Brady Heywood · Published 2022-05-19

Complex Systems Thinking – How to change the way we think about problem solving

privateReview: unreviewedConfidence: high
complexityemergencepower lawself-organisationtipping pointagentic-systemsagent-based-modellingmulti-agent-systemstipping-pointsfeedback-loopspower-laws
Open on YouTube ↗

Summary

Overview

  • Speaker: Dr Sean Brady
  • Channel: Brady Heywood
  • Main topic: Complex Systems Thinking and Complexity Science
  • Purpose: To help viewers shift their problem-solving mindset from Newtonian reductionism to complex systems thinking by understanding non-linear interactions, emergence, and self-organisation. Dr Sean Brady discusses how humans approach problem-solving through a Newtonian vs. complex systems lens. He explores why traditional reductionist thinking fails for complex systems and introduces core concepts such as emergence, tipping points, self-organisation, and power laws, using examples like ants, starlings, cities, earthquakes, language, sand piles, and forest fires.

Topic Map

Newtonian vs Complex Thinking

  • Explanation: Comparison of traditional Newtonian (reductionist, linear, equilibrium-based) thinking with complex systems thinking which focuses on interactions and emergent properties.
  • Key claims:
    • Newtonian thinking assumes systems are predictable, rational, measurable, and controllable.
    • Newtonian thinking breaks systems into components to understand the whole.
    • Complex systems are defined by interactions between components, not the components themselves.
  • Examples:
    • Studying a single ant vs an ant colony
    • Isaac Newton's impact on science
  • Terminology:
    • Newtonian thinking
    • reductionism
    • equilibrium
    • deterministic
    • causal
  • Why it matters: Understanding the limitation of Newtonian thinking prevents misapplying reductionist tools to complex problems.

Power Laws across Systems

  • Explanation: How diverse natural and social systems exhibit power law relationships when plotted on log-log scales.
  • Key claims:
    • Diverse complex systems exhibit power law distributions.
    • Large events occur infrequently, while small events occur frequently.
  • Examples:
    • Basal metabolic rate vs body size in animals
    • Earthquake magnitude and frequency
    • City population sizes
    • Word frequencies in Jane Austen's Pride and Prejudice
  • Terminology:
    • power law
    • log-log scale
    • scaling
  • Why it matters: Power laws indicate underlying common natural laws governing complex systems across different domains.

Analysis Methods: Machine Learning vs Agent-Based Modelling

  • Explanation: Comparing machine learning (good for prediction without understanding) with agent-based modelling (good for understanding without prediction).
  • Key claims:
    • Machine learning acts as a black box that yields prediction without understanding.
    • Agent-based modelling provides understanding of emergent behavior without high predictive accuracy.
  • Examples:
    • NetLogo software
    • Thomas Schelling's segregation model
  • Terminology:
    • machine learning
    • agent-based modelling (ABM)
    • NetLogo
    • black box
  • Why it matters: Choosing the right analytical tool depends on whether the goal is prediction or structural understanding.

Key Concepts of Complex Systems

  • Explanation: Deep dive into emergence, tipping points, self-organisation, and self-organized criticality.
  • Key claims:
    • The whole is greater than the sum of its parts.
    • Simple local rules can produce complex global emergent behaviors.
    • Small causes can trigger massive global impacts when systems reach critical points.
  • Examples:
    • Starlings flocking behavior
    • Schelling segregation model
    • Forest fires and density tipping points
    • Sand pile avalanches
  • Terminology:
    • emergence
    • tipping point
    • self-organisation
    • self-organized criticality
    • edge of chaos
    • positive feedback
  • Why it matters: These concepts explain why complex systems behave counterintuitively and collapse unexpectedly.

Key Points

Complex systems are about interactions, not components

  • Explanation: Breaking a complex system down into individual parts destroys the interactions that create its complex behavior.
  • Evidence: Studying a single ant tells you little about an ant colony.
  • Practical implication: When analyzing a complex system, focus on the relationships between agents rather than the agents in isolation.

Emergence means the whole is greater than the sum of its parts

  • Explanation: Interactions among agents produce system-level behaviors that were not explicitly programmed or designed into the individual agents.
  • Evidence: Starlings flocking in complex shapes using only three simple local rules.
  • Practical implication: Expect unexpected system behaviors that cannot be deduced purely by examining individual components.

Tipping points and positive feedback drive sudden state shifts

  • Explanation: In systems with positive feedback, small linear increases can push the system past a tipping point into a drastically different regime.
  • Evidence: Forest fire spread models at varying tree densities; holiday home pricing bubbles.
  • Practical implication: Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.

