# Foundation Model A **foundation model** is a large AI model trained on broad, diverse data at massive scale that can be adapted (fine-tuned, prompted, or otherwise specialized) for a wide range of downstream tasks. The term was coined in a 2021 Stanford HAI report to describe the paradigm shift from task-specific AI (a model trained to do one thing) to general-purpose AI (a model trained on everything, then steered toward specific uses). Examples: GPT-4, Claude, Gemini, Llama, Mistral, [[Stable Diffusion]]. Each is trained on enormous datasets (internet text, images, code) and can then be applied to translation, summarization, coding, image generation, scientific reasoning, or military targeting — depending on who deploys it and how. ## Why the Term Matters The foundation model paradigm creates a **power structure**. Building one requires: - **Compute:** Hundreds of millions to billions of dollars in GPU clusters. Training GPT-4 reportedly cost $100M+; frontier models in 2026 cost $1B+. - **Data:** Trillions of tokens of text, images, code — scraped from the open internet, licensed from publishers, or generated synthetically. - **Talent:** A few hundred researchers and engineers with expertise concentrated at a handful of labs. This means only a small number of organizations can build foundation models from scratch: OpenAI, [[Google DeepMind]], Anthropic, Meta, Mistral, and a shrinking list of others. Everyone else — governments, militaries, businesses, individuals — is a *user* of foundation models built by these organizations. The foundation model paradigm concentrates the means of AI production in fewer hands than any previous technology platform. ## The Dual-Use Problem A foundation model is inherently [[Dual-Use AI]]. The same model that writes poetry can generate propaganda. The same model that helps doctors diagnose diseases can help militaries identify targets. The same model that tutors students can be fine-tuned for surveillance. This is not an accident — it is a structural feature of the "foundation" concept. Because the model is general-purpose, its applications are limited only by the intent and infrastructure of the deployer. The KB's documentation of [[Llama Military Adoption]], [[Project Maven]], and the Anthropic-Palantir partnership all flow from this structural feature. ## KB Significance 1. **The new means of production.** Foundation models are to the AI era what oil refineries were to the industrial era and what semiconductor fabs are to the information era — the capital-intensive infrastructure that determines who can produce and who must consume. The companies that build foundation models (OpenAI, Google, Anthropic, Meta) and the companies that provide the compute to train them (NVIDIA, Microsoft, Amazon, Google) constitute the new industrial oligarchy documented across the News logs and the defense-tech sections of this KB. 2. **The open vs. closed debate.** Whether foundation models should be open-weight (Meta's Llama, Mistral) or closed (OpenAI's GPT, Anthropic's Claude) is the defining governance question of AI. Open-weight models democratize access but eliminate the deployer's ability to prevent military, surveillance, or extremist use. Closed models maintain control but concentrate power. This debate maps directly onto the KB's broader power-concentration analysis. 3. **The defense-tech bridge.** Every defense-tech AI system documented in this KB — [[Palantir AIP]], Anduril's targeting systems, [[Scale AI]]'s data labeling — runs on foundation models built by commercial AI labs. The foundation model is the bridge between Silicon Valley and the Pentagon. ## See Also - [[Frontier Model]] — The most capable foundation models at any given time - [[Large Language Model]] — The text-specific variant - [[Scaling Laws]] — Why foundation models keep getting bigger - [[Open-Weight AI]] — The access-vs-control debate - [[Dual-Use AI]] — The military application problem