Preferred Networks Develops Generative AI System to Support Self-Defense Force Operational Planning

It has come to light that Preferred Networks (PFN) has been awarded a demonstration project by Japan's Acquisition, Technology & Logistics Agency (ATLA) to develop a system that uses generative AI to support operational planning for the Japan Self-Defense Forces. The case is drawing attention as an example of a Japanese AI company making a full-fledged entry into the defense sector.

The phrase "domestically built AI" is being emphasized, apparently reflecting a broader push toward greater reliance on homegrown technology for national security reasons. Details of the demonstration and any deployment timeline remain unknown at this point, but AI adoption in defense is becoming a very real topic in Japan as well.

Sam Altman: "People Don't Want AI to Serve as CEO"

OpenAI's Sam Altman shared his views on the role of AI and the future of human work. His remark that "people don't want AI to behave like a CEO" suggests a recognition that there are clear limits to the areas where AI can substitute for human judgment.

Altman's statement can also be read as a kind of line drawn against the assumption that AI will replace every job. The message appears to be that in roles requiring strategic decisions and setting an organization's direction, human accountability and contextual understanding remain essential. Meanwhile, how extensively AI is used in a supporting role continues to evolve alongside the technology.

LLM-Assisted Labeling Does Not Improve Security Classification — A CVE-ATT&CK Mapping Study

A study featured on Hugging Face Daily Papers reported that in classifiers mapping CVEs (Common Vulnerabilities and Exposures) to MITRE ATT&CK techniques, **automatic labeling by LLMs underperforms manual annotation by experts**.

The research team trained a multi-label classifier on a dataset of 1,207 expert-curated CVEs. LLM-generated labels matched expert annotations at a rate of only about 0.39, and expanding the dataset with roughly 1,000 additional CVEs reduced coverage of rare techniques. What proved decisive was that the apparent improvement was an artifact caused by evaluation noise. Under the correct evaluation protocol, the effect of LLM augmentation was statistically zero.

This result matters as a concrete example showing that AI labeling capability does not always outperform human expertise. It once again underscores a fundamental principle of machine learning: data quality matters more than data quantity.

Open-Source Self-Hosted AI Gateway "ilter" Released

"ilter", published on GitHub, is a self-hostable AI gateway tool. It provides integrated support for the MCP protocol, budget management, PII (personally identifiable information) masking, smart routing, and fallback capabilities.

For developers working with multiple LLM providers, it is a practical tool for centrally managing API keys, cost control, and security requirements. As AI usage expands within enterprises, it is drawing attention as an approach that maintains governance while avoiding provider lock-in.

It has not gained much traction on Hacker News yet, but it looks like a tool that precisely addresses practical needs.

Local Music Generation with Stable Audio 3.0 Medium

A hands-on article on local AI music generation using Stable Audio 3.0 Medium was published on Qiita. It covers composing music with an open-weights music generation AI model in a high-end GPU environment.

While music generation AI is less widely known than image or video generation, technically it is said to be "reaching a high standard." The article explains the process in concrete detail, from setting up the execution environment through to generation, making it a useful reference for developers who want to try AI music generation in a local environment.

---