What this covers

Neil Johnson, a physicist at George Washington University, discusses how complex systems—traffic jams, armed conflicts, online extremism—self-organize according to universal mathematical laws rather than the initial conditions or stated goals of their participants. The episode centers on a striking empirical finding: vastly different phenomena, from insurgent attacks across Iraq, Colombia, and Afghanistan to the growth of extremist support online, all exhibit a power law distribution with a slope of approximately 2.5. Johnson argues this universality signals a common underlying mechanism: coalescence-fragmentation dynamics in which clusters of agents form, grow, and occasionally splinter, with no constraint from physical proximity. He contrasts complex systems with merely complicated ones, emphasizing that true complexity arises when heterogeneous, adaptive agents with memory interact across scales, generating phenomena that cannot be computed from their parts alone.

The conversation ranges across applications and implications. Johnson explains how the El Farol problem—a thought experiment about bar attendance that captures why traditional rational-actor economics fails—leads to a counterintuitive solution: self-organized crowd-anti-crowd splits that emerge without coordination. He discusses infodemiology, noting that misinformation spreads like viral contagion but without the geographic constraints that let epidemiologists employ social distancing. A central claim, contested in the community, is that mitigation should nudge entire systems rather than target individual bad actors—targeting the emergence of power-law distributions before violent or extremist movements fully coalesce. Johnson also reflects on science as a crowd-and-anti-crowd system itself, arguing that trust in science means trusting the contentious, self-correcting process rather than any single result, a distinction especially fraught as scientific uncertainty plays out in real time before global audiences.

Sharpest takeaway

Johnson argues that complex systems—from traffic to insurgencies to online misinformation—self-organize into universal statistical signatures (notably a 2.5-slope power law) produced by simple coalescence-fragmentation dynamics, and that recognizing these patterns enables systems-level intervention rather than targeting individual 'bad actors.'

  • Complex systems differ from merely complicated ones because cross-scale feedback among adaptive, heterogeneous agents produces emergent phenomena that cannot be computed from parts.
  • Diverse conflicts and online extremist movements share a 2.5 power-law slope generated by clusters coalescing and occasionally fragmenting.
  • Because big clusters grow from small ones, mitigation should nudge the system (slow small-cluster coalescence) rather than hunt individual bad actors.

The argument · threads8 threads · 33 claims
0.81

Power laws with specific slopes indicate shared generative mechanisms across seemingly different systems.

6 pointscentrality 5/5
  • Plotting fatalities per event across very different modern conflicts (Iraq, Colombia, Afghanistan) and terrorist events yields not a bell curve but a power law, and remarkably these conflicts—despite different motivations, locations, and ideologies—share approximately the same 2.5 slope, implying a common underlying mechanism in how humans conduct conflict.

    “We got power laws, and they all had the same slope... they all seem to have the slope around the red slope, which in this case is a 2.5 slope.”

  • The reason power laws matter is not the debate over exact exponents but that the existence of a power law points to a physical generating mechanism distinct from one producing a bell curve; two systems with the same power-law slope (like two mountains with the same ski-run gradient in different continents) likely share a common underlying mechanism, even if measurement imperfections prevent observing a perfect power law.

    “We're never gonna get a perfect power law. So the power law is more indicative that the system is on some kind of tipping point.”

  • A power law is characterized by scale invariance—what happens at one scale is identical to another, just rescaled—so the same theory applies across scales (e.g., big and small earthquakes share one mechanism); such distributions occur when a system sits near a phase transition, making power laws an indicator that a system is at a critical tipping point.

    “the power law is characterised by the fact that what happens at one scale is exactly the same as what happens at another scale. It's just twice as much, or half as much.”

  • Power laws with exponents between two and three have a diverging standard deviation (infinite fluctuation size), so a system at 2.5 can form arbitrarily small or large clusters—an effectively infinite adaptability that confers an evolutionary advantage to a system trying to compete and win, which Johnson prefers as the explanation for the 2.5 slope over the geometric two-dimensional clustering account.

    “power laws between two and three actually have... If you were to calculate their standard deviation... it actually diverges. And so there are systems that are, in some sense, infinitely adaptable.”

  • A minimal model in which insurgent units coalesce into clusters and occasionally fragment (like fish schooling then scattering from a predator), with coordination allowed over arbitrary distances rather than constrained by physical proximity, not only produces a power law but specifically one with a 2.5 slope—explaining the observed conflict and online-extremism distributions without invoking ideology.

