NVIDIA post-training workflow and GitHub AI billing updates #85
Today's Letter
NVIDIA, one-day Cosmos 3 post-training workflow
- NVIDIA outlined a one-day post-training workflow for Cosmos 3 Nano using TAO agent skills and LoRA.
- On the Woven Traffic Safety video QA dataset, exact-match accuracy rose from 54.41% to 87.14% in one run.
- TAO AutoML pushed peak accuracy to 93.35% in 4-way multiple-choice evaluation.
- Cosmos 3 uses a mixture-of-transformers design with separate autoregressive reasoning and diffusion generation paths.
- NVIDIA said LoRA adapts the model with about 7x fewer GPU hours than full-parameter supervised fine-tuning.
- Deployment is positioned around Cosmos 3 Reasoner NIM, which serves LoRA adapters through OpenAI-compatible endpoints.
Source: developer.nvidia.com
GitHub, AI credit pools added to billing UI
- GitHub added direct AI credit pool management to the billing UI for cost centers, replacing a REST API-only workflow for this setting
- The feature is available for Copilot Business and Copilot Enterprise on GitHub Enterprise Cloud
- Admins can enable the AI credit pool while creating or editing a cost center in the billing interface
- GitHub calculates the pool limit automatically from assigned Copilot licenses and updates it as licenses are added or removed
- At the pool limit, enterprises can either block further included usage or allow continued usage as additional spend when overages are permitted
- The AI credit pool is designed to keep each cost center within the included AI credits funded by its own Copilot licenses
- The pool is separate from a cost center budget, which caps metered charges after the included credit pool is exhausted
- GitHub says both controls can be configured on the same cost center for combined usage and spend management
Source: github.blog
Jocoletter curates AI, software, and product trends for developers and builders.
#GitHub #NVIDIA