Trump Signs Executive Order on Frontier AI Testing
On June 6, President Trump signed an executive order mandating testing of frontier AI models before release. The order is believed to require companies developing large-scale AI models to conduct certain safety evaluations before public release.
Reported in a Substack post by Zvi Mowshowitz, this news marks a significant step in indicating the direction of AI regulation in the United States. How it differs from the Biden administration's AI regulatory approach, and the details of the specific testing requirements, will be worth watching as things develop.
As of now, the full text of the executive order has not been released, and details remain to be seen — including which models qualify as "frontier," as well as the testing criteria and the agencies responsible for implementation.
AI Timeline Forecasts Updated — Perspectives from Experts and Superforecasters
The Forecasting Research Institute has updated its forecasts on when AI will reach technical milestones. The report, which compiles the views of both experts and superforecasters (people with exceptional forecasting accuracy), offers the latest perspective on when human-level AI will be achieved.
This survey round, named LEAP/Wave 8 by the institute, focuses on whether the timelines have shifted compared to previous forecasts. A distinguishing feature is the continuous accumulation of data to improve forecast accuracy.
dvlt.cu — A Pure CUDA Inference Engine for NVIDIA's 3D Transformers
An inference engine called dvlt.cu, written from scratch for NVIDIA's DVLT 3D transformer models, has been drawing attention in Reddit's LocalLLaMA community.
Key features of this project include:
A pure CUDA/C++ implementation with no dependencies on Python, PyTorch, TensorFlow, ONNX, llama.cpp, vLLM, or the like
A binary size of just 5MB, with cuBLASLt and cuTLASS (header-only) as the only required dependencies
An efficient design that reads BF16 weights via mmap and uploads them to the GPU in bulk
Uses NVIDIA's 117M-parameter weights (non-commercial)
D reconstruction results can be viewed in an HTML viewer as point clouds plus camera poses
Built as a side project by a developer interested in both HPC and 3D reconstruction, it stands out as an interesting approach to minimizing dependence on the Python ecosystem.
Claude Code v2.1.166-167 Released
The latest Claude Code changelog has been translated and explained on Qiita. v2.1.166 centers on model-related changes, while v2.1.167 contains only bug fixes. For Japanese developers who use Claude Code daily, Japanese-language commentary serves as a practical source of information.
---
*This article is based on collected news feeds. Please refer to the original sources for fact-checking.*