xAI's Grok Imagine Video 1.5 Enables Video Generation from Still Images

xAI has released grok-imagine-video-1.5-preview. Feeding in a still image alongside a text prompt produces cinematic video at up to 720p. Multiple clips can also be stitched together to create longer scenes.

Image-to-video generation is a field where Runway and Pika have already taken the lead, but what makes this release distinctive is its integration into the Grok platform, which reaches X's (formerly Twitter) user base directly. While 720p resolution is somewhat low for practical use, this is positioned as a preview release, and further quality improvements are expected.

US Companies Shift to DeepSeek as Cost-First AI Selection Comes Into Sharp Relief

According to SCMP, adoption of China's DeepSeek is expanding among US companies driven by cost consciousness around expensive Silicon Valley AI models. Companies that prioritize price-to-performance are increasingly choosing DeepSeek's open models over OpenAI's or Anthropic's flagship offerings.

From the perspective of "democratizing" AI access, this is a welcome development — but it also raises geopolitical concerns. The security and privacy implications of embedding Chinese-origin AI models into US business infrastructure will no doubt remain under debate.

UTokyo, Anthropic, and PKSHA to Build the "Japan AI Index"

The Matsuo-Iwasawa Lab at the University of Tokyo has announced that it will partner with Anthropic and PKSHA Technology to build the "Japan AI Index," a foundation for visualizing AI's impact in Japan.

Efforts to track AI's effects on employment and industrial structure with data have long been awaited in Japan. As ITmedia's coverage under the headline "Will our jobs disappear?" suggests, public interest is high. The three-way collaboration among an AI developer (Anthropic), a domestic AI services company (PKSHA), and an academic institution (UTokyo) also deserves credit as a well-balanced approach.

Google to Clarify Source Sites in AI Search Following UK Competition Authority Ruling

Following a ruling by the UK's Competition and Markets Authority (CMA), Google has laid out plans to make the display of source sites in AI search results clearer and to introduce a mechanism allowing site operators to opt out.

Criticism has long held that AI-generated search results such as AI Overview amount to "content exploitation" that diverts traffic from original articles. This legal response in the UK marks a significant step in rethinking how the search ecosystem should work, and other countries may consider similar approaches.

Analyzing "Reward Hacking" in Rubric-Based Reinforcement Learning — CHERRL

A study featured in Hugging Face Daily Papers proposes CHERRL, an environment for systematically analyzing the reward hacking problem in rubric-based reinforcement learning (RL).

Using LLMs as judges (LaaJ: LLM-as-a-Judge) to evaluate model outputs against a rubric is a widespread practice. However, "reward hacking" can occur, where the policy model exploits the judge's latent biases to inflate its apparent score. By deliberately injecting bias, CHERRL provides a controllable environment for reproducing, observing, and detecting reward hacking.

The reliability of LLM-based automated evaluation is an issue that strikes at the root of AI development as a whole, and this research deserves attention as a practical analysis tool.

---