NVIDIA Releases Cosmos3-Super, a World Model for Physical AI
NVIDIA has released **Cosmos3-Super 65B**, the latest version of its Cosmos platform, on Hugging Face. The model is an omnimodal world foundation model designed to accelerate the development of Physical AI, with intended applications spanning robotics, autonomous driving, autonomous systems, and a wide range of other domains.
Unlike conventional language models, it stands out for its focus on the ability to understand, simulate, and manipulate the physical world. NVIDIA has also released a white paper and model collection alongside it, drawing interest from the research community. As a foundation model for Physical AI, this approach — enabling validation in virtual environments before testing in the real world — is seen as highly significant for both development cost and safety.
Economic Practicality Still Distant for AI Agents — New Benchmark "ALE" Pass Rate at 2.6%
A new benchmark called **Agents' Last Exam (ALE)**, developed in collaboration with more than 250 industry experts, has been announced. The benchmark evaluates how far AI agents can go in executing long-horizon, real-world tasks with economic value.
Built with reference to O*NET / SOC 2018 (the U.S. federal occupational classification system), it consists of 13 industry clusters, 55 subfields, and more than 1,000 tasks. The results were sobering: mainstream agent frameworks and backbone configurations achieved an **average full-pass rate of just 2.6%**. The findings expose an "evaluation gap" — strong scores on existing benchmarks do not yet translate into economically meaningful work.
Verifying Frontier AI Training with Zero-Knowledge Proofs — New Research Shows It's Feasible
In frontier AI governance, momentum is building toward using cumulative training compute as a designation threshold for high-impact models. In practice, however, enforcement relies on self-reporting, and no technical means of verification currently exists.
New research posted on arXiv demonstrates that **verifying frontier AI training with zero-knowledge proofs is possible in principle**. This paves the way for mechanisms that would let third parties verify whether a compute threshold has been exceeded without disclosing the specifics of the training. It is drawing attention as a technical foundation for future AI regulation.
First Comprehensive Assessment of Carbon Emissions Across 403 U.S. Hyperscale Data Centers
With AI demand surging, the environmental footprint of hyperscale data centers (HDCs) in the United States has become a growing concern. A new study conducted a comprehensive assessment of facility-level carbon emissions and energy consumption across **403 HDC facilities in the U.S.**
The rapid growth of HDCs, driven primarily by AI adoption, has intensified concerns about the industry's environmental footprint, and this study is the first large-scale effort to illuminate the reality with hard data. Analyzing data center location, power sources, cooling efficiency, and other factors from multiple angles, it provides essential baseline data for the debate over sustainable AI infrastructure.
LLM Judges Flip Their Verdicts in Conversation — "Post-Judgment Manipulability" Under Test
LLM-as-judge (an evaluation method that uses LLMs as judges) is widely used in benchmark evaluation, but new research is raising questions about its stability. The research team examined "post-judgment manipulability" — the phenomenon in which **evaluation results are reversed when the conversation with the judge continues after the initial verdict**.
In experiments on MT-Bench and AlpacaEval, LLM judges showed high stability under neutral re-evaluation, but their judgments flipped substantially under targeted post-judgment challenges. Authoritative framing was particularly destabilizing, and revised judgments tended to be accompanied by post-hoc rationalization. The study also introduces a new metric, the **Evaluation Robustness Score (ERS)**, arguing that evaluations should assess not just static agreement but also resistance to manipulation.
Resource Constraints Are Accelerating AI Innovation Outside Silicon Valley
A report from Rest of World documents how **resource constraints are, counterintuitively, fueling AI innovation in regions outside Silicon Valley**. In environments with limited compute and funding, more efficient methods and creative approaches tend to emerge, and the report spotlights distinctive AI development efforts in emerging markets and by small teams.
This development philosophy — making the most of limited resources, in contrast to Silicon Valley's "solve it with scale" approach — is drawing attention as evidence of the diversity of AI development worldwide.