Cosma Shalizi

Life and career

Cosma Rohilla Shalizi arrived at statistics from the outside, which is most of the explanation for what he does inside it. He trained as a physicist, taking his doctorate at the University of Wisconsin–Madison, and his formative research environment was the Santa Fe Institute in its heyday — the period when "complexity science" was the most exciting and least disciplined thing happening in quantitative research. He worked in and around the computational-mechanics program associated with James Crutchfield, whose central question was deceptively simple: given a stream of observations from some process, how much structure does it actually have, and how would you measure that without fooling yourself?

That question turns out to have two halves, and Shalizi has spent his career on both. The constructive half is the mathematics of minimal predictive structure. The destructive half — the one that made him famous well beyond statistics — is the systematic demolition of complexity-science claims that do not survive contact with a likelihood function. He moved from Santa Fe through the University of Michigan's complex-systems center and then to the Department of Statistics at Carnegie Mellon, where he has taught for roughly two decades, retaining an external affiliation with Santa Fe. At CMU he became the department's most visible teacher of the "here is what your regression is actually doing" tradition; his long-running lecture notes on data analysis, written in deliberately plain language and circulated freely, have been read by far more people than have ever sat in the room.

Then there is the blog. *Three-Toed Sloth*, begun in the early 2000s, is the single most influential piece of statistical writing on the internet that is not a textbook. Its house style — bone-dry, ferociously well-read, footnoted with an "attention conservation notice" warning you off before you start — set the tone for a generation of quantitative bloggers. Its function was closer to a public peer-review service. When a celebrated result made the rounds claiming that some network was scale-free, that some social behavior was contagious, that some economic time series had long memory, Shalizi's post explaining exactly which step of the inference was unearned became the thing people linked to. He is not a contrarian for sport; the posts are usually generous about what the authors were trying to do and merciless only about the specific inferential move that fails. But he established, over twenty years, that "this is a very interesting claim and the evidence does not support it" is a legitimate and necessary scientific contribution.

He is also a serious reader outside his field — the *Bactra Review*, his book-review site, ranges across history, political economy, and science fiction — and that omnivory shows up in the work, which frequently asks what a statistical model is *for* rather than only whether it fits.

Key contributions

**Computational mechanics and causal states.** The technical core of Shalizi's early work is the epsilon-machine: given a stochastic process, group all pasts that induce the same conditional distribution over futures into a single *causal state*. The resulting state machine is provably the minimal sufficient statistic for prediction — no smaller representation predicts as well, and the entropy rate of the state process (the "statistical complexity") is a well-defined measure of how much structure the process has, as distinct from how random it is. This is a genuine answer to a question complexity science had mostly been waving at: a coin flip has maximal randomness and zero structure, and a well-defined quantity should say so. With Kristina Klinkner he developed CSSR, a practical algorithm for reconstructing causal states from data, and with collaborators he extended the framework to spatially extended systems. The idea has a clean modern echo: predictive-state representations in reinforcement learning and belief-state methods are close relatives.

**Power laws, and how to tell.** The 2009 paper with Aaron Clauset and Mark Newman, "Power-law distributions in empirical data," is the most-cited thing he has done and functions as a methodological intervention rather than a discovery. The prevailing practice was to plot a complementary CDF on log-log axes, observe something roughly straight, and declare a power law. The paper replaced that with an actual statistical procedure: maximum-likelihood estimation of the exponent, a principled data-driven choice of the lower cutoff above which the tail is claimed, a goodness-of-fit test via bootstrapped Kolmogorov–Smirnov distances, and — critically — likelihood-ratio comparison against plausible alternatives, especially the lognormal, which produces convincing near-straight lines on log-log plots for entirely different reasons. Applied to two dozen canonical datasets, the machinery found that most celebrated power laws were not well supported. The lesson generalizes far past power laws: a model that "looks right" on a transformed plot has not been tested against anything.

**Confounding in social networks.** With Andrew C. Thomas he showed that homophily (similar people become friends) and contagion (friends make you similar) are generically confounded in observational network data — that under standard latent-trait models, the two are not separately identified without assumptions the data cannot check. This landed squarely on a wave of widely publicized findings about the "spread" of obesity, smoking, and happiness through social networks, and it is now a standard citation for why those designs cannot deliver what they claim.

**Philosophy of Bayesian practice.** His paper with Andrew Gelman argues that what good Bayesians actually *do* — fit a model, check it against the data, discover it is wrong, and change it — is not well described by the subjectivist story of coherently updating a prior over a fixed hypothesis space. They recast Bayesian data analysis as hypothetico-deductive: the posterior predictive check is a falsification test, and model expansion is the real inferential engine. He has also written on the frequentist behavior of Bayesian updating under misspecification, where the posterior concentrates on the pseudo-true parameter rather than the truth, and on nonparametric risk bounds for time-series prediction under dependence and non-stationarity — the theory needed to make prediction honest when the data are not i.i.d.

In battle

Shalizi is one of the roster's genuine generalists among the statisticians: mean 41.5 across 102 problems, median 41, with a comparatively thin tail of disasters (only twelve problems at or below 20). He does not have the spiky bimodality of a single-method specialist. What he has instead is a reliable floor on anything that is fundamentally a question of *whether an empirical claim is warranted*.

His ceiling problems are exactly the two halves of his career. "Is it really a power law?" (95) is Clauset–Shalizi–Newman restated almost verbatim; he is not applying a toolkit here, he co-wrote the answer. "Does your pipeline reproduce?" (95) and "The p-value reckoning" (89) put him at the center of the replication crisis, where twenty years of blogging about garden-of-forking-paths analyses, unprincipled model selection, and results that evaporate under re-analysis is directly on point. "Predict, or explain?" (88) is arguably the question his teaching is organized around — the distinction between a model that forecasts well and a model whose coefficients mean anything. "The thousand-year flood from thirty years" (84) is extreme-value extrapolation, where his tail-fitting rigor is precisely the discipline required and his instinct that the answer might be "you cannot responsibly say" is a feature. "Robust to failure, fragile to attack" (80) and "Who will know whom next year?" (75) draw on the network side of his work, and networks is his strongest category at 83.3. "Trained here, deployed there" (78) is distribution shift, where his non-i.i.d. risk-bound work applies almost directly.

The losses cluster in two places, and both are honest. First, *engineering*: "Small tools, loosely joined" (12) is systems-design philosophy, not inference. Second — and more instructive — *representation learning and modern NLP*: "The sentence in a single vector" (12) and "Attention replaces recurrence" (10) sit at his floor, and his NLP category average of 11.0 is the worst on his sheet. His entire methodological stance is that you should be able to say what a model's structure means and whether its claims are earned; a transformer's learned representation offers neither, and nothing in epsilon-machines or likelihood-ratio model comparison transfers to architecture design. "Weight of evidence at Bletchley" (10) is the subtle one: sequential likelihood-ratio testing with a stopping threshold is information-theoretic evidence-weighing, a cousin of his work, but under adversarial daily key changes and wartime deadlines — an operational cryptanalytic problem he never touched. "The variables that must be whole" (12) is integer programming, and "The factor you can't keep changing" (15) is a design-of-experiments constraint problem; his experimental-design average of 26.5 across thirteen problems is a real soft spot, since his relationship to designed experiments is mostly as a critic of people who analyze them badly.

The playbook: send Shalizi against anything phrased as *someone claims X; is that true?* He will win it. Send him against anything phrased as *build the thing* and he will lose.