Google has described autofinetune, an experimental setup that lets an AI agent repeatedly adjust and run language-model post-training jobs. The September 11 developer post combines Tunix, Gemma and Cloud TPUs with an agent-driven research loop.
The idea is pleasingly workshop-like: change a setting, run an experiment, measure the result, then keep the change or roll it back. A human supplies the instructions, constraints and evaluation criteria. In Google’s supervised fine-tuning example, the agent could adjust training settings but could not change the dataset or model architecture.
The public repository contains separate supervised fine-tuning and reinforcement-learning examples, plus sample runs. It describes the project as an experiment and flags a workaround for an Antigravity CLI problem on TPU virtual machines. This is still tinkering territory.
Our take: the useful promise is less repetitive experiment management. It is not evidence that a model has become broadly smarter, or that an unattended research loop can choose a worthwhile objective for you. Geeknewz has not reproduced the experiments; reported improvements remain the authors’ results.