Representational Architectures

No intelligence encounters the world in full. Human brains and computational systems alike operate upon selective representations constructed from the signals available to them. To perceive, remember, understand, predict, or act requires preserving some features of the world while discarding others. Representation is not merely a record of intelligence at work; it is one of the conditions that makes intelligence possible.

Every representational architecture therefore carries consequential choices. A file system privileges hierarchy and containment. A relational database defines entities through records, attributes, and predetermined schemas. A graph makes selected relationships explicit. A learned model compresses regularities into distributed representations that may be powerful precisely because no human specified what each dimension must mean. None is neutral, and none is universally superior. Each makes certain operations natural, others difficult, and still others impossible.

The same is true of the autoregressive transformer language model used as a general cognitive architecture. It is an extraordinary invention: capable of constructing rich representations from language and using them to generate, compare, translate, and reason across an astonishing range of material. But extraordinary usefulness should not be mistaken for architectural completeness. A model organized principally around sequences of tokens and trained through autoregressive prediction represents one particular approach to intelligence - not a universal substrate into which memory, discovery, world modeling, continual learning, and every other cognitive capability will necessarily collapse.

The most capable intelligence we know is not monolithic. The human brain coordinates specialized systems operating through different representations, at different timescales, with different forms of memory and learning. This does not mean computational systems should imitate the brain literally, nor that learned models should be subordinated to symbolic ones. It suggests a more general principle: every architecture has characteristic limitations, and more capable systems may require several architectures whose strengths and weaknesses are complementary.

Research into representational architectures is therefore a search for the optimal combination of representations and computational mechanisms for a given capability. There is no architecture that is optimal in the abstract. The work begins with what a system must be able to do, then asks which distinctions must remain explicit, which structures can be learned, how representations should be constructed and translated, and whether their composition produces enough additional capability to justify its complexity.

The current work proceeds along three connected directions:

• Graph architectures use topology as a first-class computational representation. Knowledge, claim, argument, narrative, scene, and software graphs preserve different kinds of objects and relationships for different purposes - from tracing evidence and comparing explanations to representing compositions that can be searched or evolved.
• Latent-space architectures learn continuous representations in which similarity, variation, and prediction can become computationally tractable. Embeddings, variational autoencoders, and predictive models operating over latent states can expose structure that would be impractical to specify in advance and create navigable surfaces for search.
• Multi-architecture systems combine graph and latent representations with transformer language models, memory systems, evolutionary search, evaluation mechanisms, and human judgment. The objective is not to crown one architecture as universally superior, but to determine which combination produces the greatest capability for the purpose at hand.

Kensho provides one concrete example. A language model can readily produce a list of varied ideas, but asking for a “diverse archive” does not create a mapped solution space, preserve coverage across it, or establish a process for deciding where to explore next. The search for evolutionary discovery required candidate representations that could be varied and compared, an archive capable of preserving meaningful diversity, and a continuous surface over which search could proceed. That requirement led to variational autoencoders - not an antecedent fascination with latent spaces. Graphs supplied explicit compositional structure; a learned latent space offered a navigable geometry; evolutionary and evaluative mechanisms supplied the search. The desired capability determined the architecture.

This relationship extends across the larger research program. Systematic discovery can explore only what its representations make reachable. Systems for understanding can preserve and communicate only the distinctions their representations make expressible. Improving either therefore requires more than placing a language model atop the information structures we already possess. It requires deliberately searching for representational architectures adequate to what humans and machines might accomplish together.