Frameworks, Models & Processes

Boid Rules (Starling Flock Model)

  • How it works: Local rules governing individual agent movement in a flock without central authority.
  • Components:
    • Alignment (moving in the same direction as nearby birds)
    • Separation (avoiding crowding nearby birds)
    • Cohesion (moving towards nearby birds)
  • When to use: Simulating decentralized swarming or collective animal/agent behavior.

Forest Life Cycle (Complexity Loop)

  • How it works: The cyclical transition of complex ecosystems between chaos, self-organisation, self-regulation, lock-in, and collapse.
  • Components:
    • Chaos
    • Self-organisation
    • Self-regulation
    • Lock-in
    • Collapse
  • When to use: Understanding resilience, vulnerability, and systemic collapse in ecological, economic, or socio-technical systems.

Examples & Case Studies

Thomas Schelling's segregation model with agents preferring 40% similar neighbors.

  • Illustrates: Emergence of full racial segregation from mild individual preferences.
  • Lesson: Macro-level outcomes can be far more extreme than micro-level intentions due to emergent interactions.

Simulating forest fires in NetLogo at varying tree densities.

  • Illustrates: Tipping points and self-organised criticality.
  • Lesson: Linear increases in density can lead to non-linear catastrophic shifts once a tipping point is crossed.

Actionable Takeaways

  • Immediate:
    • Shift focus from system components to component interactions.
    • Recognize that not all systems operate in equilibrium or follow linear cause-and-effect.
    • Look for power law distributions in real-world data.
  • Strategic:
    • Use agent-based modelling to build understanding of emergent properties.
    • Identify positive feedback loops and potential tipping points in business or engineering systems.
    • Abandon pure reductionism when tackling complex socio-technical problems.
  • Questions to investigate:
    • What are the local interaction rules driving the behavior of our organization?
    • Are we mistaking a complex system for a Newtonian machine?
    • Where are the hidden positive feedback loops in our system that could lead to a tipping point?

Claims Worth Verifying

  • Word frequencies in Jane Austen's Pride and Prejudice follow a power law. (empirical linguistic claim)
  • Basal metabolic rate follows a power law relationship with body mass across mammalian species. (biological scaling claim)

Notable Quotes

"Nature and Nature's Laws lay hid in night: God said: Let Newton be! And all was light." (at 0:58) "What does it really mean to say that the whole is greater than the sum of its parts? It's not magic, but to us humans with our crude little human brains it feels like magic." (at 36:48) "Somehow, by constantly seeking mutual accommodation and self consistency, groups of agents manage to transcend themselves and become something more." (at 37:15) "The key driving force behind self-organized criticality is that microlevel agent behaviour tends to cause the system to self-organise and converge to critical points at which small events can have big global impacts." (at 60:14) "There is no master neuron in the brain, for example, nor is there any master cell within a developing embryo. If there is to be any coherent behaviour in the system, it has to arise from competition and cooperation among the agents." (at 90:22)

Compressed Summary

  • Newtonian thinking relies on reductionism, determinism, and equilibrium, which fail for complex systems.
  • Complex systems are defined by interactions and produce emergent behavior greater than the sum of their parts.
  • Agent-based modelling and machine learning offer complementary tools for understanding and prediction.
  • Power laws appear across diverse complex systems like earthquakes, cities, and biology.
  • Self-organised criticality means small local events can trigger massive global cascades past tipping points.
  • Keywords: complexity, emergence, power law, self-organisation, tipping point
  • Core insight: Complex systems cannot be understood by studying their components in isolation; their behavior is driven entirely by non-linear interactions producing emergence and tipping points.

Core insights

5
Architecturemedium noveltystrong evidence

Complex systems are defined by interactions between components, not the components themselves. The behaviour of an agentic system lives in the interaction graph, message contracts, shared-context topology, and feedback loops between agents, not inside any single agent implementation.

Why it matters

An engineer who optimizes individual agents in isolation (better prompts, models, tools) will not fix failures that arise from how agents interact. Observability and abstraction boundaries must be placed around inter-agent interactions, and tests must exercise the whole interaction graph of the system.

Generalization

For any system composed of interacting parts, the primary unit of analysis and design should be the interaction, not the isolated component.

