NVIDIA AI Podcast What Open Source Teaches Us About Making AI Better - Ep. 278
15 snips
Oct 21, 2025 Bryan Catanzaro, Vice President of Applied Deep Learning Research at NVIDIA, and Jonathan Cohen, Vice President of Applied Research, delve into NVIDIA's groundbreaking Nemotron open models. They discuss how these models redefine accelerated computing, emphasizing collaboration and efficiency. The duo explains the importance of data quality for faster model training and how enterprises can safely customize these models. They also share insights on the future of AI with larger, multimodal models and innovative developments in low-precision training.
AI Snips
Chapters
Transcript
Episode notes
Open Foundation For Customizable AI
- Nemotron is NVIDIA's open AI technology including models, datasets, and algorithms designed for deep enterprise integration.
- NVIDIA builds Nemotron to both support community customization and to learn how to co-design models with hardware and software.
Models As Part Of The Platform
- NVIDIA treats models as a core part of its accelerated computing platform alongside chips and software.
- Co-designing models, software, and hardware yields lower latency, higher throughput, and better energy efficiency.
Data Quality Trumps Raw Compute
- Smarter pre-training datasets can dramatically accelerate convergence; NVIDIA saw a 4x speedup by improving dataset composition.
- Better data lets you train a smarter model with the same compute budget.