Knowledge
Knowledge HubConcepts and strategic researchGlossaryDefinitions across six domainsFrameworksStructured strategy methodsTopicsConnected subject areasLibrary
PlaybooksPractical strategy guidesBooksBooks and long-form researchOperator VaultMember resources and toolkits
Explore Discover, the Knowledge Hub, playbooks, and the Operator Vault.
Glossary
Foundation model
A foundation model is a large machine-learning model trained on broad data — typically at scale — that can be adapted to a wide range of downstream tasks rather than being purpose-built for a single application. Adaptation may include prompting, fine-tuning, retrieval augmentation, or tool use. The term is associated with the Stanford CRFM framing of models that serve as a general base for many applications.
Why it matters
Enterprise AI programs often confuse a single chat product with the underlying model class. Foundation models matter because they concentrate capability, cost, and risk: one base model (or a small set) can power search, drafting, coding assistance, classification, and agent workflows — which concentrates vendor dependency, evaluation burden, and governance decisions.
A foundation model is broader than a large language model. LLMs are the most visible foundation-model family for text, but foundation-model language also covers multimodal and other large pretrained systems adapted across tasks. Strategy and architecture choices should name the model class and the adaptation pattern separately.
Example (enterprise): A company standardizes on one frontier multimodal foundation model for customer-support drafting, internal knowledge Q&A, and code assistance, then differentiates applications with retrieval corpora, prompts, and tools rather than training a separate model for each use case. The shared base creates leverage and a single point of evaluation for safety, cost, and quality.
Mechanisms
- Broad pretraining — large-scale training on diverse data creates transferable representations and behaviors usable across tasks.
- Adaptation layer — prompting, fine-tuning, adapters, RAG, and tools specialize the base without rebuilding capability from scratch.
- Shared risk surface — security, bias, IP, and cost controls attach to the foundation layer and propagate to every dependent application.
Boundaries
- Foundation model ≠ a finished product or “AI strategy.” Products wrap models with UX, data, permissions, and workflows.
- Foundation model ≠ large language model as exact synonyms. LLMs are a major subclass; foundation-model language can include other modalities.
- Having access to a foundation model does not imply a proprietary data moat; differentiation usually comes from data, retrieval, workflow design, and governance.
Related topics
None linked yet.
Related concepts & frameworks
Related Discover collections
Further research
No related resources yet.