Boids and the Vanishing Margin

1. The model

Craig Reynolds published boids in 1987 as a computer-graphics solution to an animation problem: how to move a flock without choreographing it. Each agent applies three steering rules against the neighbours inside a fixed perception radius.

• Separation. Steer away from neighbours that are too close.
• Alignment. Steer toward the average heading of neighbours.
• Cohesion. Steer toward the average position of neighbours.

That is the whole specification. There is no leader, no shared map, no representation of "the flock" anywhere in the system. Flocking is a property of the interaction, not of any participant — which is why boids became the standard teaching case for emergence, and why the same three rules keep reappearing in swarm robotics, crowd simulation, traffic modelling, and particle-swarm optimization.

2. What the simulation actually demonstrates

Established. Local rules with local information can produce coherent global structure without central control, and the structure is robust to losing individual agents. This is not a metaphor; it is a reproducible result you can perturb with the sliders above.

Established. The regime is a knife edge. Alignment near zero gives you a gas. Cohesion too high gives you a knot the flock cannot escape. Separation too high shatters it. The interesting band is narrow, and nothing inside the system is monitoring whether you are in it.

Not established. That real starlings implement these exact rules. Field studies of murmurations support topological neighbour rules — roughly seven nearest neighbours rather than a metric radius — which is a different model with different stability properties. Boids is a sufficiency proof, not a description of a bird.

Not claimed. That emergence explains cognition, or markets, or civilisation. The demonstration is narrow and it earns only what it shows: coordination does not require a coordinator.

3. The counter-entry

Roman Yampolskiy puts the window in blunt terms: a modal estimate of two to three years to superintelligence. Below human level, we can still evaluate and correct. Above it, the models increasingly generate the feedback that trains their successors. Character training, scalable oversight, and monitoring have no strong guarantee of surviving that phase shift. A capabilities failure is visible and repeatable. A serious alignment failure may be neither.

The careful moment to apply the brakes, he says, already lies in the past. The remaining useful moment is now. And the sharpening is in the phrasing: the opportunity itself is irreversible and nontransferrable. Once a threshold is crossed, the option to have slowed earlier cannot be recovered, and it cannot be handed to a later actor. There is no secondary market in past caution.

Why the boids belong on the same page

Two reasons, one structural and one that has to be stated carefully.

The structural one: a flock has no unit capable of deciding to slow down. Deceleration is not in the rule set. Each agent is doing something locally reasonable — match the neighbours, do not collide, stay with the group — and the aggregate trajectory belongs to no one. That is a fair description of a competitive capability race in which every individual decision is defensible and the sum is not steered by anybody. It is also why "someone will stop this" is not an argument; in a system of local rules, there is no someone.

The one to state carefully: the analogy does not license a prediction. Boids has three parameters and no adversaries. AI development has firms, states, incentives, regulators, and agents that model each other — which is precisely the machinery boids omits. The model gives you an intuition about the absence of a steering unit. It gives you nothing about timelines, and no timeline in this essay is derived from it.

4. The arithmetic of the closing window

The successive-squaring exercise from earlier in this sequence — 2 to 18,446,744,073,709,551,616 in six steps — is the reason the margin behaves the way it does. Ordered surplus that compounds produces both the beauty and the compression. The same growth curve that lets one person match a fifty-person company also shortens the interval in which any deliberate choice is still available.

This is the counter-entry in the accountant's ledger: acceleration in one column, vanishing margin in the other, and a note that some entries cannot be rewritten. The books stay open. That is not optimism; it is the honest state of the account.

And a refusal, since it is easy to slide here: nothing about this is ordained. A closing window is a description of a rate, not a destiny. The rate is produced by decisions with budgets attached, and decisions with budgets attached can be made differently. Claiming inevitability would concede the argument on arrival.

5. Falsifiers

• If a boids implementation with correct parameters fails to produce stable flocking, the sufficiency claim in section 2 is wrong and the simulation above is mis-specified.
• If field data shows starling coordination requires a leader or global signal, the "no coordinator" reading loses its cleanest biological illustration.
• If capability growth turns out to be sublinear in compute and data over the next several years, the compression claim in section 4 is overstated and the window framing should be softened.
• If oversight methods demonstrably transfer across a capability jump — verified on systems more capable than their evaluators — then the irreversibility claim weakens, because the option would in fact be transferable.
• If someone shows that a decelerating coalition can form and hold inside a competitive field, the "no unit can decide to slow" reading is refuted, and that would be the most welcome refutation in the essay.

Return to the essay index — or continue to The Proxy That Ate the Purpose.