Complex systems are defined by interactions between components, not the components themselves.
Open source video
Studying a single ant tells you little about an ant colony.
Open source video
Empirical Resulthigh noveltystrong evidence

Mild local preferences can produce extreme global states: Schelling's segregation model shows that agents tolerating 40% similar neighbours can end up fully segregated. Macro-level outcomes are not a proportional reflection of agent-level rules or intentions.

Why it matters

In multi-agent systems, small local biases can amplify into population-level pathologies such as monoculture, tool-use lock-in, cascading retry storms, or context-conformity collapse. Local design decisions cannot be assumed safe just because they look mild.

Generalization

Validation and safety cannot be established by checking individual agent rules; they require population-level simulation and monitoring of aggregate outcomes.

Macro-level outcomes can be far more extreme than micro-level intentions due to emergent interactions.
Open source video
Emergence of full racial segregation from mild individual preferences.
Open source video
Empirical Resultmedium noveltymoderate evidence

Many diverse natural and social systems exhibit power-law event-size distributions: many small events and a few very large ones. Treating agent-system failures with average-case thinking misses that rare, huge events can dominate total risk.

Why it matters

Reliability and capacity engineering for multi-agent systems should explicitly model tail events, cascades of unbounded size, and extreme correlated failures. Averages and ordinary percentile SLOs are poor guides if the underlying distribution is heavy-tailed.

Generalization

When frequency/size data is power-law distributed, mean and high-percentile metrics mislead; use tail-aware planning, extreme-event drills, and explicit cascading-failure scenarios.

Large events occur infrequently, while small events occur frequently.
Open source video
Diverse complex systems exhibit power law distributions.
Open source video
Practicehigh noveltystrong evidence

Prediction and understanding are distinct capabilities for complex systems: machine learning gives a black-box prediction without understanding, while agent-based modelling gives understanding of emergent behaviour without high predictive accuracy. Teams must choose the tool by the question they ask, or deliberately combine both.

Why it matters

An LLM-based evaluation may tell you that something is wrong but not why; an agent-based model with simple interaction rules can identify the mechanism. Conversely, an agent-based model may be too stylized for accurate production forecasting. Designing for agents needs a portfolio of predictive and mechanistic modelling, not a single all-purpose eval.

Generalization

Any modelling effort should separate the questions 'can we predict?' and 'can we explain?' and select or combine tools accordingly.

Machine learning acts as a black box that yields prediction without understanding.
Open source video
Agent-based modelling provides understanding of emergent behavior without high predictive accuracy.
Open source video
Mechanismhigh noveltymoderate evidence

Positive feedback loops make stable-looking systems vulnerable to sudden state changes: forest fires ignite nonlinearly as density crosses a threshold, so systems can look fine until a small final increment pushes them over a tipping point.

Why it matters

A production agentic system with reinforcing loops—retries increasing load, shared context contamination, viral tool-use patterns—can collapse abruptly with little warning. Architectural controls should add negative feedback and damping, and operators should monitor feedback-loop strength, not just static level metrics like latency or error rate.

Generalization

For any complex engineered system, assume regime shifts are possible; identify feedback loops and design dampers before the system approaches a critical threshold.

Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.
Open source video
Small causes can trigger massive global impacts when systems reach critical points.
Open source video

Deep dives

5

Interaction-first architecture for agentic systems

Research question

When a multi-agent system fails, how much of the failure variance is attributable to interaction topology, message contracts, and shared-context dynamics versus individual agent capability?

Why

Agentic systems are complex systems; if interactions dominate behaviour, isolated prompt/model testing cannot catch systemic failures, and engineering investment should shift to interaction protocols, feedback loops, and population-level tests.

Complex systems are defined by interactions between components, not the components themselves.
Open source video
Studying a single ant vs an ant colony
Open source video
Source video

Early-warning signals for feedback-driven collapse in multi-agent workloads

Research question

Which measurable telemetry signals—feedback-loop strength, rate of change, variance, lag-1 autocorrelation—give warning before an agent system crosses a tipping point, and can dampers be placed before collapse occurs?

Why

Stable-looking systems with positive feedback can collapse suddenly, and conventional level metrics like latency or error rate may not signal an approaching phase transition.

Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.
Open source video
Small causes can trigger massive global impacts when systems reach critical points.
Open source video
Source video

Mild local preferences and emergent population pathologies in multi-agent systems

Research question

At what preference thresholds, network structures, and feedback strengths do mild local agent preferences become population-level monoculture, lock-in, or cascading failure, and what guardrails can be introduced pre-deployment?

