Data-centric AI: A new paradigm
summary
Claims that model development time now shifts from architecture tuning to data-centric AI, showing a concrete LLM workflow: web crawl, deduplicate Associated Press copies, correct gendered noun bias, hire humans to label underrepresented holidays (added Diwali in Google Photos), then fine-tune for dermatology when testers surface missing skin-condition images. Notes that value attribution relies on costly ablation studies (2^n combinations infeasible), and adversarial datasets are built to find jailbreaks and safety gaps.
AI-written summary · about
chapters (8)
- 0:00Introduction to Data-Centric AI
- 0:57The Shift in Development Focus
- 2:10Building and Preparing Data
- 4:25Evaluating Model Quality
- 5:31Fine-Tuning and Data Value
- 7:00Cost of Data and Limitations
- 8:32Safety and Adversarial Data
- 9:30The Future of AI Development
related talks
30mData-Centric AI Cookbook-let's prep that data
33mKeys to Data-Centric AI: Asking AI to Improve its own Dataset.
41mModels Schm-odels: Why you Should Care about Data-Centric AI
23mDeveloping an intersectional framework to analyze biases in artificial intelligence and deep neural
28mData Curation: Transforming Bytes into AI Gold
41mWhen AI Failures Reflect Our Own Social Biases
26mHow to develop useful and unbiased AI
20mDecision Intelligence: The Next Phase in AI