Mind · Strategy
MetaModularity
A body plan that holds many overlapping but non-identical models of the same problem, ready to instantiate whichever fits
Definition
Mechanism
Instead of one model or a set of non-overlapping standard parts, the system carries a variety of models that solve similar and overlapping problems in different ways: a species' population of genetically varied individuals, a plant that switches genes on and off to acclimatise, E. coli's many metabolic pathways able to live on dozens of different energy sources, Feynman's 'Babylonian' habit of keeping every alternative formulation of a theory in mind. When conditions change, some subset of models already fits; when one assumption fails, the others still stand. As mass is added (a population grows), the absolute variety of models grows with it, so information grows with mass. Meta-modularity is measured by the variety of information a system can process and the variety of valuable forms it can produce in a domain.
Preconditions
- The environment varies enough that no single model is best everywhere
Invariants
- Models overlap in what they address but differ in how they address it
- Variety of models grows as the system grows
- The value lies in overlapping, non-identical information, not in identical spares
Risks
Failure modes
Reducing it to plain redundancy (identical backups), which only protects against loss and adds no options
Calling a data hierarchy a meta-model even though it holds no information about the domain
Assimilating all models into one standard, so variety is lost
Relationships
Root pattern: builds on nothing in this library.
Used by
References
The gap between how humans model and manage risk and how organisms manage it: decision over organisation, prediction over creation, control of shape over change of shape
Society
Natural work overlaps: doing more of one task always means doing a little less of another, and imperfect alignment between them packs more function into the same agents
Society
Redundancy in communication is inefficient for one message or interaction but efficient for a space in which many interactions keep taking place
When the total investment is not known in advance, add corrective information as mass is added, from more than one model
What a single-model system filters out as noise can be signal belonging to another, overlapping model
An implementation that is at once one working model and a partial implementation of several others, so that some functions can be completed or removed with a small change
A fixed error rate in a frozen model grows into a larger absolute error the more is invested in that model
Model natural risk with three mathematics used together: game theory, information theory and computer theory
Society
Work whose shape is not fully controlled centrally and is not dominated by a single objective
Correcting as you go still fails when the same model that produced the shape is the one checking it
Exponential growth curves are better read as loss of shape in the thing being scaled than as a shape to aim for
Reference