AI Infrastructure

2026 年 Jul 26 日
AI engineering

Synthetic Data and Pre-Training at Scale: poolside’s Lessons

Poolside's synthetic data pipeline pairs templates with supplementary context, while the training team crashes on weight mismatch to catch silent failures at scale during pre-training.

2026 年 Jul 26 日
AI engineering

How one power line failure exposed a growing AI data center grid problem

A power line failure near D.C. caused 3 GW of data centers to disconnect from the PJM grid simultaneously, flickering lights and highlighting a growing grid stability problem.

2026 年 Jul 26 日
AI engineering

Co-operative time-slicing for RL post-training in llm-d

Co-operative time-slicing lets RL jobs interleave on shared GPUs, raising duty cycles from 40% to 70% without hurting convergence, by checkpointing idle phases to DRAM and swapping in active ones.

2026 年 Jul 26 日
Business

Debate on restricting Chinese open-weight AI models

Open-weight models offer cheaper intelligence than frontier labs, threatening returns on massive investments, but advocates argue they accelerate innovation and U.S. leadership.

2026 年 Jul 26 日
Tools

AI on Your Lakehouse: Context Comes in Shapes, Not Queries

Zach Blumenfeld shows how three graph shapes — trees, communities, and paths — give lakehouse agents context they can't get from queries alone.

2026 年 Jul 26 日
AI engineering

Cloud Run Multi-Region Services: Automatic Failover with Readiness Probes

Cloud Run adds readiness probes and service health for automatic multi-region failover within seconds, with no additional cost beyond standard probe compute.

2026 年 Jul 26 日
Business

Panasonic vSkipGen validated on Google Cloud Axion C4A-metal for SDV

Panasonic's vSkipGen on Google Cloud's C4A-metal creates digital twins for cockpit software, using crosvm and VirtIO to run AAOS images without physical prototypes.