Summary
AI Search’s weekly roundup covers a dense slate of releases spanning open-source models, video generation, robotics, and AI agents. The headline open-source story is GLM 5.2 from Zhipu AI, described as seizing the top spot among open-weight models and capable of running locally on high-end consumer hardware. Also featured is DreamX World, an open-source world model trained on a mix of Unreal Engine gameplay and real-world video that generates explorable, persistent 3D environments from prompts — available now at 5B parameters, with a 14B version planned and the current release fitting in approximately 21GB.
On the video tooling side, Permavid introduces a dual-memory architecture — separating appearance memory from structural/geometric memory — to solve long-standing consistency problems in AI video editing, while Kling’s Omni Director enables camera motion cloning from reference videos rather than text descriptions alone. Alibaba’s new open unified science model adds understanding of proteins, small molecules, and chemical reactions to the open-source ecosystem, and the roundup also covers humanoid robot demos and Midjourney’s pivot into physical spa hotel experiences.
A standout segment covers OpenAI Codex’s new skill-learning feature on macOS: users record themselves completing a workflow once — such as uploading a YouTube video with metadata — and Codex extracts a reusable “skill” that it can execute autonomously on future sessions using computer use. The feature is currently Mac-only and unavailable in the EU, but represents a meaningful shift toward demonstration-based agent training rather than purely prompt-based instruction.
📺 Source: AI Search · Published June 21, 2026
🏷️ Format: Roundup
