Mind · Strategy
MonoModularity
A body plan in which one model processes every situation, treating each situation as just more data for the same process
Definition
Mechanism
A single model, algorithm or decision procedure is applied across all the environments it meets, and the particulars of each problem enter only as input data: one Bayesian update rule for every inference, one language model for every topic, one decision-maker freezing one plan at one moment. Because processing is uniform it scales well and pairs well with standardised lower-level parallel hardware. But it has little room to take in genuinely new information: changing how anything is done means changing the whole model at once, and systematic errors in the model are not corrected by feeding it fresh data. By the No Free Lunch theorems, no single model beats specialists across all problem classes.
Invariants
- The same model is used regardless of domain; domains differ only as data
- A systematic error in the model is not corrected by updating its inputs
- Gains on one class of problems are offset by losses on another (No Free Lunch)
Risks
Failure modes
Applying the model far outside the conditions it was built for, so its error grows with use
Mistaking a hierarchy of data types for a meta-model of the domain
Correcting the model with the same perspective that produced its errors
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
What a single-model system filters out as noise can be signal belonging to another, overlapping model
A fixed error rate in a frozen model grows into a larger absolute error the more is invested in that model
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
Reference