Keynote: AI & the stuff built for AI - are they actually useful for data science?
summary
Trendy tools get judged by concrete results: PyTorch adds numpy compatibility and a compiler that shows 57× CPU and 200× GPU toy speedups on microbenchmarks, JAX offers jit/pmap compilation, and NVIDIA RAPIDS targets GPU dataframes; LLaMa/GPT and diffusion models excel at pattern-matching, generation and faster lookup but fail on deep reasoning and robust explainability. The practical rule is clear: use GPUs and compiled kernels for hot, repeated pipelines that actually run faster; use LLMs for drafting, search and brainstorming but do not automate critical decisions.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 1:24Bridging AI and data science
- 4:12AI vs robust data science
- 7:49PyTorch and JAX workflows
- 12:08GPUs in data processing
- 17:13Mental model for LLMs
- 22:55LLM applications and limits
- 27:29Diffusion models in data science
- 28:10Q&A
related talks
24mKeynote: Next-generation frameworks for Large-scale AI
12mKeynote: What's New in AI
44mKeynote: Tracking the Pulse of Generative AI- 12mKeynote: What's New in AI
12mKeynote: Build Accelerated LLM Applications with Open Source Tools from NVIDIA
26mKeynote: Artificial Innovation Accelerates Artificial Intelligence
16mKeynote: Accelerating AI/ML Development with Open Source Software
12mKeynote: AI Infrastructure