    “that simple, simple, simple rule of coalesce, coalesce, and then occasionally fragment, remarkably gives exactly... Not only does it give a power law, it gives a power law with a 2.5 slope.”

  • Tracking the growth of online support for ISIS in 2014-2015 revealed online communities (clusters) coalescing and being periodically fragmented by social-media moderators removing them for violating terms, producing a near-perfect 2.5 power law identical to that found in physical conflicts, and preliminary results suggest far-right radical groups follow the same pattern.

    “When we went and looked online at the growth in support for ISIS during 2014, 2015, there, the clusters are, they are communities online and we found a pretty perfect 2.5 power law.”

0.76

Systems-level interventions like nudging clusters and self-organized strategies outperform top-down removal of bad actors.

5 pointscentrality 5/5
  • Social-media efforts to remove individual bad influencers are hard and counterproductive: you must first find them (if they even exist on a good-to-bad spectrum), avoid being sued, and risk that removing a powerful person generates even more support for them, which is why system-level nudging is preferable.

    “then you'd have to stop being sued by them... maybe they're quite powerful people... by you removing them, that will create even more support for them.”

  • Current mitigation strategies from the social sciences focus on individual 'bad actors' or bad apples, which is like trying to stop water from boiling by removing the fastest molecule—futile because another always replaces it; a systems view recognizes that any future extremist movement will be preceded by clusters building into a 2.5-slope power law, so the alarm signal is the emergence of that distribution rather than any individual.

    “Which is a little bit like saying, "I'm gonna stop water boiling by taking out the bad molecule."”

  • A good solution can emerge from individual agents acting in their own interest without central planning or explicit cooperation, and forcing the outcome top-down (telling people which group to join) would produce larger fluctuations and a worse result than letting competition self-organize.

    “the fluctuations that will come out of trying to force people to do something would be even worse... Would be worse than the ones that come out of competition.”

  • An emergent solution to the El Farol problem arises when agents inadvertently split into a 'crowd' following one strategy and an 'anti-crowd' following the opposite, without any coordination or local information; this self-organized 50/50 split yields zero fluctuation in system output, which is even better than randomized (coin-toss) attendance that still produces large fluctuations—suggesting complex systems could be controlled by nudging agents into paired opposite strategies.

    “without any coordination, without any local information, just global information fed back to them... they've managed to split themselves into a 50/50, which means that... you're hitting 50, 50, 50, 50, 50. You have no fluctuation.”

  • Because large extremist clusters grow from small ones and are harder to tackle once robust, the effective strategy—analogous to preventing big sand-pile avalanches by managing small ones—is to nudge the system by slowing the coalescence of small clusters, which is less legally fraught and less likely to backfire than removing powerful individuals who can sue or gain sympathy when removed.

    “To prevent the big avalanches in the sand pile, just take care of the small ones so that they don't build into big ones, 'cause the big ones need the small ones to build them.”

0.72

Complex systems differ from physics by featuring heterogeneous adaptive agents with memory that cannot be simplified via classical symmetries.

4 pointscentrality 5/5
  • The simplifying assumptions that let physics make progress—such as interactions not depending on time, conserved energy, and stable symmetries—do not hold in complex systems like traffic, where there is no conserved energy and behavior at midnight differs from 6 PM, so those symmetries that would normally help solve the problem are absent.

    “The traffic, there's no conserved energy. There's no sense in which it's the same at midnight as it is at 6:00 PM. So somehow, all these trends, all these symmetries that would have helped us in the problem are just not there.”

  • What distinguishes complex systems from ordinary physics systems is that they are assembled from individual pieces that are heterogeneous, carry memory of the past (making them non-Markovian), and adapt their strategies—often to worse ones—rather than being dumb, memoryless particles that simply bounce off each other.

    “So we're heterogeneous, we have memory of the past, and we adapt. So if it doesn't work, I'm gonna change my system, and probably to a worse one, but I will change.”

  • A system becomes complex rather than merely complicated when its parts depend on prior states and on simultaneous actions of others, so that feedback across scales (e.g., other drivers congesting your route) makes outcomes impossible to compute without actually running the system—out of which surprising phenomena like traffic jams and stock crashes can emerge.

