AI Model

Engineering AI Models

What is a Language Model?

A language model is an AI program trained on vast amounts of text to understand human language and predict the most likely next word in a sequence.

Real World Example

Think of it like supercharged autocomplete on your phone.

Input: “The cat sat on the…”

Language Model Prediction: It calculates probabilities based on patterns it learned from reading books and articles:

Mat (85% probability)
Rug (10% probability)
Banana (0.01% probability)

By repeating this prediction process word by word (or token by token), the model generates full sentences, answers questions, translates languages, or writes code.

What is a token?

A token is the basic unit of text that a language model reads and processes. It is not always a full word. It can be a single character, a word segment, or a whole word.

Think of tokens like Lego blocks for words.

If you input the sentence: “Unbelievable!”

A language model doesn’t process it as one big word. Instead, it breaks it into smaller chunks (tokens):

Un (prefix)
believ (root)
able (suffix)
! (punctuation)

On average, for English text:

1 token ~ 4 characters or 0.75 words
100 tokens ~ 75 words

Different Kinds of AI Models

AI models can be categorized in several ways. One useful classification is by the type of problem they solve.

Classification models

Classification models assign an input to one or more categories.

Examples:

Email -> Spam / Not Spam
Transaction -> Fraud / Legitimate
Ticket -> Billing / Technical / Account

Regression models

Regression models predict continuous numerical values.

Examples:

House features -> Estimated price
Historical traffic -> Expected traffic
Customer activity -> Expected revenue

Embedding models

Embedding models convert data into numerical vectors that capture semantic relationships.

For example:

"How do I reset my password?" -> Embedding vector

Similar concepts tend to produce vectors that are close together in vector space.

Embeddings are fundamental to systems such as:

Semantic search
Recommendation systems
Document retrieval
Retrieval-Augmented Generation (RAG)
Duplicate detection

Generative models

Generative models create new content based on learned patterns.

They can generate:

Text
Code
Images
Audio
Video

Large language models are one important category of generative models.

3.5 Multimodal models

Multimodal models can process more than one type of input or output.

For example:

Text + Image -> Reasoning -> Text

A multimodal system may accept a screenshot and a question about that screenshot, rather than requiring the image to be converted into text first.

How LLMs Work

A Large Language Model (LLM) is a language model trained on very large datasets using neural-network architectures, most commonly transformer-based architectures.

At a high level:

Tokenization

The input text is converted into tokens.

"Design a scalable API" -> [tokens...]

The model operates on these numerical representations rather than raw strings.

Embeddings

Tokens are transformed into numerical representations called embeddings.

These representations allow the neural network to work with semantic and contextual information.

Transformer architecture

Modern LLMs are primarily based on the Transformer architecture.

One of the most important components is self-attention.

Self-attention allows the model to determine which parts of the input are relevant to each other.

For example:

The service returned an error because it exceeded the timeout. <- relevant context

The model can learn relationships between service, error, exceeded, and timeout rather than treating every word independently.

Training

During pretraining, the model is exposed to a large corpus of data and learns to predict tokens.

A simplified example:

Input:
The API request returned a

Target:
timeout

The model produces a probability distribution over possible next tokens.

Training adjusts the model’s parameters to reduce prediction error over many examples.

This process requires substantial compute, data, and optimization infrastructure.

Inference

After training, the model can be used to generate responses.

Generation happens token by token.

This is why concepts such as context window, latency, throughput, temperature, and token cost matter when designing LLM-powered applications.

What an LLM Does and Doesn’t Do

An LLM is fundamentally a model for generating outputs from input context.

It does not automatically have:

Current information about your private systems
Access to your database
Access to your APIs
Permission to execute arbitrary operations
Reliable knowledge of events that occurred after its training data
A guaranteed source of truth

This distinction is critical in production AI systems.

Instead of expecting the model to contain everything, production systems often provide the model with context, tools, retrieval systems, and controlled execution environments.

That leads to agentic AI.