Concept · foundation-model

Foundation model

A single large pre-trained model that serves as a general-purpose base for many downstream uses.

◷ Share view, as of June 2026: confidence, rationales and priors are hidden. The questions, stances, claims, tests and evidence stay visible.

Settled

Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.

What it is

Older AI built one model per task. A foundation model inverts that: one very large pre-trained model is adapted, by prompting, fine-tuning, or tool wiring, to thousands of tasks it was never specifically built for. "Frontier model" means a foundation model at the current capability edge.

The industrial structure of the field follows from this idea: a few labs spend enormously to train general-purpose bases, and an application layer builds on top. Whether value accrues to the foundation layer or the application layer is one of the Atlas's central economic questions.

Understand first

pre-trainingPre-training · The compute-heavy first phase where a model learns general capabilities by predicting tokens across vast text corpora.transformerTransformer · The neural-network architecture, built around attention, that underlies virtually all modern AI models.

Builds toward

agentAgent · A system that runs a model in a loop: planning, calling tools, and reacting to results to pursue a goal.benchmarkBenchmark · A standardized test set used to score and compare model capabilities.distillationDistillation · A training technique that teaches a smaller student model to mimic the outputs of a larger teacher model, producing compact models with disproportionately high capability.emergent-capabilityEmergent capability · An ability that appears in larger models without being present, or predictable, in smaller ones.fine-tuningFine-tuning · Further training of a pre-trained model on a narrower dataset to specialize or steer its behavior.hallucinationHallucination · Confidently produced output that is fluent but factually wrong or fabricated.model-commoditizationModel commoditization · The process by which competitive pressure and open-weight alternatives erode the pricing power of any individual AI model or lab.open-weight-modelOpen-weight model · A model whose trained weights are publicly released, allowing anyone to run, fine-tune, or modify it without API access or per-token payment.ragRAG (retrieval-augmented generation) · Pairing a model with a search step that retrieves relevant documents into its context before it answers.switching-cost