Twenty-six articles on the data behind modern artificial intelligence, one per newsletter issue, covering annotation, evaluation, feedback, provenance and language.
In 2022, a 1.3 billion parameter model fine-tuned on roughly 77,000 carefully collected prompts was preferred by human evaluators over a model more than a hundred times its size. The difference was not architecture and it was not compute. It was data. That is what it means to say the bottleneck moved, and it is where this series begins.
Across eight parts, The Data Layer of AI looks at what makes training and evaluation data good, why annotators disagree and what to do about it, how generative AI changed the work, where synthetic data helps and where it harms, and what all of this means for building AI for India.
Talk to us about your annotation, evaluation and feedback data needs.
Contact us →The data foundation enterprises trust to build reliable AI, since 2021.