The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?

arXiv:2012.12689 · eess.SY, cs.AI, cs.SY, nlin.AO · Submitted 2026-08-15 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?".

Jane: The paper was written by Guido Fioretti from University of Bologna.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title and Big Question: Tom: Welcome back to the show, everybody. Today we're digging into a paper with a title that honestly sounds like a philosophical riddle: "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?" I'm Tom, and as always, Jane is here with me. Jane, that title alone — it's asking whether dumb parts make a smart system, or whether that's just a myth.

Jane: It really is, Tom. And the author, Guido Fioretti from the University of Bologna, he's not just being clever. He's tackling a genuine debate in the world of agent-based modeling. You know, when we build computer simulations of societies or markets, we have to decide how smart to make the individual simulated people. The old school says keep them simple, the "KISS" principle. The newer school says make them detailed and realistic, the "KIDS" principle.

Tom: Right, and the paper's title is basically asking which one actually works. Does a crowd of simpletons somehow produce genius group behavior, or do you need smart individuals to get a smart collective?

Jane: Exactly. And Fioretti's approach is really neat. Instead of just arguing about it abstractly, he took one specific, famous model — the Lotka-Volterra predator-prey model — and started tinkering with the "intelligence" of the simulated wolves and sheep.

Tom: The classic ecological model, right? The one that shows how predator and prey populations oscillate over time.

Jane: That's the one. But here's the twist: he's not just interested in animals. The Lotka-Volterra model is also used to describe business cycles, technological competition, even ideological struggles. So when he's asking whether predators and preys can learn to coexist, he's really asking whether capitalists and workers, or competing firms, can find a stable balance.

Tom: And that's where it gets fascinating, because the basic model is actually unstable. In the classic setup, the system almost always collapses — either both species die out, or the preys explode and the predators starve. Coexistence is incredibly rare.

Jane: Right, and that's the puzzle. The paper starts with this question: if the simple version fails, can we make the agents smarter so they figure out how to coexist? And the surprising answer is... mostly no. Smarter doesn't automatically mean better.

Tom: So we've got a paper that's questioning a core assumption in simulation science. I love it. But before we get into the weeds of what worked and what didn't, let me ask you — what's the vibe you're getting from the results so far?

Jane: The vibe is that there's a sweet spot. Simple agents fail, but so do hyper-smart agents. The interesting stuff happens in the middle, and that's what we're going to unpack next.

Tom: Perfect. So stick around, because we're about to find out what kind of "middle intelligence" actually saves the wolves and sheep from extinction.

The Core Findings: Tom: So we're back, and we're still talking about "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?" Jane, last segment we set the stage. Now let's get into the meat. What did Fioretti actually do with his wolves and sheep?

Jane: Okay, so he took the standard NetLogo model — you know, the one where wolves chase sheep — and he started swapping out their behavioral rules. He tried a bunch of different "intelligence levels." First, he tried simple adaptive rules, like predators reproducing based on how many preys are around. That failed. Then he tried what he calls "meta-cognition" — agents that understand the whole system and make collective agreements.

Tom: Like the predators agreeing to stop reproducing if they outnumber the preys, to avoid wiping out their food source?

Jane: Exactly. And that one actually worked. When the predators showed that kind of foresight, the two species coexisted ninety-six point seven percent of the time. That's a huge jump from the basic model where coexistence happened less than one percent of the time.

Tom: But here's the kicker — when the preys tried the same trick, it backfired spectacularly. When preys agreed to stop reproducing if they outnumbered the predators, they ended up taking over completely ninety-nine point eight percent of the time. The preys went extinct in the long run because they grew without limit.

Jane: Right, and that asymmetry is one of the most interesting parts of the paper. Predators, being the ones who control the consumption, can use intelligence to stabilize the system. Preys, even with the same intelligence, can't. They just end up destroying themselves with their own success.

