
I collected every public review on OpenReview for ICLR, NeurIPS, and ICML. ICLR submissions grew 19-fold in eight years, and one in five now has 'LLM' in the title, yet those papers are accepted at exactly the average rate. Plus: the score that gives a coin-flip chance of acceptance, what rebuttals are worth, and the rise and fall of 'delve'.
Oct 7, 2026

I decomposed a year of hourly district-to-district flows in the Seoul metropolitan area into a handful of hidden travel motives, using flow counts only. Six motives match a static model with 463,000 parameters, the commuting motives move with subway smart-card counts (correlation 0.66–0.77), and on festival and rally days the model leaves +44–56% unexplained arrivals in Yeouido-dong. Recovering the provider's trip-purpose labels, however, barely beat time of day alone.
Oct 7, 2026