Anthropic Opens "Mythos 5" to 100+ US Organizations — From Negotiation to Approval

On June 26, the Trump administration approved access to Anthropic's state-of-the-art AI model "Mythos 5" for more than 100 US companies and government agencies. In a letter addressed to Anthropic's Chief Compute Officer, Commerce Secretary Howard Lutnick stated that "Anthropic has worked with the US government to address risks related to Covered Models. These efforts have brought significant progress," clearing the model for release to "certain trusted partners."

According to multiple outlets reporting the same day, the scope also includes non-US employees. TechCrunch reported that "more than 100 companies and government agencies are authorized to use Mythos 5."

The move comes roughly two weeks after a June 12 export-control directive halted availability worldwide (including nationality filtering on AWS Bedrock, Google Cloud, and Microsoft Foundry). It marks a clear step forward from the previous "under discussion" status to "approved and released," and it's fair to say the relationship between AI companies and regulators has reached a milestone. That said, we'll have to wait for future official announcements on the conditions and oversight framework under which the model will be provided.

Predicting Sudden Death Risk with AI — A New Approach from UC Berkeley

A research team at UC Berkeley has announced a new method for detecting the risk of sudden cardiac death using AI. The study, which also gained traction on Hacker News, suggests that AI can identify patterns that existing risk assessments have struggled to capture.

Sudden cardiac death claims many lives each year, yet predicting it in advance has long been considered extremely difficult. By extracting subtle abnormal patterns from data such as electrocardiograms, AI could improve the accuracy of identifying high-risk groups. As for the detailed methodology and the timeline for clinical application, we await the publication of the paper.

BlackBerry CEO: "Safety-Certified Software Won't Be Replaced by AI"

In a June 26 Bloomberg Television interview, BlackBerry CEO John Giamatteo said that, at this point, AI cannot replace the company's safety-certified products.

The company's QNX operating system is currently embedded in 275 million vehicles worldwide, and it is expanding beyond the automotive sector into industrial automation, robotics, and medical devices. Giamatteo also noted that QNX is BlackBerry's fastest-growing division.

In domains where safety directly affects human lives, the inherent uncertainty of AI is hard to accept. This is a telling example of how traditionally safety-engineered software remains strongly competitive in edge domains that demand certification and assurance.

Masayoshi Son Signals Interest in Investing in TEPCO — "Power" Is the Key to Attracting Data Centers

According to ITmedia, SoftBank Group's Masayoshi Son addressed "securing power" — the biggest bottleneck for AI infrastructure — at the company's shareholders meeting on June 26. He also touched on the fact that subsidiary SoftBank has put itself forward as a candidate to become the next owner of Tokyo Electric Power Company (TEPCO), and spoke of a long-term vision of "net asset value of ¥1,000 trillion."

Bringing state-of-the-art data centers back to Japan requires a stable, large-scale power supply. If a partnership with TEPCO materializes, it could strengthen the competitiveness of Japan's domestic AI infrastructure. However, any acquisition of or investment in TEPCO faces significant hurdles, including regulatory approval and political consensus, so the concrete shape and timeline remain unclear at this point.

LocalLLaMA Topic: Vulkan Build of llama.cpp Records 2x ROCm Speeds

An interesting configuration using the Vulkan build of llama.cpp on an AMD Radeon RX 7900 XTX was reported in Reddit's LocalLLaMA community.

The setup runs Qwen3.6-35B-A3B (IQ4_XS quantization) with a 262K-token (roughly 260,000 tokens) context length, with VRAM usage fitting within about 22 GiB. Most notably, it achieved roughly 2x the token generation speed compared to an optimized ROCm 7.14 setup, with lower memory usage and footprint as well. The Vulkan backend has long been treated as experimental, but this suggests it may be reaching production-level readiness.

Other topics shared in the community included troubleshooting results for P2P communication in a bifurcation setup with 4x RTX 5060 Ti, and voices considering a shift from cloud models to in-house distillation.