Screencasts could be scalable data + evals for single-user emulation (Guardian Angels)

Researchers exploring personalized AI have a new data-collection approach: according to LessWrong, one researcher is building a system that continuously captures screencasts, webcam footage, and user inputs at four frames per second to train "Guardian Angel" models. These personalized AI systems could predict which tweets you'll click on, which research papers you'll read through, or which technical concepts you'll investigate—without requiring manual labeling of training data. The estimated cost: less than one hundred fifty dollars per year, and less than ten terabytes of storage per year.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton