Is model breeding a standard technical term?
It is an understandable umbrella metaphor, but established literature more often uses precise terms such as evolutionary model merging, neuroevolution, architecture search, population-based training, model soups, distillation, adapter composition, and quality-diversity. Use “model breeding” as a category name, then name the exact operator.
Does a model breeder train foundation models from scratch?
Not necessarily. The most practical systems create descendants from approved parents through adapters, fine-tuning, distillation, quantization, pruning, routing, or composition. Training from scratch is one expensive operator among many.
Does the system need a database?
No. A small or read-heavy implementation can use immutable files, manifests, content-addressed directories, and append-only records. A database becomes useful when concurrent writes, large telemetry volumes, transactional workflows, or complex authorization justify it. This website intentionally demonstrates a file-backed content architecture.
Can model breeding happen online in production?
Candidate generation can be continuous, but unrestricted in-place production mutation is difficult to audit and roll back. The recommended enterprise starting point is offline generation with shadow, canary, and explicit promotion. Online routing adaptation is safer than online weight mutation when bounded and observable.
Is a mixture of experts the same as a model ecology?
No. A mixture-of-experts model typically has internal experts trained and served as one architecture. A model ecology can include independent artifacts, runtimes, owners, contracts, release cycles, and deterministic components.
Can any two models be merged?
No. Direct parameter merging normally requires close compatibility in architecture, tensor shapes, tokenizer, normalization, and training lineage. Heterogeneous models are more safely combined through contracts, routing, ensembles, or distillation.
Why keep an ensemble if a merge is smaller?
The ensemble is a behavioral baseline and may preserve complementary strengths that a merge destroys. The merge must demonstrate enough resource or deployment benefit to justify any quality loss.
What prevents runaway model population growth?
Explicit complexity and resource budgets, minimum promotion margins, package leases, retirement reviews, duplicate detection, bounded archives, and no-op as a valid action. Every active specialist must justify ongoing cost.
Should the model control its own evaluator?
No. Candidate generation and evaluator governance must be separate authority domains. Allowing a candidate to modify its own acceptance criteria creates a direct path to reward hacking.
How is teleodynamics used here?
As a control-system analogy: structural organization persists only when it continues to repay resource and risk costs. The site does not claim that software is biologically alive, conscious, afraid, or intrinsically self-preserving.
What is the smallest useful model-breeding project?
One approved parent, one bounded operator such as quantization or adapter training, a frozen evaluation suite, an immutable lineage manifest, a champion–challenger comparison, and a manual promotion decision. That is enough to establish the core discipline.
What should be automated first?
Artifact hashing, manifest validation, reproducible evaluation, resource profiling, and rollback. Automate candidate generation only after evidence and release controls are reliable.
How should safety be scored?
Critical safety and security conditions should be hard gates, not merely negative terms in an aggregate fitness score. Aggregate scores are appropriate only after non-negotiable invariants pass.
Can a small specialist outperform a much larger model?
On a narrow, well-defined task, yes. That does not imply general superiority. Compare under the same contract, dataset, hardware, and failure requirements.
How do you prevent deskilling?
Design for capability transfer, explanations, verification, periodic unassisted operation, user export, and graceful fallback. Measure whether people and institutions remain capable when the AI is unavailable.
What is a good first deployment pattern?
An offline champion–challenger factory with immutable packages and a deterministic or simple router. Add learned routing, federation, and structural automation only when measured needs justify them.
Source reports used for this guide
These reports are preserved verbatim in the site archive. The guide above is an editorial synthesis and may narrow, qualify, or reorganize claims from the source material.