As part of a technical experiment conducted by my team and Nvidia engineers, the Cosmos 3 Nano model demonstrated an impressive performance leap. Initially, the model's accuracy on a four-choice test was 54.41%. However, in less than 24 hours, thanks to the autonomous operation of the AI agent Codex, this figure was raised to 93.35%.
The key factor in this success was the use of the specialized dataset Woven Traffic Safety from Woven by Toyota, containing over 8,000 video examples of road situations with questions. The Codex agent, acting on pre-prepared instructions from the Nvidia TAO toolkit, independently evaluated the base model and selected the LoRA (Low-Rank Adaptation) fine-tuning method.
Interestingly, Codex did not simply launch the process. It demonstrated analytical capabilities: it detected a missing frame rate parameter in the dataset configuration, corrected it, loaded the model weights, and launched the training container. The first LoRA run on eight NVIDIA A100 accelerators (80 GB) took only 30 minutes and raised accuracy to 87.14%.
In the second stage, the developers used TAO AutoML for Bayesian optimization of hyperparameters. The system conducted 43 parallel trials, adjusting the learning rate, batch size, and LoRA parameters. The best result of 93.35% was achieved in 19.5 hours on several A100 nodes in Oracle Cloud Infrastructure.
It is important to emphasize: although the agent operated autonomously, its capabilities were limited by predefined instructions. It did not generate new strategies but merely executed a predetermined workflow. Furthermore, the 93.35% figure pertains to a specific research dataset and is not an indicator of real-world autonomous driving safety.
My expert assessment: This experiment demonstrates the maturity of machine learning automation technologies. LoRA combined with AutoML can save up to 7 times the GPU hours compared to full fine-tuning. However, the key takeaway here is not so much the accuracy but the iteration speed. The ability to go from a base model to a high-precision specialized solution in a single day is a breakthrough in MLOps that will significantly accelerate the deployment of AI in industrial scenarios.