Unleash your Private Datasets to Power up AI
summary
Argues that private data, not prompts alone, makes models useful, comparing two concrete paths: RAG pipelines with Qdrant or Milvus that ingest daily crawls and run vector search (GPU-accelerated, or Postgres with extension for small scale) versus fine-tune Llama 2/3 on single or multi-node GPUs as a one-time snapshot. Gives numbers and tradeoffs: fine-tunes become stale, RAG adds up-to-date books but incurs a 1–5% (or worse if unoptimized) performance penalty, recommends Argo/Kubeflow for ingestion and Nvidia-style guardrails for safety.
AI-written summary · about
chapters (9)
- 0:00Introduction to AI deployment
- 1:04Challenges of AI hallucinations
- 2:54Limitations of basic prompting
- 4:39RAG pipelines explained
- 7:07RAG vs fine-tuning trade-offs
- 9:34Vector databases and pipelines
- 11:41Fine-tuning vs RAG balance
- 13:48AI regulation and guardrails
- 15:36Guardrails for safety and trust
related talks
26mUnlocking the Power of Data-Centric AI.
58mAI-Enhanced Big Data: Integrating Private LLMs and Vector Databases
17mDevOps for Private AI: A Scalable Platform Using Kubeflow
12mDigital & Data/AI Strategy and Transformation
19mScalable Data and AI Solutions with Cloud-Native Architecture
22mGenerative AI Model Data Pre-Training on Kubernetes: A Use Case Study
13mBuilding High-Performance AI Literature Processing
32mKeynote: AI & the stuff built for AI - are they actually useful for data science?