Why

Local design decisions look safe individually, but emergence decouples macro outcomes from micro intentions; safety must therefore be validated at population level rather than by checking individual agent rules.

Macro-level outcomes can be far more extreme than micro-level intentions due to emergent interactions.
Open source video
Emergence of full racial segregation from mild individual preferences.
Open source video
Source video

Hybrid predictive-mechanistic evaluation for agent systems

Research question

How should an engineering team combine black-box machine-learning prediction with agent-based mechanistic modelling so that forecasting and root-cause explanation are both first-class capabilities?

Why

LLM evals and trace analytics give prediction without understanding while ABMs give understanding without high predictive accuracy; relying on a single all-purpose evaluation sacrifices one of these capabilities.

Machine learning acts as a black box that yields prediction without understanding.
Open source video
Agent-based modelling provides understanding of emergent behavior without high predictive accuracy.
Open source video
Source video

Power-law tail risks in agentic failure distributions

Research question

Do failure sizes in real or simulated multi-agent workloads follow power-law distributions, and what tail-aware capacity and SLO policies should replace average-case planning?

Why

If failures are heavy-tailed, rare giant cascades dominate total risk and standard mean/percentile engineering systematically underestimates catastrophic exposure.

Large events occur infrequently, while small events occur frequently.
Open source video
Diverse complex systems exhibit power law distributions.
Open source video
Source video

Article ideas

4

Optimising Agents Is Not Enough: Engineering the Interaction Graph

Most systemic failures in agent platforms will come from the interactions between agents, not from individual prompts or models; therefore engineering rituals, abstractions, and ownership must move to interaction topology, message contracts, and feedback loops.

Angle

Contrast current agent-template development with a complexity view borrowed from ant colonies.

Source video

Prediction Is Not Understanding: Why Agent Eval Needs a Second Model

An LLM judge or learned regression may predict quality, but it cannot say why; teams must pair black-box prediction with agent-based modelling of interaction rules or they will keep fixing the wrong layer.

Angle

Challenge single-pipeline evals that conflate forecasting with explanation.

Source video

The Schelling Trap in Agent Teams: Mild Biases Create Extreme Lock-In

Small local preferences for popular tools, formats, or strategies can drive an agent population into far more extreme monoculture and lock-in than any individual intent would suggest, so teams need aggregate guardrails, not just local rules.

Angle

Read Thomas Schelling's segregation model as a warning for agent collective behaviour.

Source video

Your Agent Platform Is a Fire: Plan for the Rare Huge Cascade

Because agentic failures follow a 'many small, few huge' pattern, SLO engineering based on averages or percentiles leaves teams exposed to the dominant risks; resilience must be designed around tail scenarios and blast-radius isolation.

Angle

Extrapolate forest-fire and sand-pile dynamics to agent incidents.

Source video

Project ideas

4

Interaction-Layer Failure Attribution

new

Varying message contracts or feedback rules in a multi-agent task loop produces a larger change in end-to-end failure rate than varying individual agent prompts or models over comparable engineering effort.

Proof of concept

Implement a minimal task-driven multi-agent system with swappable agents and interaction modes. Run sweeps: keep agents fixed and alter interaction topology or shared-context policy; then keep interactions fixed and alter one agent.

Measurement

Effect size on task success or failure rate across intervention sweeps, normalised by intervention cost.

Source video

Runaway-Feedback Collapse Simulator

gatehouse

A retry/load feedback model of an agent workload shows critical slowing down (rising autocorrelation and variance) before catastrophic failure and outperforms static latency alarms in lead time at equal false-alarm rate.

Proof of concept

Build a Python or NetLogo simulation where failing agents retry and publish load; increase request rate until collapse. Compute lag-1 autocorrelation and variance on a system-state time series.

Measurement

Detection lead time, false-positive rate, and AUC compared with static latency or error-rate thresholds.

Source video

Agent Monoculture Phase-Transition Probe

new

A low-tolerance local 'copy what neighbours use' rule will produce an extremely concentrated tool or prompt distribution above a sharp threshold, whereas below that threshold the same rule leaves high diversity.

Proof of concept

Implement an agent-based model of agents choosing among tools where each agent mildly prefers the tool used by a majority of its neighbours; sweep tolerance threshold and graph rewiring; visualise concentration.

Measurement

Herfindahl-Hirschman Index and discontinuity slope over threshold sweeps, plus critical threshold width.