Tom: And then he tried the really smart stuff. Agents with perfect foresight — they can predict the future of the whole system. And what happened? Total disaster. When both species could see their inevitable doom, they all committed suicide. Literally. one hundred percent extinction.

Jane: It's almost poetic. The smarter they get, the more they realize the system is doomed, and they just give up. So that's the "Possibly Not?" part of the title — more intelligence doesn't always mean better outcomes.

Tom: But then there's the one case that really blew my mind. The middle ground. Instead of perfect foresight, the agents just used simple linear extrapolation — looking at recent trends and projecting them forward. Like, "preys are increasing, so I'll reproduce more."

Jane: And that's where things get wild. When both predators and preys used this simple prediction rule, they coexisted seventy-nine point seven percent of the time. But here's the twist — both populations grew without limit. The wolves and sheep were coexisting, but they were also exploding in numbers. They formed these massive moving columns, chasing each other across the simulated landscape.

Tom: It's like a perpetual growth machine. And the author connects this to capitalism — you know, the Lotka-Volterra model applied to business cycles. He suggests that economic growth might only emerge once enough people adopt this extrapolation-based way of thinking. Saving, investing, projecting trends into the future.

Jane: Which is a profound idea. The simple act of predicting the near future, based on what just happened, can transform a system from oscillating around a fixed point into one that grows forever. That's not something anyone would have guessed from the math alone.

Tom: So we've got simple agents failing, super-smart agents failing, and these middle-ground extrapolators creating explosive coexistence. What does that tell us about the original question?

Jane: It tells us the answer isn't a simple yes or no. It depends on what kind of intelligence, and who has it. And that's exactly what we're going to dig into next.

Improvements and Implications: Tom: Welcome back. We're still on "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?" and Jane just dropped the bombshell about extrapolation creating endless growth. Now, what does this mean for the field? What's the actual improvement this paper suggests?

Jane: Well, Tom, the paper doesn't propose a new algorithm or a new model. It proposes a new way of thinking about agent intelligence. Instead of asking "how smart should agents be?" it asks "what kind of intelligence actually changes the collective outcome?" And the answer is surprisingly specific.

Tom: And that answer is — first-order derivatives, right? The ability to sense the rate of change of another agent's population.

Jane: Exactly. Fioretti's argument is that when individuals can compute first-order derivatives — basically, "is the other species growing or shrinking?" — that's enough to create higher-order effects at the collective level. The individuals don't need to understand acceleration or curvature. The interactions between them generate those higher-order dynamics for free.

Tom: So it's like emergence. The collective computes the second derivative, even though no individual knows what a second derivative is.

Jane: Precisely. And he's pretty confident that this generalizes beyond the Lotka-Volterra model. Any system where agents interact and multiply their effects — snowball effects, feedback loops — could show this pattern. Simple individual predictions, combined with interaction, can produce complex collective dynamics.

Tom: But he also says higher-order individual intelligence doesn't help. He bets that if agents could compute second derivatives, it wouldn't add anything useful. It might even cause overfitting, like those sophisticated stock market agents that perform worse than simple ones.

Jane: Right, and that connects back to the earlier findings. The perfect foresight agents — they were essentially trying to compute the whole future, all derivatives at once. And it paralyzed them. They couldn't act. The extrapolators, with just one derivative, they could act and create something new.

Tom: Now, Lu and Meng are joining us. Lu, you're the AI researcher — does this resonate with what you see in machine learning?

Lu: Absolutely, Tom. This reminds me of the difference between overparameterized models that memorize noise and simple models that capture the trend. The extrapolators are like a simple linear regression — they capture the slope, not the curvature. And in many real-world settings, that's enough to be effective. The paper gives a concrete simulation showing why that might be true in social systems too.

Meng: But from an engineering standpoint, I have to ask — how do you actually implement this? The paper uses a parameter called "Timespan" to decide how many past steps the extrapolation looks at. And the sensitivity analysis shows that if you set it too long, the magic disappears. So the improvement is real, but it's fragile. You need to tune it right.

