Carrie Bickner

In our common way of describing the world, we all intuitively understand what we mean when we say Leaves of Grass or ChatGPT. These are intellectual units—discrete entities of thought, creation, and knowledge that are recognized as cohesive objects in their respective contexts.

But intellectual units, whether books or software, are rarely static. Just as Walt Whitman revised and expanded Leaves of Grass over multiple editions, AI models like Claude or ChatGPT are constantly updated, iterated upon, and customized for new contexts. The question of how we describe, track, and preserve these units across their lifecycles is at the heart of the MARTI Framework (Metadata for AI Responsibility, Transparency, and Integrity).

The Inspiration: Books, Software, and Intellectual Units

The connection between versions of software and editions of books crystallized for me while listening to Lex Fridman’s November 11, 2024 interview with Dario Amodei. At minute 38, Lex asked about the differences between various versions of Anthropic’s AI model, Claude. The way Dario described the updates and evolution of the model mirrored the way new editions of books emerge: each iteration reflects improvements, changes, and the evolving intentions of its creators.

This parallel solidified the core idea for this project: how can we use metadata frameworks to describe and preserve intellectual units across domains, from books to AI? The MARTI Framework provides a structure for answering this question, emphasizing provenance, structural metadata, and domain-specific adaptability.

Cataloging Books: An Illustration of Metadata Relationships

Let’s start with a familiar intellectual unit: Walt Whitman’s Leaves of Grass.

If you search your local library catalog, you’ll find the bibliographic record for Leaves of Grass: The Death Bed Edition. This record acts as a “parent,” describing the work in general terms—its title, author, publication details, and themes.

But this bibliographic record branches into specific “children”: item records. In this catalog, there are five item records for the so called Death Bed Edition. Each one represents a specific physical copy of the book, with unique metadata such as barcodes, condition, and history.

This distinction between bibliographic and item records reflects structural metadata: the relationships between objects. It demonstrates how metadata frameworks capture the dual nature of distinct intellectual units—the conceptual whole and its individual manifestations.

Expanding Metadata to AI: Intellectual Units and Their Iterations

The same principles of structural metadata apply to AI-generated objects. Consider ChatGPT-4.0:

• The parent intellectual unit is the core model, which encompasses the architecture, training data, and overarching capabilities.
• The children are the instances of ChatGPT-4.0 deployed on specific devices or applications, each with its own unique attributes and context.

Using MARTI’s morphology, these relationships can be encoded and preserved just as we document the bibliographic and item-level metadata for Leaves of Grass. By treating AI models as intellectual units, MARTI ensures that each version or instance can be tracked across its lifecycle, preserving its identity, context, and evolution.

The MARTI Morphology: A Framework for Intellectual Units

The MARTI Framework provides tools to describe intellectual units and their iterations across domains. In its Framework Overview, MARTI emphasizes the following principles:
Provenance: Documenting the origin and evolution of an intellectual unit.

Structural Metadata: Encoding relationships between parent and child objects, such as editions of a book or versions of an AI model.

Domain-Specific Adaptability: Allowing metadata frameworks to adapt to the unique needs of different fields, from libraries to AI development.

By applying these principles, MARTI makes it possible to describe both tangible objects like books and intangible digital creations like AI models within a unified framework.

Why Intellectual Units Matter

Whitman’s Leaves of Grass offers a vivid example of the intellectual unit in action. Over his lifetime, Whitman published six (or nine) editions of the book, each reflecting his evolving vision. Yet despite these differences, all editions remain connected through their bibliographic identity.

AI models operate similarly. Each version of an AI model, like Claude or ChatGPT, is part of a larger intellectual unit, tied together through provenance and metadata. Without a framework like MARTI, these relationships risk being lost, making it harder to trace the origins, updates, or adaptations of the model.

Where Metadata Meets Meaning

Intellectual units, whether literary works or AI models, are connected by intricate relationships that define their identity and context. Like the branches of a tree, these relationships extend outward, linking a core concept to its specific iterations.

Walt Whitman’s Leaves of Grass offers a compelling example. Each edition represents a branch of his creative vision, while individual copies carry the marks of their own histories. Similarly, in AI, parent models like ChatGPT-4.0 give rise to unique instances, each shaped by its context and use.

The MARTI Framework provides the structure to document these relationships. By focusing on provenance, structural metadata, and adaptability, MARTI ensures that intellectual units remain coherent and traceable across their lifecycles. Whether mapping the editions of a book or the versions of an AI model, MARTI helps us preserve the connections that make these units meaningful.

From haikus to sonnets, from books to algorithms, the MARTI morphology offers a way to understand and preserve the evolving forms of knowledge.

O I say these are not the parts and poems of the body only, but of the soul,
O I say now these are the soul!
(Whitman, 1855. https://whitmanarchive.org/item/encyclopedia_entry9)