Google introduces autonomous LLM post-training on TPUs

What is autonomous post-training?

The "autofinetune" project shared on the Google Developers Blog proposes a research loop that runs large language model (LLM) post-training without human intervention. It covers Supervised Fine-Tuning (SFT) and Reinforcement Learning via GRPO.

How it works

A developer defines boundary conditions and evaluation metrics in a single Markdown specification. An AI agent then takes over:

• Iteratively edits training scripts,
• Launches experiments,
• Automatically commits verified hyperparameter optimizations to Git.

This removes manual tuning cycles.

What infrastructure it uses

The system is built on Google's AI stack: Tunix, Gemma models, and Cloud TPUs. The project demonstrates hands-off performance gains in both function calling and math reasoning models.

What it means

By automating much of the fine-tuning process, the approach lets developers focus on experiment design and evaluation. However, the project is still a research effort, and how far it will move into production remains unclear.