Aaron Clauset

@aaronclauset.bsky.social

NO KINGS. NO FASCISTS. FUND SCIENCE. Professor of Computer Science @ BioFrontiers Institute at University of Colorado, Boulder and External Faculty @ Santa Fe Institute orcid: https://orcid.org/0000-0002-3529-8746

How to get the most out of this class, in the age of AI -- advice from an expert

Coding agents and foundational AI models are incredible tools now. In the hands of an expert, they can dramatically amplify your ability to collect, analyze, and understand biological network data. But, you are not yet an expert and these tools can make it harder for you to build that expertise.

In this class, my goals for you are to (i) develop the network intuition and knowledge of an expert, and (ii) develop the expert skill to solve real problems with biological network data. It's possible to complete all of the problem sets with zero effort, and even parts of the final project---AI coding agents are that good. However, you will learn much more and build greater expertise if you instead embrace the struggle of figuring things out on your own.

That does not mean never using coding agents. It does mean waiting to use them until you're expert enough to know: when are they amplifying your own knowledge vs. substituting theirs for yours? It means accepting that sometimes you will feel lost, you'll get things wrong, you'll have to look things up manually, and you will need to study without AI tools. That's the price of becoming an expert yourself. For that, I am here to help. Attend lecture, ask questions, take notes, come to office hours, do the problem sets, form a study group, etc.

Money quote from the preprint's conclusion : "The choice being presented to the scientific community — discovery or appli-cation, curiosity or impact — is a zero-sum illusion. The researchers who pursue fundamental questions are the same ones who end up transforming medicine, defence and industry — not because they planned to, but because transformative discoveries do not respect funding categories. Fundamental research is the foundation of the building. No one sees it while the skyscraper is standing, but remove it and everything above comes down."

Title slide for the talk, "Escalation dynamics and the severity of wars", by Aaron Clauset

I cannot decide whether the souls of our learned overseers are merely stunted, or if they have been utterly hollowed out by the winds of their own indifference. I had thought the threshold of their competence was set so low that it brushed the very heather, yet even there, they stumble and fall headlong into the mire.

They speak now of these "Artificial Intelligences"—soulless phantoms they’ve invited to haunt the halls of the mind—proving that even a shallow grave is too high a hurdle for those who have forgotten what it is to possess a human heart.

Screenshot of the top of my 2025 Year in Review post, with a look back at 2025, by the numbers

Figure 3: NAS election and gendered effects on individual productivity λ and prominence θ for physicists. a,b, Gender and the annual individual productivity λ and individual prominence θ of physicists that have become NAS members, before and after their election. c,d, Gender and the individual productivity λ and individual prominence θ of co-authors of physicists that have become NAS members, before and after their election. Shaded areas and error bars represent 95% confidence intervals.

Figure 2: Gender and NAS membership effects on individual productivity and prominence of physicists. a, Temporal trends of the election of men and women NAS members from 1970 to 2023. The upper panel shows the probability of an active publishing physicist being elected to the NAS. b, Gender and individual productivity λ of NAS members and other physicists over their careers. c, Gender and individual prominence θ of NAS members and other physicists over their careers. Shaded areas represent 95% confidence intervals.

Figure 1: Gender and individual productivity and prominence of physicists.
a,b, Gender and temporal trends of the raw productivity, denoted by the number of papers, and prominence, using total citation counts, of mid-career physicists from 1970 to 2010, using the year that these researchers published their first paper. c, Illustrative scheme to show how to untangle the collaboration network effects of productivity and prominence using first-/last-author publications. d,e, Gender and dynamical patterns of individual productivity λ and prominence θ of mid-career physicists. Shaded areas represent 95% confidence intervals.

Screenshot of the top of the paper "The undervaluing of elite women in physics",
Abstract: "Elite women in physics wait longer than
men for recognition. Once elected to the US National Academy of Sciences, however, their prominence surges — evidence that their work was undervalued all along."
Body: "Physics has changed dramatically over the past 50 years, both in the range of phenomena it studies and its demographic composition. For example, the share of publishing women physicists worldwide has more than doubled — from 4.5% in 1970 to 9.2% in 2020 — and continues to increase. Yet parity remains distant, raising the question: to what degree has the physics community progressed toward the meritocratic ideal of recognizing scientists purely on the basis of their contributions?

We address this question in two ways. First, we compare broad trends in the productivity and prominence of men and women physi- cists worldwide over the past 50 years. Second, we examine the careers of elite physicists, asking whether the high honour of being elected to the US National Academy of Sciences (NAS) impacts the careers of men and women in physics differently. Our analysis leverages a recently developed Bayesian network model1 to estimate individual measures of scientific productivity and prominence from large-scale bibliographic data. We apply this model to a global physics collaboration network constructed from first- and last-author pairs in 9.1 million physics journal articles (omitting single-author papers) published between 1950 and 2023 and recorded in the OpenAlex bibliographic database. Within this global network, we focus on 93,456 established physicists with at least 30 years between their first and most recent publication and with at least 10 first- or last-author publications..."

Screenshot of a slide from the talk: "little in academia makes sense except in the light of prestige",
• understanding epistemic and social inequalities requires understanding how prestige drives and maintains them
• untangling the effects of social biases can make science better & more meritocratic
talk outline:
• prestige shapes who joins the scientific workforce
• prestige shapes which scientific ideas spread
• prestige shapes who persists as a scientist, and what they study

Screenshot of the cevianlabs.io website. Banner says "Accelerating Academic Hiring", followed by The Cevian Advantage, Cevian Labs provides advanced tools to faculty and administrators to make better decisions, faster. Founded by faculty who spent years wishing for a way to automate the tedious parts of faculty hiring (summarizing CVs, tracking publications, synthesizing candidate information) and finally just built the solution we always needed. Our secure platform automates manual error-prone processes to deliver...

