Nvidia has conducted a successful technical experiment on the automation of advanced training. For this task, they used the advanced AI agent Codex. The object of the study was the Cosmos 3 Nano model. The accuracy of this neural network on the four-choice test has increased significantly. The indicator increased from 54.41% to an impressive 93.35%. The whole process took less than one day. It was artificial intelligence that performed the main stages of complex retraining. The agent acted strictly according to the TAO system's pre-prepared instructions. The experiment was conducted on a specialized Woven Traffic Safety dataset. This dataset is developed by Woven by Toyota. It contains videos of various traffic situations. It also has four-choice questions. More than eight thousand examples were used for full-fledged training and subsequent verification.

Initially, without additional adaptation, the Cosmos 3 Nano model produced modest results. She answered only 54.41% of the questions correctly. The developers then instructed the Codex agent to evaluate the basic version. He was also given the task to carry out additional training using the effective LoRA method. This method does not change all the parameters of the huge model. It trains only small additional adapters. According to calculations by Nvidia specialists, this approach required seven times less resources. This is a significant saving of GPU hours compared to full retraining. The first run of LoRA took only about thirty minutes. The calculations were performed on eight NVIDIA A100 accelerators with 80 GB of memory. The accuracy of the model immediately increased to 87.14%. The agent independently selected the instructions, checked the annotations and found an error. He found a missing frame rate parameter in the configuration. After the correction, the agent loaded the weights and launched the training container.

The second step was to launch the TAO AutoML system. She selected the learning rate, package size, and LoRA parameters. The system conducted forty-three parallel tests to find the perfect settings. The best result using Bayesian optimization was 93.35%. This important stage took nineteen and a half hours. The calculations were performed on several A100 nodes in the Oracle cloud. The Codex agent in this experiment was not limited to simple analysis. He selected the workflow, checked the data, and fixed the configuration. However, his autonomy remained limited by the scope of the instructions given. It is important to note that the 93.35% indicator reflects only the accuracy on a single dataset. It does not measure the actual safety of autonomous driving. Earlier in June, Nvidia researchers and leading universities introduced the ENPIRE framework. It allows AI agents to improve robot management policies on real hardware.