Going In Deep On Data | YC Paper Club
Data inspection beats architecture changes for improving ML model performance, per YC talk
“The correct answer is you look at the data.”
A YC Data Club talk argues that when a model underperforms, the first step should always be analyzing false positives and negatives in the data — not trying new architectures or reading papers. The speaker, drawing from nine years running a deep learning team at Focal Systems, used this as a filter interview question to identify engineers with the right instincts. The talk also highlights how data businesses went from being dismissed as zero terminal value in 2016 to creating roughly $100B in market cap, validating the data-centric AI thesis.