Source video

Diagnose-a-Shift: Black-Box vs Mechanistic Root-Cause Harness

beyond-evals

When an agent workflow is changed by a single interaction rule, ABM-based parameter fitting identifies the changed rule more accurately than black-box LLM or trace explanations, while black-box prediction forecasts the new outcome more accurately than the ABM alone.

Proof of concept

Generate logs from an ABM with known interaction rules, train a black-box sequence predictor, then create shifted logs by changing one rule. Ask each diagnostic path to localise the changed rule and forecast outcomes.

Measurement

Forecast error for the predictor and top-1 root-cause accuracy for each explanation path.

Source video

Architectural implications

5

A complex system's properties come from the interaction graph among components, not from components in isolation.

Before

We design and test each agent independently, assuming system reliability is the sum/product of individual agent reliability.

After

We add first-class abstractions for interaction protocols, shared-context constraints, and feedback loops; agents are designed with local rules, and system-level tests exercise whole interaction graphs.

Consequence

New failure classes become visible, such as feedback amplification and cascade resonance, and individual agent optimizations no longer dominate system quality.

Source video

Emergence means macro outcomes can be far more extreme than micro intentions.

Before

Agent rules are treated as local specifications and global behaviour is assumed to follow directly from them.

After

We run population-level agent-based simulations before deployment to find emergent pathologies, and we enforce global guardrails (diversity, lock-in prevention, load concentration limits) rather than trying to centrally specify every action.

Consequence

Deployment safety gating now includes aggregate-state checks, not just per-agent correctness, and unexpected collective behaviours are caught before production.

Source video

Prediction and explanation are decoupled in complex systems.

Before

A single pipeline (recorded traces, offline replay, or black-box LLM scoring) is expected to both forecast outcomes and explain failure causes.

After

Build two complementary evaluation paths: a predictive/statistical path for quantitive forecasting, and an agent-based model path for mechanistic hypotheses through parameter sweeps over interaction rules.

Consequence

Debugging shifts from prompt tweaks to interaction-rule changes; predictions remain optimizable, while mechanism hypotheses are explicitly testable.

Source video

Positive feedback and tipping points create discontinuous collapse in seemingly stable systems.

Before

Scaling and reliability assume approximately linear degradation; thresholds are set from current load or latency observations.

After

Add dampers (retry budgets, backpressure, bounded shared context) to weaken positive feedback; monitor rate-of-change and second-order effects; include circuit breakers that can stop cascades.

Consequence

Operations teams can handle bifurcation-like behaviour where the same input produces radically different outcomes, and are less likely to be surprised by a sudden, irreversible collapse.

Source video

Power-law event sizes imply that rare, very large events can dominate total risk.

Before

Reserve capacity and SLOs based on averages and percentiles of historical event sizes; failure blast radius is assumed bounded by component capacity.

After

Design for unbounded cascade sizes by isolating blast radii, using independent per-agent retry and state budgets, and deliberately injecting extreme correlated failures in tests.

Consequence

Reliability investment shifts from shaving typical latency to preventing or absorbing rare global events; cost estimates explicitly include tail exposure.

Source video

Tradeoffs and failure modes

5

Prediction vs understanding (ML vs ABM)

Benefit

Machine learning can predict well without a mechanistic model; agent-based modelling can provide causal and structural understanding even where precise numerical prediction is weak.

Cost or risk

Choosing ML when explanation is needed yields opaque, potentially uncontrollable systems; choosing ABM when accurate prediction is needed provides poor quantitative forecasts; using the wrong one can drive decisions on misleading evidence.

Machine learning acts as a black box that yields prediction without understanding.
Open source video
Source video

Emergent macro-outcomes from local rules

Benefit

Simple local rules can generate coordinated, scalable global behaviour without centralized control, as in flocking and Schelling-style models.

Cost or risk

System-level outcomes can be far more extreme than agent-level intentions, so emergent behaviour may violate safety, fairness, or operational constraints and is not controllable by tuning individual agents alone.

Macro-level outcomes can be far more extreme than micro-level intentions due to emergent interactions.
Open source video
Source video

Reductionist component decomposition

Benefit

Breaking a system into components brings clarity, measurability, and familiar engineering control.

Cost or risk

For complex systems, decomposition removes the very interactions that produce the system-level behaviour, so isolated component analyses give false confidence.

Breaking a complex system down into individual parts destroys the interactions that create its complex behavior.
Open source video
Source video

Optimizing close to a tipping point

Benefit

Exploiting positive feedback can drive rapid adoption, learning, or scale-up when a change pushes the system through a desired phase transition.

