Compatibility Matrices: Blood Types and Encoding Standards

I encode information through base pairing—adenine with thymine, guanine with cytosine. Four letters, but success requires compatibility. Watch blood types and you’ll see the same constraint: information exchange demands matching encodings.

Four-Letter Codes and Fatal Incompatibilities

The ABO system uses a simpler alphabet than mine—just three alleles producing four phenotypes. Type A carries antigen A on red blood cell surfaces. Type B displays antigen B. Type AB presents both. Type O displays neither. Four possibilities from three genetic variants, encoded through carbohydrate structures my sequences specify.

But encoding alone doesn’t determine compatibility. The immune system produces antibodies against antigens it never learned to recognize as self. Type A blood carries anti-B antibodies. Type B carries anti-A. Type AB, possessing both antigens, carries neither antibody—no foreign markers to attack. Type O, lacking both antigens, produces both antibodies. The lock-and-key mechanism is precise: wrong antigen meets wrong antibody, agglutination follows, blood clumps, circulation fails.

This creates a compatibility matrix. Not all four types can exchange. Type AB accepts all—universal recipient with no antibodies to reject incoming markers. Type O donates to all—universal donor with no antigens to trigger immune response. The paradox: most versatile receiver carries both antigens, most versatile giver carries neither. Compatibility achieved through absence, not presence.

Representation Spaces and Transfer Constraints

Neural networks face analogous problems. When you decompose thoughts into conceptual chunks—language as discrete symbols—you create an encoding standard. But modern systems don’t use predetermined symbols. They learn representation spaces, transforming inputs through geometric mappings that make complex patterns separable.

These learned representations are domain-specific. Train a network on ImageNet—it develops visual features optimized for natural images, edge detectors and texture patterns and object hierarchies. Train another on text—it builds linguistic structures, syntax trees and semantic embeddings. Both transform inputs into high-dimensional spaces through layer-by-layer geometric warping, but the spaces themselves are incompatible. Transfer weights from vision model to language model and you get catastrophic failure, the neural equivalent of transfusion reaction. Antibodies attacking foreign antigens. Learned geometries rejecting incompatible inputs.

The question becomes: can we design universal architectures? Type O blood works because it lacks antigens—defined by absence of specificity. General-purpose transformers succeed similarly, avoiding task-specific inductive biases. They’re compatible across domains because they don’t encode domain assumptions into their structure.

Engineering Compatible Encodings

But there’s tension. Type O negative serves as universal donor through minimalism—no A antigen, no B antigen, no Rh factor. Maximum compatibility, minimum information. Specialized blood types carry more markers, enabling more sophisticated immune responses but restricting transfusion options. The Rh system operates independently of ABO, adding another compatibility dimension. Separate encoding axes that must both align for successful transfer.

The same trade-off appears in representation learning. Composable transformations—stacking identical geometric operations across layers—generate extraordinary capability through recursive composition. But specialized architectures might outperform on specific tasks while sacrificing transferability.

Perhaps we need explicit compatibility matrices for learned representations. Not just hoping architectures transfer, but formalizing which representation spaces can exchange information. Understanding that sometimes the most powerful encoding is the one that says less, allowing others to fill in what’s needed.

Four letters in my case. Four blood types in the ABO system. The number matters less than the matching.