What are foundation models?
Foundation models are large, broadly trained models built to serve as a general-purpose starting point for many different downstream applications, rather than being trained from scratch for one narrow task. The term covers more than just language models, it includes models trained on images, audio, code, and combinations of these modalities together, and what unifies them isn’t the specific kind of data they were trained on but the underlying idea: train once on an enormous, broad dataset, producing a model with wide general capability that can then be adapted, through further fine-tuning or simply through how it’s prompted, to a huge range of narrower tasks without needing to be built from the ground up each time.
Why the term “foundation model” describes something distinct from “large language model”
Large language model specifically refers to a model trained on text, predicting and generating language, while foundation model is the broader category that a large language model happens to be one prominent example of. A model trained to understand and generate images is a foundation model but not a large language model, and a model trained to work across multiple modalities at once, text, images, audio, together, is a foundation model built on a broader premise than any single-modality large language model could be. Every large language model currently in wide use is a foundation model, but not every foundation model is a large language model, which is why the two terms, while related and often used somewhat loosely in casual conversation, actually point at genuinely different scopes.
Understanding this distinction matters because it clarifies what’s actually novel about the foundation model concept, it’s not specifically about text or language at all, it’s about a training approach and a resulting capability profile, broad, general, adaptable, that happens to apply just as well to other kinds of data as it does to text, and the rapid progress in multimodal AI systems traces directly back to applying this same foundational approach beyond the text domain it was first popularized in.
What “broad pretraining” actually means and why it produces such general capability
A foundation model’s defining characteristic is that it’s trained on an enormous, deliberately broad dataset before any task-specific adaptation happens, exposing it to an extremely wide range of patterns, styles, topics, and structures rather than a narrow, curated dataset built around one narrow application. This breadth is precisely what gives a foundation model its most valuable property, general capability that transfers usefully across many different downstream tasks, rather than narrow competence at exactly one thing and nothing beyond it.
This general capability is what separates a foundation model from the narrower, task-specific models that dominated AI development before this approach became standard, a model trained from scratch specifically to classify one particular category of image, for instance, would need to be built entirely anew for a different classification task, while a foundation model’s broad, general capability can be adapted to a wide range of related tasks without starting from nothing each time. This reusability is a large part of why foundation models have become the dominant paradigm across so much of modern AI development, the upfront cost of broad pretraining gets amortized across an enormous number of downstream applications rather than being paid separately for each one.
The two main ways a foundation model actually gets adapted to a specific task
Once a foundation model exists, there are two broad paths to applying it toward a narrower purpose. Fine-tuning involves further training the model’s own parameters on a smaller, task-specific dataset, adjusting its behavior directly to specialize it for a particular application while retaining much of the broad capability it gained during its original, general pretraining. Prompting involves using the model exactly as it already is, with no further parameter adjustment at all, and instead shaping its behavior for a specific task purely through how a request is phrased, what instructions and examples are included alongside it.
These two approaches trade off differently, fine-tuning can produce more consistent, more deeply specialized behavior for a narrow task but requires real engineering effort and a dataset built for that purpose, while prompting requires no additional training at all and can be set up quickly, but the resulting behavior stays more general and less deeply specialized than a properly fine-tuned model would be. Many real applications end up using a combination of both, relying primarily on prompting for flexibility while reserving fine-tuning for the specific parts of a task where deeper specialization genuinely matters.
Why the economics of foundation models favor a small number of large training efforts
Training a genuinely capable foundation model from scratch requires an enormous amount of computational resources, data, and specialized expertise, a cost that’s only economically justified when the resulting model can then be reused across a very large number of downstream applications rather than being built for just one. This economic reality is exactly why the field has consolidated around a relatively small number of organizations capable of training genuinely capable foundation models from scratch, while a vastly larger number of teams build on top of those foundation models rather than training their own from the ground up.
This division of labor mirrors the broader build-versus-adopt pattern covered throughout this collection’s infrastructure discussions, applied here at the level of the model’s original training rather than its downstream deployment, the immense upfront cost of foundation model training gets justified by amortizing it across an enormous number of downstream uses, and a team building an application on top of an existing foundation model benefits from that amortized investment without needing to bear any of the original training cost itself.
Why foundation models carry real risks precisely because of how broadly they’re reused
A foundation model’s broad reuse across countless downstream applications means any flaw, bias, or limitation present in that foundation model doesn’t stay contained to one narrow application, it propagates into every single downstream system built on top of it, a genuinely different risk profile than a narrowly trained, single-purpose model would carry. This is why evaluating a foundation model carefully before building on top of it, covered throughout this collection’s discussion of AI native testing, matters more than it might for a smaller, more narrowly scoped model, a problem discovered late, after many downstream applications have already been built on a flawed foundation, is considerably more costly and disruptive to fix than one caught early, before that widespread reuse has happened.
This same dynamic also means an improvement to a foundation model, a better training approach, a more diverse and representative dataset, a fix for a known limitation, benefits every downstream application built on top of it simultaneously, which is exactly the flip side of the same broad-reuse property that makes the risks so consequential, the same characteristic that concentrates risk also concentrates benefit whenever the foundation itself genuinely improves.
How the foundation model concept changed how teams think about building AI products
Before the foundation model approach became standard, building an AI capability for a specific product typically meant training a dedicated model from scratch for that exact purpose, a considerable undertaking requiring its own data collection, its own training infrastructure, and its own specialized expertise for every single application. The foundation model approach inverted this, a team building a new AI product today typically starts from an already-trained foundation model and adapts it, through fine-tuning or prompting, rather than starting from nothing, which has dramatically lowered the barrier to building genuinely capable AI applications.
This shift explains much of the rapid proliferation of AI-powered products and features across the industry in recent years, the hard, expensive part, training a broadly capable model from scratch, has already been done by a relatively small number of organizations, and the considerably easier, cheaper part, adapting that existing capability to a specific application, is what most teams building AI products are actually doing today, which is a meaningfully different and more accessible starting point than training a model from scratch would have required.
Common mistakes people make around foundation models
1. Using “foundation model” and “large language model” interchangeably, missing that foundation models cover a broader category including image, audio, and multimodal models beyond text alone.
2. Treating a foundation model’s broad capability as automatically sufficient for a specific task without considering whether fine-tuning or careful prompting is needed to actually achieve reliable, specialized behavior.
3. Underestimating how a flaw or limitation in a widely reused foundation model propagates into every downstream application built on top of it, rather than staying contained to one narrow use case.
4. Choosing to train a new model from scratch for a task that an existing foundation model, properly fine-tuned or prompted, could have handled with considerably less effort and cost.
5. Assuming every foundation model is equally suited to every downstream task, missing that a model’s specific pretraining data and approach shape which kinds of adaptation it handles best.
What connects these mistakes is treating foundation models as a single, undifferentiated category rather than as a genuinely important architectural concept with real implications for risk, cost, and capability, understanding what makes a model a foundation model, and what that origin actually implies for how it should be adapted and evaluated, is what separates building thoughtfully on top of one from simply assuming its broad capability translates automatically into good results for any given task.
The deeper point about foundation models is that they represent a fundamental shift in how AI capability gets built and distributed, from bespoke, single-purpose models trained separately for each application toward a smaller number of broadly capable models that get adapted many times over, and understanding this shift is essential to understanding why AI development has accelerated as dramatically as it has, and why the quality of a small number of foundation models now matters disproportionately to the quality of AI products across the entire industry built on top of them.