Jane: That's a fair point, Meng. The paper acknowledges this. With a Timespan of seven instead of five the coexistence rate drops from seventy-nine point seven percent to sixty-five point one percent. So the specific implementation matters. But the qualitative finding — that simple prediction enables explosive coexistence — is robust.

Tom: So the improvement isn't a recipe. It's a principle. And that principle is: don't make your agents too smart, but give them just enough predictive ability to react to trends.

Lu: And I'd add — the paper also suggests that heterogeneity matters. The agents aren't all identical. They have different positions, different local information. That heterogeneity is what allows the collective to compute those higher-order derivatives.

Tom: Alright, so we've got a principle, we've got a warning about tuning, and we've got a big idea about emergence. What does this mean for the world outside the simulation? Let's wrap this up.

Conclusion: Tom: We're in the final stretch now, still talking about "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?" Jane, give us the big picture. What should our listeners take away from this paper?

Jane: The takeaway, Tom, is that the relationship between individual intelligence and collective intelligence is not linear. It's not monotonic. There's a sweet spot. And in the Lotka-Volterra model, that sweet spot is simple prediction — looking at recent trends and acting on them.

Tom: And the explosive growth result — that's the part that's going to stick with me. Two species coexisting while both populations explode. It's not sustainable, but it's stable in the short term. And the author connects that to our own economic system, which also seems to be growing without limits.

Jane: Right, and that raises a deep question. Is that explosive coexistence a form of collective intelligence, or is it collective madness? The paper says it's undecidable within the model. Environmentalists would say it's heading for collapse. Economists would say innovation will save us. The model can't tell us which is right.

Lu: And that's actually the most honest conclusion a paper can reach. It identifies the mechanism — extrapolation-based prediction — that generates the explosive dynamics, but it doesn't pretend to know whether that dynamics is good or bad. That's a value judgment, not a scientific one.

Meng: From my side, I appreciate that the paper is transparent about its limitations. The sensitivity analysis shows where the results hold and where they break. That's the kind of rigor we need more of in simulation research.

Tom: So, final thoughts. This paper doesn't give us a definitive answer to the title's question. But it gives us a framework for thinking about it. Simple agents can create smart collectives, but only if they have the right kind of simple intelligence — the ability to sense and react to change.

Jane: And that's a beautiful idea. The collective can be smarter than any individual, but only if the individuals are just smart enough to pay attention to each other. Not too smart, not too dumb. Just enough to extrapolate.

Tom: Well said, Jane. That's a wrap on "The Less Intelligent the Elements, the More Intelligent the Whole. Or, Possibly Not?" by Guido Fioretti. Thanks to Lu and Meng for joining us. And to our listeners — keep asking questions, keep challenging assumptions, and we'll see you next time with another paper from the arXiv.

Jane: Goodbye, everyone. Stay curious.

Guido Fioretti

University of Bologna

eess.SY, cs.AI, cs.SY, nlin.AO

Submitted: 2026-08-15

Updated: 2026-08-18

Comments: 29 pages, 3 figures, 3 tables

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 40/100

Key concepts

Agent-based modeling: This involves building computer simulations of systems like societies or markets by defining the rules for individual simulated entities (agents) and observing their collective behavior over time. The paper uses this to test how different levels of agent intelligence affect system outcomes.
Lotka-Volterra model: A classic ecological model used to describe how predator and prey populations oscillate over time. In the episode, it is applied to business cycles and competition between firms to test whether simple or complex agents can achieve stable coexistence.
First-order derivatives: This refers to the ability of an agent to sense the rate of change in another agent's population—whether it is growing or shrinking. Fioretti argues that this simple predictive ability is sufficient for individuals to generate higher-order collective effects without needing complex understanding.
Extrapolation: The middle ground intelligence discussed is using simple linear extrapolation, which means looking at recent trends and projecting them forward. This simple prediction rule, when combined with interactions, leads to explosive growth in the model.

