China's Orca: A World Model That Handles Robotics Tasks Without Action Labels

The Beijing Academy of Artificial Intelligence (BAAI) has released a new world model called Orca. Orca's distinguishing feature is that it predicts "abstract world states" rather than tokens or pixels.

It was trained on 125,000 hours of video data without action labels (ground-truth data for operations), and reportedly matched the performance of the specialized π0.5 across five robotics tasks. It is drawing attention as a step toward solving the "real-data shortage" that the robotics field faces.

Ghost Font, a Typeface AI "Cannot Read," Gains Traction

Ghost Font, released by MixFont, is a font that humans can read but AI (OCR and text recognition) cannot. It has racked up 50 points and 46 comments on Hacker News.

It is attracting interest as a countermeasure against AI-driven automated scraping and data collection, and as a new approach to content protection. By showing text on the web to humans while keeping it unreadable to crawlers, this approach exploits a new "asymmetry of visibility" in the AI era, pointing to one possible direction for future debates over content protection.

Microsoft's Emissions Rise 25% Amid AI Data Center Expansion

According to Microsoft's latest report, its emissions have increased by 25% due to the expansion of AI data centers. The company previously relied on carbon offset credits (environmental credits), but it appears to have reversed course and will now focus on direct emissions reductions.

This is a notable example of how the rapid buildout of AI infrastructure translates directly into environmental impact, made visible through a big tech company's financial reporting. As a hidden cost of the generative AI boom, it could become a challenge for the entire industry going forward.

Nobel Laureate Chemist Omar Yaghi Heads to Tsinghua University's AI Materials Lab

Omar Yaghi, known for his pioneering work on MOFs (metal-organic frameworks) and a recipient of the Nobel Prize in Chemistry, has reportedly moved to Tsinghua University to lead an AI materials science lab.

Efforts to apply AI to materials discovery are accelerating, and the move of a top scientist to a Chinese university could also shift the geographic balance of AI research.

A Japanese Field Test: Comparing Code Review Accuracy Across Seven AI Models

An article comparing the code review performance of seven AI models, including GLM and DeepSeek, under identical conditions has been published on Qiita. Each model reviewed three files, and an audit with Codex revealed a wide gap in accuracy — anywhere from 52.1% down to 8%.

By asking "how many real bugs can each model actually find" rather than comparing spec-sheet numbers, this practical evaluation provides useful reference data for choosing models in real-world settings.

---