AI models and organisational knowledge

Selected third-party reading to accompany What the company knows, Open models. Competing ambitions. and Some assembly required.

The Signals examine how organisational knowledge becomes useful through AI, whose interests shape the result and how businesses develop value from available models. The companion explains models, the software and services around them, and the choices involved in using them. This collection supplies additional vocabulary and technical depth for readers who want to examine the choices behind a proposal.

The companion provides short definitions of LLMs, weights, hosting and open weights. Start with the first two entries here for more depth on those distinctions. The later entries explain ontology and the software around a model. There is no expectation that a board member should read every technical section. Inclusion identifies useful material, rather than endorsing every claim or the publisher’s products.

Understanding the model

What are large language models?

Publisher and provenance: IBM (educational material from an AI supplier).

Best for: Readers seeking a first explanation of tokens, training, inference and the difference between adapting a model and supplying information for a response.

What it adds: The sections on using and deploying LLMs explain how retrieved information enters the model’s context without retraining it, and how applications connect models to other software. These distinctions help a reader ask where company knowledge is being held.

Limitations: This is a simplified introduction with product promotion. Its opening overstates the novelty of language processing, and its descriptions of model architecture should not be treated as universal. It is useful vocabulary, not an assessment of a proposed system’s reliability.

Open Foundation Models: Implications of Contemporary Artificial Intelligence

Publisher and provenance: Georgetown University’s Center for Security and Emerging Technology (research and policy analysis, March 2024).

Best for: Readers trying to distinguish access to weights from access to training data, development methods and reuse rights. Begin with the “Terminology” section.

What it adds: Explains why openness has several dimensions. Being able to obtain weights creates deployment and adaptation options, but licences, supporting components and documentation still matter. The wider discussion considers how different interests change the assessment of benefits and risks.

Limitations: The policy setting and model examples are from 2024. Use it for the distinctions and competing perspectives, not as a current account of regulation or a model-buying guide. It predates the Open Source Initiative’s later AI-specific definition.

Understanding differences between hosting services

Endpoint Accuracy Index: same model, different accuracy

Publisher and provenance: Artificial Analysis (a company that evaluates AI models and services). The index launched on 4 August 2026; the methodology explains its tests and reference deployments.

Best for: Readers comparing services that offer the same named model, or deciding how to run a model themselves.

What it adds: Shows how response limits, numerical precision and other operating choices can change results. The comparison is with the same model run using its developer’s recommended configuration, rather than with a different model.

Limitations: Results are dated snapshots, with statistical uncertainty, from selected tests. They do not establish performance on a business’s own work or the total cost of completing it. The configuration actually used still needs testing.

Understanding the software around the model

Ontology Development 101: A Guide to Creating Your First Ontology

Publisher and provenance: Natalya F. Noy and Deborah L. McGuinness, Stanford University (academic development guide).

Best for: Readers who want to understand ontology independently of Palantir’s product. The opening three pages explain the purpose and basic concepts; the later methodology is for practitioners.

What it adds: Shows how explicit concepts, relationships and assumptions support shared understanding and reuse. It also explains that an ontology supplies knowledge for applications; it is not itself the whole application. The wine-and-food example makes the modelling choices visible.

Limitations: A foundational guide using older tools and terminology. It does not address modern LLMs, Palantir’s operational implementation or the contractual rights needed to move a working service.

Effective context engineering for AI agents

Publisher and provenance: Anthropic’s Applied AI team (engineering guidance from a model developer, September 2025).

Best for: Leaders and technical colleagues examining what must be built around an LLM to make it useful over sustained work.

What it adds: Explains the roles of instructions, tools, retrieved information and persistent notes. It makes clear why supplying more material is not automatically better, and why useful knowledge can be maintained outside the model’s weights.

Limitations: Written for builders, with Claude examples and a supplier’s view of the design choices. Context management does not by itself establish data ownership, enforce access permissions or decide whose priorities should govern an action.

Palantir’s own architecture and model documentation

Read: Ontology overview and AIP Model Catalog.

Publisher and provenance: Palantir (product documentation).

Best for: Readers needing to distinguish the general concept of ontology from Palantir’s implementation, or to understand how its platform uses LLMs.

What it adds: The overview describes the connection between organisational concepts and operational functions. The catalog separates the model’s creator from the company serving it. This makes the layers in an actual supplier arrangement easier to identify.

Limitations: Documentation of a particular platform, not independent evidence of business returns or superiority. Model availability depends on the deployment. These pages do not establish that an ontology export will run elsewhere or that the customer has secured the necessary reuse rights.