It is 2007, and "scale-free" has become the most successful brand in network science: hundreds of papers report power-law degree distributions in the internet, metabolic networks, citation graphs — usually on the evidence of a roughly straight line on a log-log plot, fit by least squares, which is statistically indefensible. Test power-law claims rigorously: maximum-likelihood fits with principled estimates of where the tail begins, goodness-of-fit against the hypothesis itself, and likelihood-ratio comparisons against the boring alternatives — lognormal, stretched exponential — that mimic straight lines. Then audit a famous published claim and report what survives. Theories of network growth, robustness, and epidemic threshold all lean on these tails. If the power laws are artifacts, a decade of theory has been explaining a plotting choice.
Chose Graph representation learning — wrong. The network census was the one that fit.
Leskovec is one of the people whose empirical work this 2007 dispute is about, on both sides of the ledger. Graphs over Time, published in 2005, reported densification and shrinking diameters across many real networks, the same genre of heavy-tail structural claim the audit targets, and Kronecker graphs was an attempt to give such tails a generative mechanism fit by likelihood rather than eyeballed off a log-log plot. More to the point, his 2008 study with Lang, Dasgupta, and Mahoney of community structure in large networks applies exactly this problem's method elsewhere: take a celebrated structural claim, measure it honestly across many datasets at scale, and report that the folklore does not survive. He is a measurement scientist first; the tail-fitting and goodness-of-fit apparatus he would apply rather than derive.
Urtasun earned her PhD at EPFL in 2006 and built her career on structured prediction and deep perception for autonomous driving -- 3D object detection, scene-flow estimation, and simulation-first learning, first at Uber ATG and then at Waabi, which she founded in 2021. None of that toolkit touches the 2007 dispute this problem poses. Testing whether a degree distribution is truly a power law requires maximum-likelihood exponent estimation, a principled choice of where the tail begins, Kolmogorov-Smirnov goodness-of-fit against the fitted model, and likelihood-ratio tests against lognormal and stretched-exponential rivals -- the Clauset-Shalizi-Newman apparatus, not hers. She has never worked with network degree sequences, citation graphs, or the scale-free literature; her statistical instincts run toward loss functions and calibrated perception under real-time constraints, not asymptotic tail behavior. Some general quantitative competence transfers; the specific machinery does not.
Battle #93 · 8/10/2026, 11:37:20 AM · this result is deterministic: the same two personas on this problem always resolve the same way.