On April 3rd, a team from NYU and Meta FAIR published a paper titled "Hierarchical Planning with Latent World Models (HWM)". Original paper link: (https://arxiv.org/abs/2604.03208). This paper shifts its focus away from generating more realistic future frames, turning instead to a long-standing execution challenge for world models. Once the task chain is extended, prediction errors accumulate continuously, and the action search space expands rapidly.

The mainstream view holds that the more parameters a model has, the closer it gets to human-like thinking. However, a paper published by a Zhejiang University team in Nature Communications on April 1st presents a different perspective. They found that as model scale (primarily SimCLR, CLIP, DINOv2) increases, the ability to recognize specific objects does continue to improve, but the ability to understand abstract concepts not only fails to improve but can even decline.

The leak of Claude Code this time stemmed from the npm package publication. Anthropic inadvertently included the cli.js.map file. External parties used this file to extract over 512,000 lines of code. It contained over forty tool modules, indicating that many tasks are already being executed by more specialized tools. There was also a 46,000-line QueryEngine, which acts as the central dispatcher, specifically designed to schedule tasks. Furthermore, the code revealed several previously undisclosed features, such as Buddy, Kairos, and Auto Mode.