20 bottleneck and user habit resistance led to no usage. The lesson learned is to integrate general Skills into existing toolchains rather than building standalone Agents, emphasizing deep integration with developer workflows.

The article systematically explains the two core concepts of MCP (Model Context Protocol) and Agent Skills, noting MCP addresses connectivity while Skills encapsulate domain knowledge and operational processes. Introducing a progressive disclosure mechanism for Skills effectively mitigates context explosion. It proposes that a layered MCP+Skills hybrid architecture is a crucial direction for Agent development.

The latest LangChain survey indicates that by 2026, AI Agents have entered the practical implementation phase, with 57% of enterprises officially deploying them in production. Customer service and R&D analysis are core scenarios. The biggest implementation barrier is output quality, not cost. Key technology trends include observability tracing and multi-model hybrid architectures.

The AI workplace co-pilot survey jointly released by Lenny and Figma, based on 1750 samples, shows over half of practitioners save at least half a day per week due to AI, with entrepreneurs benefiting the most and designers perceiving the least impact. Engineers are shifting from Copilot to Cursor and Claude Code. The AI opportunity is moving from content production towards strategic thinking.

Notion founder Ivan Zhao interprets the AI revolution from a historical perspective, likening AI to the "infinite brain" following steam and steel. He analyzes its profound impact on individuals, organizations, and economies, proposing to stop viewing AI merely as a co-pilot and instead reimagine entire work methodologies.

Three Google DeepMind Gemini leads shared the stage, revealing that the Flash model has reached or even surpassed the level of the previous Pro generation, with Pro's main role shifting to distilling Flash. Post-training offers the greatest breakthrough potential, while latency and speed are severely undervalued. Future priorities are open-ended tasks and continuous learning capabilities.