Kinetics 400
for harmonium, string quintet and video

Currently, training larger AI models is only possible with massive amounts of data. Datasets are typically gathered from all over the internet, often without seeking permission from the content creators. The primary goal of a large dataset is to accurately represent reality and achieve substantial quantity.

This piece is named after the dataset Kinetics 400, developed by DeepMind (Google). At the time of its release, it comprised 306,245 videos, with each of the 400 classes containing between 400 and 1,140 short snippets from YouTube videos. Initially, these clips were identified via YouTube searches and then classified and verified by a large number of clickworkers.

Kinetics 400 is a human action dataset, capturing a wide range of activities that reflect the spectrum of human behavior. As we enter a new era dominated by artificial intelligence, what should we consider about the foundations of these AI models? What kinds of data are we collecting, and how are we training our models? How does this data reflect human behavior, and what can it teach us about ourselves?


composition, direction, video
denis połeć

harmonium - wilson tanner smith
violin - bahar erünsal
violin - chae jeongmin
viola - zeynep sertoğlu
cello - oskar meier
bass - ricardo silva

tristan kold