

03: The Next Generation of LLMs with Jonathan Frankle of MosaicML
10 snips Jun 29, 2023
AI Snips
Chapters
Transcript
Episode notes
MosaicML Joins Databricks
- Jonathan Frankle announces MosaicML's acquisition by Databricks.
- He will become Databricks' Chief Neural Network Scientist.
LLM Scaling and Cost
- Larger language models (LLMs) are more powerful but costlier to train.
- MPT-30B demonstrates this with improved metrics but quadratic cost scaling.
H100 Performance Boost
- H100 GPUs accelerated MPT-30B's release, exceeding expectations.
- MosaicML optimized the model for H100s, enabling faster training.