Building and Using Large Language Models

Suppose your organization wants a language model to help staff answer questions. You could rely on a broad, general-purpose service that handles almost any topic, or you could build a model shaped around your own field and your own data. The second path is more work, so the reason to take it has to be concrete. Building a custom model offers three main advantages: stronger results on a narrow job, better control over sensitive information, and the ability to run the model close to where it is used. Each advantage answers a different practical concern.

To see why deployment matters, follow a single prompt through two setups. In the server-hosted setup, the user's request travels over a network to a remote server, the model runs there, and the response travels back over the network. The user reaches the model only through that connection.

A smaller custom model can instead be deployed directly on a customer device such as a laptop or smartphone. The source notes that companies such as Apple are currently exploring this kind of deployment. In the local setup, the model runs where the user is, so the prompt does not need to leave the device to reach it. The diagram below traces both paths and marks the network hops that disappear when the model runs locally.

A general-purpose model is designed to be useful across a wide range of requests, from writing poetry to summarizing a legal memo. A custom-built model is tailored for a particular task or domain. The central claim of this part of the material is that, within that task or domain, a tailored model can outperform a general-purpose one. Being narrower is not a weakness. Focusing the model on one area is what allows it to do well there.

The reason lies in what the model is built to do. In the source's examples, the purpose is defined before the model is built, and that definition shapes what the model is trained to do. BloombergGPT is specialized for finance. LLMs tailored for medical question answering are built to handle questions in the medical domain. A model with a clearly defined target is aimed at one kind of work, while a general-purpose model is aimed at many kinds at once.

The second advantage concerns data privacy. When an organization sends a request to a model run by an outside provider, the request and its contents leave the organization's control. Many organizations hold information they do not want to disclose, such as customer records, internal strategy documents, or patient details. The source points to third-party providers such as OpenAI as an example, where confidentiality may lead companies to prefer not to share their sensitive data.

A custom-built model changes this picture because it can be developed and run under the organization's own arrangements. Sensitive material can stay inside the company rather than being passed to a third party for every request. This is why data privacy is listed as a distinct advantage, separate from performance. A model can perform well and still be unsuitable if the organization cannot keep its data under its own control.

The third advantage follows from a practical feature of smaller custom models. Because they are small enough to be developed for deployment on customer devices, they can run directly on laptops and smartphones. Local deployment means the model runs where the user is, rather than only on a remote server reached over a network.

The source ties local implementation to two benefits. First, it can significantly decrease latency, which is the delay between a request and its response. Second, it reduces server-related costs, because the organization does not need to supply as much server capacity to handle the same work. The source states these as effects of local implementation without promising particular numbers. The advantage is therefore a direction of improvement whose size depends on the situation.

The last advantage concerns who decides how the model evolves. The source states that custom LLMs give developers complete autonomy, which allows them to control updates and modifications to the model as needed. With a model the team builds and maintains, the team chooses when to change it, what to change, and how to respond when the domain or the data shifts.

This is a different relationship from depending on a service whose behavior is set by its provider. In that arrangement, the application's behavior follows the provider's decisions. The timeline below compares where the decision sits in each case.