AI / ML

Onur Satici introduces Vortex, a Linux Foundation open-source columnar format designed to stream data directly from S3 to GPUs. By leveraging cascading lightweight encodings, layout-based segment pruning, and zero-copy memory pipelines, the system bypasses traditional CPU and NVMe bottlenecks. This approach achieves streaming speeds up to 60 Gbps without requiring upfront data reprocessing.

• Vortex enables direct S3-to-GPU streaming, removing intermediate CPU/NVMe bottlenecks.
• Zero-copy memory pipelines and lightweight encodings drive throughput up to 60 Gbps.
• Layout-based segment pruning allows efficient data access without upfront reprocessing.
• Open-source format under Linux Foundation simplifies adoption for ML training pipelines.

Cohere has released Parse 5, a 2.3-billion-parameter multimodal foundation model focused on extracting structured data from complex enterprise documents. The system processes visually rich PDFs into Markdown format and provides bounding box coordinates for visual grounding. It achieved an average score of 79.2 across key performance areas when evaluated against over 2,000 enterprise pages.

• Parse 5 converts visually rich PDFs into structured Markdown output.
• Includes bounding box coordinates for precise visual grounding.
• Evaluated on 2,000+ enterprise pages with a 79.2 average score.
• 2.3B parameter model targets complex enterprise document workflows.

The $12.9 billion acquisition of Hugging Face by Nvidia raises significant concerns about cementing the chipmaker's market dominance. Critics argue the deal would harm competition within the AI ecosystem. Regulators are urged to examine the potential anti-competitive effects before approving the transaction.

• Nvidia's $12.9B acquisition of Hugging Face triggers antitrust concerns.
• Deal risks cementing Nvidia's dominance in the AI hardware and software stack.
• Regulators should scrutinize the impact on market competition.
• The move could limit choices for developers relying on open-model infrastructure.

Google Research has released TimesFM 3.0, a pretrained decoder-only foundation model designed specifically for time-series forecasting. The model is available via PyTorch checkpoints and is integrated into Google’s ecosystem, including BigQuery ML and Google Sheets. This release follows previous versions and aims to provide scalable, enterprise-grade forecasting capabilities.

• Decoder-only architecture optimized for time-series prediction tasks.
• New 3.0 checkpoint available for immediate PyTorch integration.
• Native support in BigQuery ML allows SQL-based forecasting at scale.
• Accessible via Google Sheets for rapid prototyping and daily use.