Codex Boosts Cosmos3 Nano with Two Prompts
kimmonismus · x · 2026-07-15
NVIDIA stated that Codex, using just two prompts, improved the accuracy of Cosmos 3 Nano on the Toyota Woven Traffic Safety dataset from 54.41% to 93.35%.
During the experiment, the agent skills of NVIDIA TAO allowed Codex to automate multiple workflows:
- Detecting and fixing missing video metadata
- Running zero-shot baselines
- Generating LoRA configurations
- Launching training and evaluation
- Executing AutoML hyperparameter searches
- Reporting the best model
NVIDIA also noted:
- A single LoRA training takes about 30 minutes using 8 A100s, reaching an accuracy of 87.14%.
- The second prompt concurrently launched 43 AutoML trials, running for 19.5 hours across multiple A100 nodes to ultimately reach 93.35%.
- LoRA saves approximately 7x GPU-hours compared to full-parameter training.
The core conclusion of the post is that agent skills are becoming the interface for general-purpose coding agents to invoke specialized ML infrastructure.
Related event: NVIDIA Shows TAO/LoRA Boost for Cosmos 3 Nano(4 posts)→
More from coding & agent
- Devin adds e2b sandboxes for remote agent execution — badphilosopher · 2026-07-22
- Hermes Agent Refactoring Proposal: Decoupling via Event Bus and Monorepo Slicing — Promptmethus · 2026-07-22
- ty now reads Pydantic config keywords and field metadata — charliermarsh · 2026-07-22
- Pensar Launches AI Security Agent to Autonomously Discover and Patch 0-Days — andriy_mulyar · 2026-07-22
- ty adds first-class Pydantic support, including strict and lax field handling — charliermarsh · 2026-07-22
- Google launches Gemini 3.5 Flash Cyber for CodeMender, with limited access for governments — GoogleAI · 2026-07-22