Home
Fast-tracking AI industry hot events with concise ~1000-word analysis.
The AI Agent benchmark trust crisis: SWE-bench gaming, inflated self-reported scores, and how to spot them. Third-party independent boards (Odysseys, where Browser Use ranks #1 at 87.4%) vs vendor self-reports, plus four checks for finding trustworthy benchmarks.
OpenAI's in-house inference chip Jalapeño cuts reliance on Nvidia GPUs. Benchmarked against Google TPU, Amazon Trainium, and Microsoft Maia-the compute self-build wave's impact on API pricing and developers.
On the evening of July 27, Moonshot AI open-sourced Kimi K3's weights: a 2.8-trillion-parameter MoE with a 1-million-token context, the world's largest open-source model, benchmarking against Anthropic's Fable 5. From the July 16 API launch to tonight's weight release, Yang Zhilin used ten days to compress the China-US model gap from 6-9 months to 3-5. Breakdown of specs, benchmarks, the comeback story, and the White House accusation.
Gemini 1.5 pushed context to 1M tokens in 2024; by mid-2026 "million-level" is standard for top models, and the Chinese camp is pushing tens of millions. What long context solves, what it doesn't (lost in the middle), cost, and "needle in a haystack"-plus three tips for choosing.
In October 2025 Anthropic released Computer Use, letting Claude see the screen, click the mouse, and type to operate a computer directly. Over half a year later OpenAI's Operator, Google's Mariner, and China's Zhipu AutoGLM followed; agents are moving from "talking" to "doing." Three convergence conditions and three takeaways for regular people.
ByteDance's Doubao and Alibaba's Qianwen simultaneously shut down user-built AI character features due to the "Interim Measures for AI Personified Interactive Services" taking effect July 15. Platforms cannot comprehensively review massive UGC agents, resulting in blanket shutdowns. User digital asset property rights issues highlighted.
Musk says China will likely become the AI leader; Huang says China is destined to produce excellent AI technology. Hugging Face data shows Chinese model downloads at 17.1% surpass US 15.8% for the first time. China considers AI export controls—role reversal.
Microsoft AI head: 12-18 months to reshape white-collar work. Multi-agent collaboration (Tyrion Orchestra 8 roles + persistent memory) 300% efficiency. 60% lawyer contract review AI-handled, 55K accounting jobs gone. From bricklayer to foreman, build AI legions on Coze/Dify.
847 AI agent study: 91% have holes, 94% poisonable. Tencent Zhuque vLLM CVSS 9.8 + 360/Tsinghua 20+ agent project holes. AI-generated code 40% has security defects. AI-native security (AI writes + AI audits + AI blocks) + guardrails + sandbox isolation.
Hugging Face hit by autonomous AI attack; US closed-source models' safety rails refused to assist; China's open-source GLM locally deployed completed forensics. Open source holds 62.8% market, Qwen3/Llama 4 rival GPT-4o, DeepSeek at 1/10 cost. Guardrail paradox + data sovereignty + iteration speed drive the reversal.