EvolutionStrong architectural inferencev1.22.1

In plain English

This page explains how AI systems can change over time through updates, tests, retraining, memory, and approvals even when no single model rewrites itself.

  • Why this matters: AI risk can come from the whole arrangement, not one obvious model.
  • What to look for: data, memory, routes, adapters, tools, evaluators, updates, and rollback paths.
  • Technical version below: the expert terminology remains available and is linked through the glossary.

Evolution

AI systems can change over time even when no single model rewrites itself. Updates, tests, retraining, memory, and human approval processes can still create a selection loop where some behaviors survive and others disappear. Feed means the system receives data or feedback. Fork means a candidate is created. Fight means candidates are tested. Flee means failed candidates are removed or bypassed. The decision not to change the system. Open glossary definition means the system correctly chooses not to change.
Feed → Fork → Fight → Flee, with No-op preserved as a valid control action

The diagram shows a bounded evolutionary loop. Feed supplies evidence and resources. Fork creates candidates. Fight evaluates candidates. Flee removes unsafe or wasteful states. No-op is shown at the center to preserve non-growth as a legitimate outcome.

animated flowchart · transition graph

The unsafe unit can be the transition graph.

The relevant safety boundary includes every permitted transformation, not only the current model artifact.

01Fine-tune new local behavior
02Attach LoRA small deltas carry strategy
03Merge capabilities recombine
04Route capability appears only on path
05Evaluate metric becomes selection
06Promote alias changes identity
07Persist memory and descendants retain residue
08Rollback must restore ecology, not one file

A changing AI system made from many connected parts, not just one model. Open glossary definition can behave evolutionarily even when every model artifact is immutable at runtime and every release requires human approval.

Evidence levelStrong architectural inferenceTechnical label: Architectural inference

Variation can be created by an external pipeline. Evaluation can be performed by independent gates. Selection can preserve the artifacts that score well. Inheritance can occur through fine-tuning, merging, distillation, synthetic data, or routing rules. Succession can replace a carrier while retaining a behavior.

Read the flagship page: Nothing Has to Reproduce Itself.

Added evolution guides

New evolution expansion

v1.8.0 report-driven pages

v1.21.4 controlled-evolution additions

v1.21.9 model-breeding operators

v1.22.0 risk-side evolution