Google DeepMind
·
4 months ago
AlphaGo defeated world champion Lee Sae Dol at Go in 2016, marking the beginning of modern artificial intelligence and demonstrating that AI could discover novel strategies beyond human expertise. The system's techniques have since been applied to protein folding (AlphaFold 2 solved a 50-year challenge and now serves over 3 million researchers), mathematical reasoning (achieving silver-medal performance at the International Mathematical Olympiad in 2024), and algorithm discovery (AlphaEvolve found novel matrix multiplication methods). These applications are now converging toward artificial general intelligence, with DeepMind integrating AlphaGo's search and planning methods into its Gemini models alongside specialized scientific tools.
Import AI
·
4 months ago
Researchers from GovAI and the University of Oxford proposed 14 metrics to measure AI R&D automation—the degree to which AI systems are building other AI systems—as a prerequisite for governing recursive self-improvement. The metrics include tracking AI performance on R&D tasks relative to humans, measuring oversight effectiveness, and monitoring compute distribution across AI development. These measurements would help companies, governments, and third parties understand and control the pace of AI progress before it becomes difficult to oversee.
OpenAI Blog
·
4 months ago
OpenAI is acquiring Promptfoo, a platform that tests AI systems for security vulnerabilities during development. Promptfoo provides automated scanning and remediation tools for enterprise customers building AI applications. The acquisition will integrate Promptfoo's testing capabilities into OpenAI's product offerings for developers.
Last Week in AI
·
4 months ago
Anthropic received official notification from the Department of Defense designating it a supply chain risk, while OpenAI faced growing calls for a boycott after signing a military contract. A newsletter covering recent AI industry developments reported these two major events involving AI companies and government relationships. The designation and boycott trend signal increasing scrutiny of AI firms' relationships with military and defense sectors.
Hugging Face Blog
·
4 months ago
LeRobot released v0.5.0 with support for the Unitree G1 humanoid robot, new vision-language-action policies including Pi0-FAST and Wall-X, and faster dataset processing. The update added over 200 merged pull requests from more than 50 new contributors, expanded robot hardware support to include Earth Rover and OMX Robot, and achieved 10x faster image training and 3x faster video encoding. The modernized codebase now requires Python 3.12 and Transformers v5, enables real-time video encoding during data collection, and allows custom policies to be registered as installable packages.
Hugging Face Blog
·
4 months ago
Researchers developed Ulysses Sequence Parallelism, a technique that distributes attention computation across multiple GPUs using attention head partitioning to enable training on million-token sequences. The method requires two all-to-all communication operations per attention layer with communication volume of O(n·d/P) per GPU, compared to Ring Attention's O(n·d)—a factor of P times more data. The technique has been integrated into Hugging Face's Accelerate, Transformers Trainer, and TRL's SFTTrainer, allowing standard training workflows to handle sequences of 32,768 tokens or longer without exceeding single-GPU memory limits.