Slide title: "network effects in scientific labor"
networks mediate most scientific activities:
1. scientific training, hiring, collaboration, teaching, attention, peer review, etc.
2. networks act like a form of unequally distributed social capital

a productive collaborator --> increases your productivity
a prominent collaborator --> increases your prominence

How much does who you work with impact your productivity and prominence?

Fig 2. Five average productivity curves (publications per year) are shown, one for each strata of university prestige, showing a rapid rise to an early peak (around expected tenure time) followed by a slow decline to a stable, prestige-specific value. Averages over researchers at all levels of institutional prestige follow similar productivity trajectories, in agreement with the conventional narrative, but at differing scales of output.

Slide titled 'network effects in scientific labor', arguing that your own productivity is a network effect, because it includes the productivity of your coauthors too

Title slide: "What drives the productivity of scientific labor?" a talk by Aaron Clauset at the Oxford Summer School on Economic Networks 2025

A delightfully nerdy photo of @arash.network receiving a small karate club trophy from @nwlandry.bsky.social

Multi-panel figure from the paper. Caption: Racial diversity-association effects in the co-authorship networks of researchers. Racial diversity-association effects in co-authorship networks relative to the randomized null model, and how the racial diversity of junior co- authors in the established period is correlated with that of their early co-authors for individual researchers. The cohort of junior co-authors for a researcher is defined to be racially diverse if the racial entropy score is greater than the baseline rate predicted by the null model. a) The interplay between the racial diversity of junior co-authors in the early-career and established periods of researchers. b) Racial diversity-association effects in co-authorship networks over time, from 1960 to 2012, according to the first career years of individual researchers. c) Racial diversity-association effects across countries, where we select the top-20 countries with the most researchers and arrange them according to the continents they belong to. d) Statistical distribution of racial diversity-association effect Δc of countries with at least 100 researchers. Bars in a represent mean values and error bars indicate 95% confidence intervals. Lines denote mean values and shaded areas represent 95% confidence intervals in panel b. We report effect sizes using the Cohen’s d statistic and use two-sided t-tests for comparisons. P values reported in the plot are adjusted by the Benjamini– Hochberg procedure for multiple comparisons. ***P < 0.001.

Multi-panel figure from the paper. Caption: Gender diversity association in the co-authorship networks of researchers. Results of gender diversity-association analysis in the co-authorship networks relative to the randomized null model, showing how gender diversity of junior co-authors in the established period correlates with that of their early co-authors for individual researchers. The cohort of junior co-authors for a researcher is defined to have a high proportion of women if the percentage of women is greater than the baseline rate predicted by the null model. a) The interplay between the gender diversity of junior co-authors in the early-career and established periods of researchers, for men and women. b) Gender diversity- association effects in co-authorship networks over time, from 1960 to 2012, according to the first career years of individual researchers. c) Gender diversity-association effects by the number of junior co-authors in the early-career period. d,f), Gender diversity association across countries, for men (d) and women (f). e,g) Statistical distribution of gender diversity-association effect Δc of countries with at least 100 researchers, for men (e) and women (g). We select the top-20 countries with the largest number of established researchers and arrange them according to the continents they belong to. Bars in a and c represent mean values and error bars indicate 95% confidence intervals; the error bars in c are too small to be graphically visible. Lines denote mean values and shaded areas represent 95% confidence intervals in panel b. We report effect sizes using the Cohen’s d statistic and use two-sided t-tests for comparisons. P values reported in the plot are adjusted by the Benjamini–Hochberg procedure for multiple comparisons. ***P < 0.001.

Multi-panel Figure from paper. Caption: Comparing gender diversity of advisors’ groups and researchers’ own groups. The association effects of gender diversity relative to the null model using the mentorship data. a) The interplay between the gender diversity of advisees in the group of a researcher’s advisor and that of his/her own group, for men and women. The cohort of advisees in a research group is defined to have a high proportion of women if the real percentage of women is greater than the baseline rate predicted by the null model. b) The gender diversity- association effects over time from 1970 to 2010, based on the researchers’ first year as advisees. As data are less available for researcher groups in the early period, we merge data points for every 5 years instead of displaying yearly data. c) The interplay between the research group size of researchers’ advisors and the gender diversity-association effects. Bars represent mean values and error bars indicate 95% confidence intervals in panels a and c. Lines denote mean values and shaded areas represent 95% confidence intervals in b. We report effect sizes using the Cohen’s d statistic and use two-sided t-tests for comparisons. ***P < 0.001.

$Multi-panel figure from the paper. Caption: Diversity-association effects in the social networks of researchers in the early-career and established periods. a) An illustrative example of the gender diversity-association effect for two researchers during the early-career and established periods. The early social network of the red author comprises mostly women, and his network in the established period continues to be dominated by women. The green author interacted more often with men in their early career, and her connections in the established period are mostly men researchers as well. b) The group size distribution of advisors for researchers in the mentorship data. c) The fraction of an advisor’s advisees who are women as a function of advisor gender and the gender composition of their own early-career environment (n = 17, 917). d) Number of early junior co-authors of researchers in the collaboration network. e) The correlation between the proportion of women junior co-authors for researchers in the early-career and established periods (n = 562, 494). Bars represent mean values and error bars indicate 95% confidence intervals in panels c and e. We report effect sizes using the Cohen’s d statistic and use two-sided t-tests for comparisons. ***P < 0.001.$