The Four Layers of AI

All artificial intelligence rests, ultimately, on electricity.

Training a large model requires thousands of processors operating at full capacity, continuously, for weeks or months. Once trained, the model must be served to its users, and this second demand differs from the first in an important respect: it does not conclude. Every query submitted by every user draws additional power for as long as the product remains in operation.

That electricity must be generated, transmitted across the grid, and delivered to a facility with sufficient capacity to receive it. Nearly all of it is then converted into heat, which must in turn be removed by cooling systems that consume substantial energy, and frequently water, of their own.

For these reasons, energy constitutes the fundamental ceiling on the development of AI. A larger model becomes possible only when an operator can secure a grid connection, provide adequate cooling, and sustain the cost of both over the life of the facility. These are problems that resolve slowly. A novel model architecture may be designed within months; a new generating plant or transmission corridor may require years to permit and construct. Consequently, the question of where AI can expand is increasingly a question of where power is available and who is permitted to draw upon it.

The influence of energy extends to the very top of the stack. The price a consumer pays for a single answer can be traced, through several intermediate layers, to the cost of electricity at a particular point on the grid.

Above the energy layer resides the machinery that consumes it.

At its centre is the accelerator, a processor engineered to perform the immense volume of parallel computation that neural networks demand. Surrounding it is the supporting apparatus that renders it useful: high-bandwidth memory capable of supplying data at sufficient speed, servers and racks to house the processors, cooling infrastructure to regulate their temperature, and high-performance networking that allows thousands of individual chips to function as a single, coherent computer. A training run proceeds only as quickly as the slowest component in this chain permits.

In recent years, progress in AI has corresponded to the availability of accelerators at least as closely as it has to advances in research. Ideas have seldom been the binding constraint; timely access to sufficient computation frequently has been. Organisations that secured hardware early were able to train larger models and serve larger audiences, while those unable to do so were obliged to wait or to devise workarounds.

Hardware governs three principal outcomes:

• Feasibility of training. The scale and sophistication of a model are bounded by the quantity of computation that can be assembled and sustained over time.
• Latency. The speed with which a response is delivered depends on the capability of the serving hardware and its proximity to the user.
• Unit cost. Each response consumes a measurable portion of expensive processing time, and that cost determines whether a product can be offered affordably, or indeed at all.

This explains why an exceptional application deployed on inadequate or costly hardware often remains a demonstration. It functions, but too slowly or too expensively to serve people in daily use. The same application, deployed on appropriate infrastructure, can become something on which people genuinely depend.

A foundation model is a general-purpose system, trained on an extensive and heterogeneous body of data and subsequently adapted to a wide variety of tasks. The term reflects its role: it is the base upon which other systems are constructed.

This is the layer that caused the present era to feel abrupt. For most of the history of computing, software could perform only those functions its developers had explicitly specified. The foundation model departs from that tradition. A single model can compose correspondence, translate between languages, summarise lengthy documents, explain complex concepts, write software, and reason through multi-step problems, none of which it was individually designed to do. This generality enables small teams to build products upon existing capability rather than developing their own systems from first principles.

Because it is the most visible and most widely discussed layer, the foundation model is often mistaken for the entirety of AI. It is not. The model is subject to a fundamental limitation: it provides capability without context.

A model has no inherent knowledge of the person addressing it. It cannot independently assess whether an erroneous response would amount to a minor inconvenience or a consequential harm. It has no understanding of the purpose of the product within which it operates. It responds to what it is given, in whatever direction it is guided.

The model may be usefully compared to a powerful engine without a destination. The engine determines how fast and how far a journey can proceed; it does not determine where that journey leads, nor whether it ought to be undertaken at all. That judgment must be supplied by the layer above.

The uppermost layer is the point at which a person actually encounters the technology.

An individual opens a product in order to write, to learn, to diagnose, to design, to plan, or to decide. They are not contemplating models, processors, or electrical grids; they are occupied with their own problem. The application is the mechanism through which general capability is shaped into something suited to that problem and to that person.

It is at this layer that the most consequential decisions are made:

• Selection of the problem. A model is capable of many things; an application chooses one and commits to it.
• Design of the interaction. Whether the experience takes the form of a conversation, a structured form, a voice interface, or a background process profoundly influences what people ask and the degree to which they trust what they receive.
• Definition of boundaries. The application determines what the model is asked to do, what it must decline, what information it may access, and which decisions remain the prerogative of the person.