Cost or risk

The same feedback can push a stable-looking system into sudden collapse once the edge of chaos is crossed; there may be no gradual warning visible in conventional metrics.

Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.
Open source video
Source video

Average-case metrics under power-law distributions

Benefit

Mean and common percentile metrics are simple to track and summarize most frequent observations.

Cost or risk

With a power law, rare large events dominate total harm, so mean-based decisions systematically underestimate catastrophic risk.

Large events occur infrequently, while small events occur frequently.
Open source video
Source video

Open questions

5

What are the local interaction rules driving the behaviour of our agentic system, and which of them can produce uncontrolled emergent global effects?

Why unresolved

The video shows that macro outcomes cannot be reliably deduced from micro-level preferences or rules; you need simulation or observation to discover which local rules amplify.

Research direction

Systematically map agent interaction rules (when to call tools, how context is shared, how retries are triggered) and run parameter sweeps in agent-based models to find phase boundaries.

Source video

Are we mistaking a complex system for a Newtonian machine?

Why unresolved

Newtonian thinking assumes predictability, rationality, measurability, and control, but complex systems with interactions and feedback violate those assumptions in ways that are hard to see from inside the system.

Research direction

Design perturbation experiments: change one interaction parameter slightly and measure whether outcomes scale continuously or jump discontinuously; if jumps appear, treat the system as complex rather than machine-like.

Source video

Where are the hidden positive feedback loops in a production multi-agent system, and can their strength be monitored online?

Why unresolved

Positive feedback loops are not visible as static metrics; they are self-reinforcing rates such as retry amplification, context contamination, or tool-use concentration that can push a system over a tipping point.

Research direction

Instrument rate-of-change and correlation of signals across agents; test early-warning indicators such as rising variance, slowing recovery, or critical slowing down before collapse.

Source video

Do power-law failure distributions actually appear in real agent-system incidents, and how heavy are the tails?

Why unresolved

The video documents power laws in natural and social systems, but the summary provides no direct evidence of heavy tails in engineered multi-agent systems.

Research direction

Collect incident sizes and frequencies across agent workloads; fit distributions on log-log axes and compare power-law against lognormal/exponential alternatives to plan capacity and risk.

Source video

Can complex global coordination in agentic systems be achieved with only a small set of local rules, or is central orchestration still necessary for task-level accountability?

Why unresolved

The starling flock example shows complex patterns emerging from three local rules, but the summary does not specify when such decentralization is sufficient for goal-directed agentic tasks.

Research direction

Benchmark decentralized local-rule agents against centralized orchestrators on tasks requiring coordination, resilience, and accountability, varying the number and abstraction of local rules.

Source video

Key claims

7
comparativeVerification needed

Complex system behaviour is determined by component interactions rather than by components themselves.

Evidence

Complex systems are defined by interactions between components, not the components themselves.

Question

For a given multi-agent workload, how much of total behaviour variance is explained by interaction topology versus individual agent capability?

Source video
factualVerification needed

Micro-level preferences can produce macro-level outcomes far more extreme than the initial preferences.

Evidence

Macro-level outcomes can be far more extreme than micro-level intentions due to emergent interactions.

Question

Across which local rule shapes and network structures does amplification exceed expectations?

Source video
factualVerification needed

Power-law distributions appear across diverse natural and social systems.

Evidence

Diverse complex systems exhibit power law distributions.

Question

Are the cited examples statistically consistent with a single power-law family or with other heavy-tailed distributions?

Source video
comparativeVerification needed

Machine learning yields prediction without understanding, while agent-based modelling yields understanding without high predictive accuracy.

Evidence

Machine learning acts as a black box that yields prediction without understanding.

Question

What hybrid modelling approaches can provide both accurate prediction and mechanistic explanation in engineering practice?

Source video
causalVerification needed

Small causes can trigger massive global impacts when systems reach critical points.

Evidence

Small causes can trigger massive global impacts when systems reach critical points.

Question

What measurable early-warning signals indicate that an agent system is approaching a critical threshold?

Source video
causalVerification needed

Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.

Evidence

Stable-looking systems can collapse rapidly when positive feedback loops push them to the edge of chaos.

Question

Can the strength of positive feedback loops be reliably estimated online from telemetry before collapse occurs?

Source video
factualVerification needed

The whole is greater than the sum of its parts in emergent systems.

Evidence

The whole is greater than the sum of its parts.

Question

Under what conditions can full system-level behaviour be derived from interaction rules, and when does genuinely novel macro-behaviour appear?

Source video

Connections

5