Reproducing OLMo 3 7B on TPUs reveals hidden data loader bugs
A Google case study reproduces AI2's OLMo 3 model on TPUs, matching performance metrics while uncovering critical data sharding errors that masked memorization as improvement.
Results for “go”
Daily coverage of AI, developer tools and infrastructure. Each story explains what happened and why it matters, with a link to the original source.
Google demonstrates autonomous LLM fine-tuning loops on TPUs
Google Developers Blog details autofinetune, an agent-driven system that automates hyperparameter tuning for SFT and RL post-training on Cloud TPUs.
Behavioral evaluations offer clearer insights for AI coding agents
Google Developers Blog outlines how behavioral evaluations provide actionable feedback for AI agent development, moving beyond opaque end-to-end benchmark scores.
Four engineering patterns from top AI agent submissions
Google analyzed winning entries from its 2026 AI Agents Challenge to identify four reusable engineering patterns for building robust multi-agent systems.
Optimizing long-context embedding inference on Cloud TPU with vLLM
Google details how native TPU support in vLLM enables elastic scaling and high-precision embedding inference for Qwen3 models.
Optimizing HeyGen's Avatar IV diffusion model for Google Cloud TPUs
HeyGen and Google Cloud engineers detail how they ported the Avatar IV video generation pipeline to Trillium v6e TPUs, achieving a 1.86x speedup through kernel optimization and parallelism strategies.
Running Gemma LLMs on Raspberry Pi 5 with LiteRT
Google's 2026 guide details using LiteRT to run Gemma models on Raspberry Pi 5, achieving real-time inference for robotics and edge agents without cloud dependency.
How controller-runtime cache prevents API server overload
A deep dive into the list-watch pattern and local caching in Kubernetes controllers, explaining why reads are cheap but consistency is eventual.
Efficient data cloning in Kubernetes using external volume snapshots
A 2021 guide explains how to bypass namespace restrictions by importing cloud provider snapshots as golden images for fast, isolated development environments.