Terminology

Lyapunov function — 9× in this paper · explained in 2 episodes
Actor — 6× in this paper · explained in 3 episodes
Higher-order derivatives — 6× in this paper · explained in 2 episodes
Benchmarking — 4× in this paper · explained in 3 episodes
Cooperation — 4× in this paper · explained in 2 episodes
Agent-Based Simulation — 3× in this paper · explained in 1 episode
Core — 3× in this paper · explained in 2 episodes
Bifurcation Analysis — 2× in this paper · explained in 2 episodes
Fast and Frugal Heuristics — 2× in this paper
Metacognition — 2× in this paper · explained in 5 episodes
Abundance — 1× in this paper · explained in 1 episode
Adaptation — 1× in this paper · explained in 1 episode
Adaptive Strategy — 1× in this paper · explained in 1 episode
Agent-based modeling — 1× in this paper · explained in 1 episode
Chaotic Dynamics — 1× in this paper · explained in 2 episodes
Consistency — 1× in this paper · explained in 3 episodes
Generalization — 1× in this paper · explained in 13 episodes
Robustness — 1× in this paper · explained in 5 episodes

Summary

Summary

This paper investigates the debate within the agent-based modelling (ABM) community regarding the appropriate level of intelligence for artificial agents and how their local intelligence relates to the emergence of collective intelligence. The author approaches this debate by endowing the preys and predators of the Lotka-Volterra model with behavioral algorithms characterized by different levels of sophistication, ranging from the KISS principle (Keep It Simple, Stupid) to the KIDS principle (Keep It Detailed, Stupid).

The paper begins by reviewing theoretical insights and empirical evidence from various disciplines, including computational science, social sciences, neurosciences, group psychology, and psychoanalysis. From this survey, the author distills three propositions and three technical qualifications. Proposition 1 states that There exist settings where lack of individual intelligence is necessary to reach collective intelligence. Proposition 2 states that Individual intelligence, if employed in order to predict and coordinate behavior, can improve collective intelligence. Proposition 3, which comes in three versions, states that Single individuals may be able to envisage the collective intelligence of their organization. The three versions are: Proposition 3.1 (double-loop learning, where individuals can steer the organization), Proposition 3.2 (deutero-learning, where individuals understand but cannot steer), and Proposition 3.3 (where the majority of individuals agree to steer). The technical qualifications are: Qualification 1 (Element heterogeneity is necessary to reach collective intelligence), Qualification 2 (Given a stable equilibrium generated by a collective agreement, individual intelligence accelerates convergence), and Qualification 3 (Any degree of individual intelligence can trigger non-linear interactions that ultimately generate a substantial impact on collective intelligence).

The author uses the Lotka-Volterra prey-predator model, noting its relevance beyond ecology to applications such as the business cycle, technological substitution, ideological struggles, and market-share dynamics. The basic model has three outcomes: E1 (no preys, no predators), E2 (only preys, no predators, with unlimited growth), and E3 (preys and predators coexist in a limit cycle). The research question is whether individual behavioral algorithms can make E3 sustain itself. The author starts from Wilensky and Reisman's agent-based Wolf Sheep Predation model, which almost invariably ends in either E1 (41% of the time) or E2 (58.7% of the time), with E3 occurring only 0.3% of the time.

The author experiments with twelve behavioral configurations. Configurations 1, 2, and 3 involve reproduction proportional to the other species' fraction, but these fail to generate co-existence. Configurations 4, 5, and 6 involve collective agreements to stop reproducing if one's own population becomes larger than the other species. Configuration 4 (predators stop reproducing if their population exceeds preys) is highly successful, reaching E3 96.7% of the time. Configuration 5 (preys stop reproducing if their population exceeds predators) leads to E2 99.8% of the time. Configuration 6 (both stop reproducing) reaches E3 77.1% of the time. Configurations 7, 8, and 9 involve perfect foresight and suicide, but these are disastrous: configuration 7 reaches E2 100% of the time, while configurations 8 and 9 reach E1 100% of the time.

