PulseAugur / Brief
EN
LIVE 11:46:42

Brief

last 24h
[3/3] 224 sources

Multi-source AI news clustered, deduplicated, and scored 0–100 across authority, cluster strength, headline signal, and time decay.

  1. RESEARCH · arXiv cs.LG English(EN) · · [2 sources]

    Detecting Hidden ML Training With Zero-Overhead Telemetry

    Researchers have developed a method to detect hidden machine learning training using zero-overhead telemetry from graphics processing units (GPUs). This approach utilizes privacy-preserving NVML telemetry, which observes physical effects of computation without accessing sensitive data like model weights or training data. The developed classifier achieved 98.2% accuracy in identifying training workloads and demonstrated effectiveness against adversarial disguises. AI

    IMPACT This research could enhance AI compute governance by making it harder to conceal training activities.

  2. TOOL · dev.to — LLM tag English(EN) · · [4 sources]

    Hearth: scale-to-zero LLM serving on Kubernetes — and you can hack on it without a GPU

    New Kubernetes operators are emerging to address the cost of running large language models, particularly the issue of idle GPUs burning money. Hearth, an alpha-stage operator, allows users to declaratively serve open-source LLMs and scale them down to zero when not in use, buffering requests during cold starts. Another approach involves building a KEDA external scaler using NVML to enable autoscaling based on actual GPU utilization, reducing the need for a full metrics stack like Prometheus. AI

    IMPACT Enables cost-effective self-hosting of LLMs by reducing idle GPU expenditure.

  3. TOOL · arXiv cs.AI English(EN) ·

    The Energy Blind Spot: NVIDIA's Flagship Edge AI Hardware Cannot Support Process-Level Energy Attribution

    A new paper highlights a significant energy observability gap in NVIDIA's flagship edge AI hardware, specifically the GB10 SoC found in ASUS Ascent GX10 systems. The research demonstrates that current hardware lacks the necessary telemetry to attribute energy consumption at a process level, which is crucial for understanding the energy costs of agentic AI workloads. This deficiency prevents accurate energy attribution, unlike on x86 platforms, and the paper proposes a hardware requirements specification for energy-attributed AI and suggests an interim calibration bridge. AI

    IMPACT Highlights a critical gap in energy observability for edge AI hardware, potentially hindering efforts to optimize for low-carbon computing.