Jason Says
Two signals worth watching today: OpenCut baking MCP server into an open-source video editor, and ByteDance's world model team quietly pivoting to autonomous driving—both point to the same trend: MCP is becoming the new infrastructure layer for all software, and world model companies are racing to find killer applications.
OpenCut, the open-source CapCut alternative, is being rewritten with a plugin-first architecture featuring a built-in MCP server for AI agent control, headless batch rendering mode, and a unified Rust core for web/desktop/mobile. A significant step toward automating video editing workflows with AI agents.
HKUDS open-sourced Vibe-Trading, a personal trading agent framework offering API/MCP integration, shadow account simulation, and one-command setup for full trading capabilities. Built by HKU researchers, it's ready infrastructure for financial AI agents—though the team has issued warnings about impersonator accounts and fake token contracts.
Bindu Reddy (ex-Perplexity CEO) is building AutoBots—fully autonomous, cost-optimized self-improving agents that use a mixture of 6 LLMs to continuously iterate toward goals with zero human-in-the-loop. Signals a shift from AI as tool to AI as autonomous executor.
Meta announced a $40B additional investment in its Louisiana data center—its largest single infrastructure project to date. As the AI compute arms race intensifies, this signals that hyperscaler infrastructure spending shows no signs of peaking in 2026.
36Kr exclusive: ByteDance's Seed world model team, led by Zhou Chang, is quietly exploring autonomous driving with a focus on unmanned logistics. ByteDance officially denies commercial plans but doesn't deny early research. Given world models' overlap with AV tech, this could reshape China's embodied AI landscape.
ABot-AgentOS introduces a middleware layer above low-level robot controllers, providing planning, skill isolation, multi-stage verification, and lifelong multimodal memory. For developers, it bridges the gap between powerful VLMs and real-world robot deployment—think of it as an 'OS' that makes robots actually usable in long-horizon tasks.
NeuroCogMap applies cognitive neuroscience methods to map reproducible functional systems inside LLMs, revealing how different cognitive capabilities are organized internally. For developers, it offers a principled framework for understanding model failure modes and could unlock more targeted fine-tuning and interpretability techniques.
European startups raised $24B in Q2 2026—up ~33% QoQ and 66% YoY, marking the continent's strongest venture quarter in four years. AI was the primary driver, with the UK leading gains. For founders targeting European markets, the funding climate is warming significantly.
AI驱动的大学转学流程自动化平台，用AI替代繁琐的后台录取工作流，切入全球规模庞大的高等教育行政市场。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free