The most interesting results come from configurations 10, 11, and 12, which involve prediction based on linear extrapolation. Configuration 10 (predators reproduce with probability proportional to the variation of preys) reaches E3 91.6% of the time. Configuration 11 (preys reproduce with probability proportional to the variation of predators) reaches E1 99.5% of the time. Configuration 12 (both reproduce based on extrapolation of the other species) reaches E3 79.7% of the time, but with a remarkable novel outcome: both populations grow indefinitely while oscillating. The author notes that in case 12 co-existence of predators and preys does not imply a sustainable future for either species. Specifically, on average, the population of predators grows from 50 to 49,530.68 individuals whereas the population of preys grows from 100 to 7,453.59 individuals.

The author concludes that "The exploration of individual algorithms more sophisticated than those expressed in the basic Lotka-Volterra model to generate collectively more 'intelligent' behavior was in general unsuccessful, except for individual algorithms based on linear extrapolation." The author suggests that if most individuals compute first-order derivatives, they can impact one another generating higher-order derivatives at the collective level, but higher-order derivatives at the individual level would add little. The author also discusses the application of this finding to capitalism and economic growth, noting that economic growth can only set in once the vast majority of actors have acquired an extrapolation-based way of thinking. Finally, the author notes that whether explosive co-existence represents an instance of collective intelligence is an undecidable question within the model, as it depends on implicit predictions about long-run dynamics that are not contained in the model.

Improvements for AI systems

Based on the paper's findings, here are specific improvements for AI systems:

Improvement: Implement dynamic prediction windows that adjust based on system state, rather than fixed lookback periods.

What the improved system can do:

Detect when linear extrapolation becomes destabilizing (as shown in the paper's sensitivity analysis where Timespan=7 degraded performance) and automatically shorten the prediction horizon
Switch between prediction-based and reactive strategies when prediction accuracy degrades
Maintain coexistence in multi-agent systems where fixed-horizon predictors cause extinction or explosive growth

Improvement: Design AI systems where predictive capabilities are asymmetrically distributed, mirroring the paper's finding that predator intelligence (behavior 10) succeeds while prey intelligence (behavior 11) fails.

Improvement: Engineer systems where agents compute only first-order derivatives of others' behavior, allowing collective emergence of higher-order dynamics.

Improvement: Add safeguards that detect when agents' predictive capabilities lead to self-destructive collective decisions.

Improvement: Create a two-tier architecture: simple reactive agents for most functions, with a small subset of predictive agents that can steer the collective.

Improvement: Implement real-time monitoring that distinguishes between sustainable coexistence and explosive, unsustainable growth patterns.

Improvement: Add a validation layer that tests whether predictions improve outcomes before committing to prediction-based behavior.

Improvement: Train systems to recognize when individual first-order predictions create collective higher-order dynamics (snowball effects).

Improvement: Design AI systems where intelligence is concentrated at predator nodes (controllers, decision-makers) rather than prey nodes (workers, executors).

Improvement: Implement automatic adjustment of the Timespan parameter based on real-time system volatility.

Abstract

The agent-based modelling community has a debate on how ``intelligent'' artificial agents should be, and in what ways their local intelligence relates to the emergence of a collective intelligence. I approach this debate by endowing the preys and predators of the Lotka-Volterra model with behavioral algorithms characterized by different levels of sophistication. The main finding is that by endowing both preys and predators with the capability of making predictions based on linear extrapolation a novel sort of dynamic equilibrium appears, where both species co-exist while both populations grow indefinitely. While this broadly confirms that, in general, relatively simple agents favor the emergence of complex collective behavior, it also suggests that one fundamental mechanism is that the capability of individuals to take first-order derivatives of one other's behavior can allow the collective computation of derivatives of any order.