    “imagine if each one of those pieces now depends in some really complicated way on what's happened before, and there are other people doing things at the same time that impact us”

  • Every system of interest—including the expanding universe and living organisms—is a non-equilibrium system, because equilibrium is effectively death; what is interesting is the behavior away from the dead state, which is also the hardest regime to model.

    “In the long run, we're all dead. Equilibrium is death... it's the behaviour away from the dead state that we're all interested in.”

0.58

Online platforms form a multiply-connected multiverse where misinformation spreads without physical distance constraints.

4 pointscentrality 4/5
  • Social media platforms (Facebook, Gab, 4chan) function as separate, non-interacting universes that people tunnel between like wormholes across both space and instantaneously, and because cached postings persist, the connection spans time as well—forming a kind of multiverse, a multiply-connected fractal space in which misinformation spreads and mutates, an ideal domain for physicists who think in universes and couplings.

    “I literally think of this as universes with some kind of wormhole between them that people can go through across space”

  • Misinformation spread is analogous to viral contagion (justifying the new field of infodemiology), but a key difference is the absence of distance: one can 'sneeze' in one place and infect someone on the other side of the planet instantly, so social distancing is not a viable mitigation strategy online.

    “I can sneeze on one place, and somebody catches the cold or whatever on the other side of the planet instantly, 'cause it's online. So the notion of distance. What does it mean to social distance online?”

  • Difficulties in countering online harm arise partly because bad actors behave like self-organizing complex-system agents forming organic clusters, while official counter-efforts are top-down and homogeneous: official pandemic-information pages were boring and repetitive, whereas misinformation clusters produced many varied messages, each more likely to hit an individual's specific vulnerability.

    “the crazy ones said a million different things so they are more likely to hit someone in their sweet spot of vulnerability”

  • Online anti-vaccine communities are persuasive because each cluster develops its own flavor and blends accurate statements (e.g., there is no five-year safety record for a COVID vaccine) with emotionally resonant content (images of babies' hands, appeals to parental love), so a generic official banner from Facebook or the CDC fails because it does not address the specific flavor of that cluster—like a parent intruding to say 'do this instead.'

    “they're talking about, "Well, there'll be no safety... There's no five years of safety for the COVID vaccine." And that's accurate, they're absolutely right. And then suddenly at the top appears where they had these lovely pictures of a baby's hand”

0.57

Fat-tailed distributions dominate social and physical systems, making average-based reasoning dangerously misleading.

7 pointscentrality 3/5
  • Many social and physical systems follow fat-tailed distributions rather than bell curves, so reasoning from averages misleads: with wealth, almost everyone is below the average because a long tail of very wealthy people pulls the mean up, and building economies, buildings, or medicine around an assumed average fails when one entity can have 1,000 times another.

    “One in a billion has more than twice, or more than three times the average wage. We all know that that is not true. In fact, that tail of wealth goes on a long, long way”

  • The El Farol bar problem—where many heterogeneous agents independently decide each week whether to attend a finite-capacity bar—captures the essence of complexity: if everyone adopts the same rational strategy, it becomes self-defeating because they all go or all stay, creating internal frustration and bouncing between metastable states, so agents must act inductively from limited memory rather than computing a fixed equilibrium.

    “imagine we all hit the same strategy... That means we either all go, or we don't... And of course, that immediately becomes the wrong thing to do.”

  • Despite the El Farol problem being introduced ~20 years ago, the economics and finance field has not embraced Arthur's inductive, adaptive agent behavior because it does not fit their paradigms: there is no rational actor to perturb and no stable equilibrium—the system instead bounces among metastable states, which is closer to a physicist's worldview.

    “still nearly 20 years later, it really... They haven't really embraced what Brian Arthur was really talking about in terms of this inductive behaviour, rather than absolutely being able to compute the ground state, and then sticking with it.”

  • The reproduction (R) number—whether an infected person infects more or fewer than one other before recovering, determining whether spread grows or dies out—is calculated in an ad hoc way in epidemiology because the actual contact network is unknown, forcing epidemiologists to model disease propagation like a well-mixed test tube of susceptibles and infecteds rather than as a real network.

    “it's calculated for... In a very... In many ways, a ad hoc way in the epidemiology literature... they just don't know the network that people are in”

  • It is speculatively plausible that scale-free power-law distributions arise because they maximize or minimize some entropy or entropy-production rate, especially if entropy is interpreted as information; in that view living systems play an information game with nature, locally borrowing and returning information while the whole system's entropy increases.

