Great report from Google and MIT.

Co-designing ML models and the accelerators is an exciting research direction.

This report introduces Coco, an agent platform TPU architects use for hardware and model co-design. Each decision depends on hundreds of gigabytes of fresh simulation sweeps that no model saw during pretraining.

Coco registers every sweep in a normalized relational database, so each number an agent reports comes from a SQL query. Agents compose typed tools on their own, run recurring analyses such as comparing systems at matched execution configurations, and read the UI's navigation state as context.

The near-term goal is halving the time architects spend setting up simulations and getting insights from them.

Paper: https://arxiv.org/abs/2610.02376

Chat with Paper: https://academy.dair.ai/papers/coco-an-agentic-copilot-for-the-hardware-software-co-design-lifecycle-2610.02376