Skip to main content
biology business computer-science economics

​
Seeding

​
Description

A small initial input that disproportionately determines the emergent shape of what grows from it. The seed is small but its choice constrains the emergent structure of everything that follows — not just the local outcome, but the topology of the developmental path. The diagnostic property: the same growth dynamic applied to different seeds produces qualitatively different mature systems, often in ways the seed’s surface size doesn’t telegraph. The concept is distinct from bootstrap (broader; bootstrap covers any self-starting process, where seeding is specifically the initial-input subcase) and from scaffolding (temporary support that gets removed once structure stands; seeding’s effect is permanent shape-determination). Seeding is also distinct from load-bearing in degree: load-bearing asks “would removal change observable behavior?”; seeding asks “would different choice produce qualitatively different growth?”

​
Triggers

User-initiated: User describes a small initial choice with disproportionate downstream consequences, or asks “what initial state should we start from?” Vocabulary cues: “seed,” “initial,” “starter,” “founding,” “day one,” “garbage in garbage out.” Agent-initiated: Agent notices a system has a small initialization input whose choice shapes the emergent behavior across many subsequent steps. Candidate inference: “what does the seed lock in; what alternative seed produces meaningfully different growth?” Situation-shape signals: Path-dependence discussions. Questions of “what should we put in version 1” knowing version N will inherit the shape. Discussions of priors, initial conditions, or starting state.

​
Exclusions

  • Memoryless / re-startable systems — when state is regenerated from scratch each cycle, “seed” is a misnomer; there’s no path-dependence.
  • Initialization that doesn’t propagate — a one-time setup that doesn’t shape subsequent dynamics; the seed is decorative, not constitutive.
  • Convergent systems — domains where the long-run behavior is independent of initialization (ergodic Markov chains converge to the same stationary distribution regardless of starting state); the concept fires in transient but not in the limit.

​
Structure

Internal structure of seeding: a table of its component slots and the concepts that fill them.

​
Relationships

Relationship neighborhood of seeding: a graph of the concepts it connects to and the concepts it is a part of.
  • shape — the seed determines emergent shape; seeding is the concept for “shape via small-initial-input.”
  • load-bearing — seed choice is high-leverage; small wrongness compounds. The load-bearing test on a seed is “would a different seed produce qualitatively different downstream behavior?”
  • trigger-rule-pair — seed = initial condition; growth dynamic = rule; together they’re a trigger/rule pair.
  • graduation-promotion — seeding sets the initial scaffolding; graduation-promotion is the move from scaffolding to adult form. Seeding is what the scaffolding contained; graduation is the transition.
  • hysteresis — seeding produces hysteresis: once seeded, the system’s behavior depends on that path, and re-seeding requires substantial rework.

​
Examples

Random number generator seeds · computer-science

same RNG, different seeds, different reproducible sequences; foundational for testing.

Fermentation starter cultures · biology

sourdough, kombucha, yogurt; the starter determines strain composition for every subsequent batch.
initialization determines convergence basin; different seeds → different cluster topologies.
the initial schema + reference rows shape every downstream query, migration, and code path.
early hires + early bylaws shape culture for decades; “founder’s syndrome” is the path-dependence pathology.
In deep learning, the choice of initial weights is a canonical instance of seeding. Glorot and Bengio’s 2010 paper introduced what is now called Xavier initialization — sampling initial weights from a distribution scaled by the layer’s fan-in and fan-out — to keep activation variances stable across layers during early training. He et al.’s 2015 paper adapted the same idea for ReLU networks (He initialization), with a different variance scaling that accounts for ReLU’s asymmetric activation.The role of these initializations as seeds matches the catalog’s diagnostic: the same training procedure applied to a network with poorly-scaled initial weights converges much more slowly (or not at all) than the same network with appropriately-scaled init, even though every other ingredient is identical. The choice of seed is small in scale (a one-time draw of weights at step 0) but constrains the entire downstream trajectory — exactly the structural shape small initial input, disproportionate emergent consequence the seeding concept names.
Paul David’s 1985 American Economic Review paper “Clio and the Economics of QWERTY” and Brian Arthur’s 1989 Economic Journal paper “Competing Technologies, Increasing Returns, and Lock-In by Historical Small Events” together established path dependence as a recognized phenomenon in economics: market outcomes are not determined solely by present-day cost and benefit but also by historically contingent choices that, once made, are reinforced by increasing returns until alternatives become structurally infeasible. The QWERTY keyboard layout is the canonical case — chosen for typewriter-mechanical constraints that vanished a century ago, persisting because the cost of coordinating a switch across all keyboards, training, and software has come to exceed the benefit of any specific alternative layout.Arthur formalized the mechanism with increasing-returns-to-adoption models showing that early adoption advantages compound through network effects, learning curves, and complementary asset investments. The early-and-small choice — the seed — gets locked in not by its intrinsic merit but by the dynamics that follow it.Inference: When evaluating any large system’s apparently-suboptimal current state, the question to ask is not “why was this chosen?” but “what was the choice environment at the time the seed was set, and what feedback dynamics made it stick?” Trying to argue the system into a better state by appealing to present-day cost-benefit usually fails because the lock-in mechanism does not respond to local arguments — it is sustained by structural forces (coordination costs, sunk investments, network effects) that operate at a level above any single actor’s choice.