    “particularly if we take the vision of... the interpretation of entropy as an information... Measure. Then in that case, I'm absolutely with you on that”

  • Complex systems research is still in a pre-paradigmatic stage where there is no agreed-upon foundational equation or canonical homework problem (no Schrödinger equation, no Hamiltonian), making it a double-edged sword: enormous unexplored territory but little concrete to show outsiders.

    “Where's your Schrödinger equation? Where's your Hamiltonian? You have nothing. You have nothing to show.”

  • Lewis Fry Richardson, a physicist, sought to understand the human condition during the First World War by tracking numbers of casualties in wars—though limited to war totals rather than per-event counts—anticipating Johnson's modern per-event casualty analysis of Iraq, Afghanistan, and Colombia.

    “Lewis Fry Richardson was a... He's a physicist... in the First World War, got in his head that somehow to understand the human condition, human conflict, he would just track numbers of people in casualties... in wars.”

0.54

Extremist movements coalesce through power-law cluster dynamics independent of shared ideology or coherent messaging.

3 pointscentrality 4/5
  • Extremist support can grow rapidly without any coherent shared message because adherents are united only by what they oppose—like spins in a two-spin system that can point at any angle in the plane yet all oppose the main direction—so far-right groups who disagree fundamentally (one wanting a united Europe, another wanting to break into tribal caves) remain allied by their common target.

    “It wasn't that there was some coherent message between them all, it's just that what they were opposing was such an obvious... target, and so you could oppose it in many ways.”

  • Traditional set-piece warfare (two armies lining up at dawn until one side is annihilated) behaves like a test tube of two reacting species that fully consume each other, whereas modern insurgencies involving a large state and an organic opposition do not, because insurgents instead form into clusters and what is observed is cluster-based attacks rather than uniform attrition.

    “that's like a test tube of two reacting species that kind of just interact all day, and there's nothing left at the end of it except the product. But these other... The modern wars, the insurgencies”

  • When complex systems like forests or earthquakes show power-law size distributions characteristic of phase transitions, it is not because the system is literally transitioning from one phase to another over time, but because it is perpetually poised between possible phases at the critical point; describing how such transitions unfold in time is the newer, harder domain of dynamical phase transitions.

    “Describing how something would happen in time is the area of dynamical phase transitions, which is a really... That's a much newer area of physics because it's a system out of equilibrium.”

0.40

Trust in science means trusting the self-correcting process of contentious debate, not individual results or outcomes.

3 pointscentrality 3/5
  • Science has never before been conducted in real time with the whole world scrutinizing every pre-print and misstep, and the central challenge is preventing the inherent unknowns of science—without which scientists would have no job—from being perceived by the public as damning uncertainty, since losing trust in science leaves people with no trustworthy source of guidance.

    “I don't think science has ever had a situation whereby it's being done in real time, and the whole world is on top of every pre-print, and every misstep”

  • Trust in science should mean trust in the process—the bottom-up, contentious, self-correcting method where scientists argue and even backstab over competing theories—not absolute trust in any single result or paper, which is appropriate because science is full of unknowns and individual scientists are biased and sometimes wrong; the crowd and anti-crowd of disagreeing researchers together constitute the system that aims, in the long run, in a good direction.

    “when I mean trust in science, I do not mean trust in one particular result in science... I mean trust in the process.”

  • Following Steven Strogatz's view that taboo or unpleasant topics are interesting precisely because they are taboo, the recent availability of large datasets on conflict fatalities makes it scientifically valuable to set aside motivational narratives and simply analyze the numbers.

    “he finds it interesting to look at topics that are somehow taboo. And that stuck with me. That that's an interesting area just because it's unpleasant.”

0.18

Game theory may be a branch of physics via mean-field approximations and Boltzmann weighting analogies.

1 pointcentrality 2/5
  • Game theory may secretly be physics, since the probability forms used by agents resemble Boltzmann weightings of two-level systems at finite temperature, and the emerging field of mean-field game theory was explicitly inspired by the physics mean-field approximation as game theorists sought to handle large numbers of players.

    “the forms of the probability that are used by "agents", look very much like two-level systems in a finite... At a particular temperature, the probabilities, the Boltzmann weightings.”