Artificial Intelligence
AI represents a job category that covers a wide range of roles and skills. At one end, researchers develop new model architectures. At the other, engineers integrate AI capabilities into products. Most roles in the job market are closer to the engineering end. The standard approach today is to start with a large pre-trained model and adapt it to your use case through prompting, fine-tuning, or building systems around it.
The core mechanism
A large language model (LLM) is a system trained to predict what comes next in a sequence of text. Given a partial sentence, it estimates the probability of each possible next word (technically, each possible next token), and generates text by sampling from those probabilities. It has processed an enormous amount of text and learned statistical patterns about how language works, which makes its output surprisingly useful across many tasks.
What it cannot do
Although an LLM can do a lot, it cannot do everything. It has no memory between separate conversations and cannot look anything up unless given a tool to do so. It has no access to real-time information unless given access to the web. Most importantly, it has no way to distinguish between something true and something that just sounds like it could be true, which means it’s capable of offering misinformation or incorrect information. Understanding the limitations of this kind of AI is as important as understanding its capabilities.
Why prompting is a learnable skill
One of the most important elements of creating a useful LLM is how you provide it with information, and the output is highly sensitive to how you frame the input. Asking the same question two different ways can produce very different results. Prompt engineering is the practice of structuring inputs to make the outputs more reliable and useful. It’s not about trying lots of random approaches. It’s about being specific, making sure any relevant context is included, and understanding how the model makes its decisions.
What actually works
Have a clear idea of what you want the output to look like and be specific about it. Providing examples of good or desired outputs when you can is referred to as few-shot prompting, and can help with prompt engineering. When asking an LLM to complete complex reasoning tasks, you can ask it to think step by step before giving an answer, a process known as chain-of-thought prompting. It’s also useful to specify what the model should not do, and make sure that the same result becomes repeatable across varied inputs rather than assuming one good result means the prompt is robust.
What an embedding is
Since computers can’t understand the pure meaning of human language or visual concepts, embeddings serve as the bridge between what a person might ‘mean’ and what a computer can understand. It’s the technique that’s used in machine learning to translate real-world data into lists of numbers (known as vectors), and to map this data into a mathematical space where items with similar meanings are positioned close together. If this sounds abstract, think of the two sentences ‘what are your opening hours?’ and ‘when do you close?’. These two sentences share none of the same words, but are semantically similar. Embeddings allow you to find semantically similar content computationally.
Why this matters for AI systems
Most commercial AI applications use some version of retrieval-augmented generation (RAG) and it’s the dominant pattern for giving a language model access to a specific knowledge base, such as a document library, a product catalogue, a codebase. The system converts all documents to embeddings, stores them in a vector database, and when a user asks a question, it retrieves the most relevant chunks and puts them in the model’s context.
The fastest way to build something useful
You don’t need to train a model yourself if you want to build an AI application. The major AI providers offer APIs that let you send a prompt and receive a response, and this is now how most commercial AI products are built. Using these APIs is the fastest way to understand what is possible and what is difficult in the sphere you’re trying to work in.
What to build first
Use the Anthropic or OpenAI API and keep your first project small enough to finish. Build a chatbot that can answer questions about a document that you provide, and this will show you how to handle API calls, manage conversation history, deal with context window limits, and think about what happens when the model gets the answer wrong. These challenges represent the core problems of almost every AI application.
What fine-tuning is
Fine-tuning takes a pre-trained model and continues training it on a smaller, task-specific dataset. The result is a model that performs better on your specific use case than the base model with prompting alone. It is most useful when you need consistent output format, specific terminology or tone, or behaviour that prompting alone cannot reliably produce. For example, creating an LLM system for summarising patients’ medical histories (data that is highly sensitive and cannot be found in public domains) would require fine-tuning.
When not to bother
Fine-tuning is expensive and slow relative to prompt engineering, so make sure you’ve exhausted all the avenues that careful prompting can achieve. Many production AI systems never fine-tune at all. It’s the right choice when you have hundreds or thousands of high-quality examples, a task where consistency matters more than generalisation, and the resources to evaluate whether it actually helped.
The problem that most builders skip
Deciding that a system ‘works’ after being tested on a few examples is a mistake when building with AI. Test cases that you have provided do not necessarily reflect the range of inputs that end users will be providing, and since LLM behaviour is probabilistic, the same prompt does not always return the same answer. This makes the evaluation process absolutely essential.
What a basic evaluation process looks like
Build a set of test cases with known correct answers. Run the system against all of them automatically. Track the pass rate over time. When you change the prompt or the model, re-run the evaluation process to check you have not regressed. This is what the AI field calls an “eval suite.” It is the difference between shipping something you really understand and shipping something you hope works.