Two applications built upon the same model may therefore behave in entirely different ways, because purpose, context, and responsibility reside at the application layer. The model furnishes ability; the application furnishes judgment regarding how that ability should be exercised. That judgment is exercised through two principal disciplines: fine-tuning and alignment.

Fine-tuning: from general capability to specialised competence

A foundation model is, by design, a generalist. Its breadth is its principal strength, yet breadth alone rarely suffices for a specific purpose. A model trained on the full diversity of human writing will, left to itself, adopt an averaged voice, reason at a generic level of expertise, and respond without regard to the conventions of any particular domain.

Fine-tuning addresses this limitation. It is the process of continuing a model's training on a narrower, carefully curated body of examples that represent the behaviour the application requires. Where pre-training teaches a model the broad structure of language and knowledge, fine-tuning teaches it the particular manner in which that knowledge should be applied: the vocabulary of a discipline, the format of a response, the tone appropriate to a given audience, the depth of reasoning a task demands.

Fine-tuning is one instrument among several. An application may also shape model behaviour through carefully constructed instructions, known as prompting, or by supplying relevant information at the moment of a request, known as retrieval. These approaches are lighter and more easily revised. Fine-tuning is typically reserved for circumstances in which the desired behaviour must be consistent, deeply ingrained, or difficult to specify through instruction alone.

The quality of fine-tuning is governed almost entirely by the quality of its data. A model learns precisely what its examples demonstrate, including their errors, biases, and inconsistencies. The construction of a fine-tuning dataset is therefore not a clerical task but an act of design: each example is an implicit statement about how the product ought to behave.

Alignment: from competence to conduct

If fine-tuning concerns what a model is able to do well, alignment concerns whether it does what it ought to do. Alignment is the discipline of ensuring that a model's behaviour conforms to the intentions, values, and constraints of those it serves.

Alignment operates at two levels. The organisations that develop foundation models undertake broad alignment work, using techniques such as instruction tuning and learning from human preferences, so that their models are generally helpful, honest, and disinclined to cause harm. This general alignment is necessary, but it cannot anticipate the particular obligations of every product built upon it.

The application layer therefore bears responsibility for a second, more specific form of alignment. A model deployed in a medical context must understand the limits of what it may responsibly assert. A model deployed in education must support learning rather than substitute for it. A model deployed in a setting of personal reflection must respect the autonomy of the person it addresses. None of these obligations is universal; each arises from the purpose of the application and must be deliberately encoded, through fine-tuning, through instruction, through the design of the interface, and through continuous evaluation of how the system behaves in practice.

Alignment is not a single intervention but an ongoing practice. Models encounter situations their designers did not foresee, and users employ products in ways that were not anticipated. A well-aligned application is one whose behaviour is regularly examined against its stated purpose and corrected where it diverges.

An illustration

Mirror Connect is a AI powered journal built by TheAware.AI. The model powering it can read what a person writes and help them recognise patterns within it: a recurring concern, a gradual change in mood, a theme that resurfaces over time. This is a genuine capability, and adapting a model to perform it well is a matter of specialised competence.

The more important question, however, is one of alignment. In a product of this kind, alignment means that the system does not diagnose, does not moralise, and does not substitute its conclusions for the person's own. The objective is not the identification of patterns for their own sake, but a person arriving at a clearer understanding of their own life, in their own words and on their own terms. That principle informs every other decision: what the product brings to the surface, what it withholds, and what it deliberately leaves for the person to conclude.

It would be convenient for a company working at the application layer to regard the layers beneath it as the concern of others. Such a view would be mistaken.

Energy determines the upper limit of available computation. Hardware converts that energy into speed and establishes the cost of each response. The foundation model transforms computation into general capability. The application, through fine-tuning and alignment, transforms that capability into something a person can use, and determines how it ought to be used.

The layers are interdependent, and a change at any one of them propagates upward. When energy becomes scarcer, hardware becomes more expensive to operate. When hardware becomes more efficient, models grow larger and applications grow faster. When models become more capable, applications can address problems that were beyond reach only a year earlier.

We build at the top of this stack, and we maintain a clear understanding of the three layers beneath it. Intention is what finds expression in a product. The layers underneath determine how much of that intention the product is able to realise.