Fireside Chat: Evolution of AI-first Data Infrastructure
summary
Discussion drills into concrete changes: a move from schematized tables to media and multimodal stores, video decoding optimizations, and a new ML-optimized file format Nimble file format; co-design of models and infra that yields Llama 3 releases and a target of 600,000 GPUs in the fleet; operational work on user-sequence redesign and real-time graphs to avoid row explosion; applied tools like English-to-SQL generation and PyTorch integrations; attention to data quality, deduplication, and synthetic data for future training needs.
AI-written summary · about
chapters (14)
- 0:00AI-first data infrastructure
- 0:48Panelists' backgrounds
- 2:32AI innovation pace
- 4:20Hardware for AI evolution
- 6:05GPU clusters and research
- 7:08Data warehouse scale
- 8:53Generative AI impact
- 11:02Redesigning AI infrastructure
- 13:10AI data infrastructure challenges
- 14:58Data quality importance
- 16:46Generative AI products
- 18:08Research and open source
- 19:32Future of research
- 21:40Closing remarks
related talks
12mDigital & Data/AI Strategy and Transformation
33mPanel Discussion
53mData Engineering and AI
17mFocus on Data, on the Road to AI
48mMastering the AI Shift: Adaptive Strategies for Software Evolution
46mPath of Data & AI Challenges
28mAccelerating AI at Scale with Data
58mAI-Enhanced Big Data: Integrating Private LLMs and Vector Databases