A black prism refracting a red laser beam into a precise output: transformation from raw input to structured, directed result.
A language model does exactly this. Raw text enters, the model transforms it through learned patterns, and a structured, directed result emerges. The precision comes from training. The direction comes from you.

Level 4 of 4

This is where everything connects

Level 4 is the destination. Infrastructure from Level 3 hosts your AI. Git from Level 2 manages your code. Terminals from Level 1 deploy it. Hardware from Level 0 runs it. This level does not introduce a new domain: it shows you how all four previous domains combine into production AI systems, and how to use AI tools to build software whether or not you know how to code.

Two articles. Each one is dense. The material here changes how you think about AI, not as a magic service or an existential threat, but as a specific technical system with known properties, known limits, and known patterns for building on top of it.

What you know after Level 4

ML fundamentals
Training Inference Parameters Tokens What machine learning actually does: optimise a function on data until predictions become accurate
Language models
Transformers Attention Context window Temperature How LLMs predict the next token, what the context window limits, and why temperature affects output randomness
Prompt engineering
System prompts Few-shot examples Chain of thought How to direct model behaviour through input structure rather than model retraining
Production AI stack
RAG pipelines Vector databases Evals Observability The infrastructure that turns an LLM API call into a reliable, auditable, production feature
Vibe coding
Cursor / Claude Code Directing AI tools Prompt patterns Building software by describing intent; the Levels 0-3 vocabulary makes AI direction precise and effective

Learning path

Start here ML fundamentals Training, inference, parameters, and loss functions. What the model is and how it got its capabilities.
→
Layer 2 Language models Transformers, attention, context windows. Why LLMs can write code but cannot count letters reliably.
→
Layer 3 Prompt engineering System prompts, few-shot examples, and chain of thought. How to direct model behaviour without touching weights.
→
Layer 4 Production AI systems RAG, evals, observability, and the infrastructure that makes an AI feature reliable enough to ship.
→
Build Your first AI product Vibe coding tools, architectural direction, and the vocabulary to brief an AI coding assistant precisely.

Before and after

Before Level 4After Level 4
AIMagic / hype / opaque APIA statistical model trained on text; inference is prediction; hallucination is confident wrong prediction
LLM limitations“AI sometimes makes things up”Context window limits what the model can see; temperature controls randomness; retrieval fixes knowledge gaps
PromptsCopy from Reddit, hope for the bestSystem prompts set behaviour; few-shot examples calibrate format; explicit constraints reduce hallucination
RAGA three-letter acronym in architecture docsRetrieval-Augmented Generation: retrieve relevant documents at query time and inject them into the model’s context
AI product costA monthly bill that grows mysteriouslyToken cost per request times volume; context length and model size are the two main levers
Vibe codingAI writes code for you, sometimes worksDirected iteration: you own the architecture; AI implements; Levels 0-3 vocabulary makes direction precise

Articles in this level

The foundation

What is AI?

The broad category: machine learning, neural networks, and why AI 2026 is different from AI 1990. Covers supervised learning, the training process, what parameters are, and an honest account of what language models cannot do.

You should read this if: You need to evaluate an AI vendor’s claims or explain to a stakeholder why an AI system made a specific decision.

What is Generative AI?

The specific type of AI that creates text, images, audio, and code. How it differs from traditional AI that classifies or predicts. Covers the four main types (LLMs, image generators, voice AI, video AI) and why the business case is strongest for high-volume creative and analytical tasks.

You should read this if: You are using ChatGPT, Claude, or Midjourney and want to understand what category of technology these belong to.

What is Machine Learning?

How AI learns from examples instead of following hand-written rules. Covers supervised, unsupervised, and reinforcement learning with concrete examples of each. Includes the standard ML workflow and when not to use ML at all.

You should read this if: Someone on your team said “we need ML” and you want to understand what that means before approving the budget.

What is a Neural Network?

The mathematical architecture behind every modern AI model. Explains layers, weights, and backpropagation without calculus. Covers the difference between CNNs, transformers, and diffusion models.

You should read this if: You hear “neural network” and “deep learning” used interchangeably and want to understand the actual relationship.

Language and text AI

What is a Large Language Model (LLM)?

The technology behind ChatGPT, Claude, and Gemini. How LLMs generate text token by token, what a context window limits, and why they are not databases. Includes a comparison of GPT-4o, Claude, Gemini, and Mistral.

You should read this if: You access AI through a chatbot but have never understood what is actually happening when you send a message.

What is ChatGPT?

OpenAI’s product explained: what GPT-4o is, what the model family looks like, what ChatGPT can and cannot do, and how it compares to Claude and Gemini. Includes a code example showing how to use the API.

You should read this if: You use ChatGPT regularly and want to understand the product and model well enough to evaluate alternatives.

What is Natural Language Processing (NLP)?

The broader field that all text AI belongs to. Covers the core tasks (classification, summarisation, translation, extraction), how the field evolved from rule-based to statistical to transformer-based systems, and which tools to use.

You should read this if: You are building or evaluating a product that processes text and need vocabulary to describe what it should do.

Agents and behaviour

What is an AI Agent?

When LLMs stop answering questions and start taking actions. The agentic loop, tool use (web search, code execution, APIs, file systems), and multi-agent coordination. Includes a comparison of agent frameworks (LangGraph, CrewAI, AWS Bedrock Agents).

You should read this if: Your team is proposing an “AI agent” and you need to evaluate whether the task is bounded enough for current AI reliability.

What is AI Hallucination?

Why language models produce confident, fluent, factually wrong output. Where hallucination is most common, how to detect it, and three concrete techniques that reduce it: source grounding, citation prompting, and RAG.

You should read this if: You have seen an AI tool produce wrong information and need to understand whether and how to trust AI output in your workflows.

Model customisation

What is Fine-tuning?

Adapting a pre-trained model on your own data. The difference between instruction fine-tuning and domain adaptation. When fine-tuning beats prompt engineering, what it costs, and how LoRA makes it affordable. Includes code examples for the OpenAI API and Hugging Face.

You should read this if: Your team is asking whether to fine-tune a model and you need to evaluate the cost and data requirements before approving the work.

Visual AI

What is Computer Vision?

AI that understands images and video: classification, object detection, OCR, and video analysis. Covers how CNNs work, when to use cloud vision APIs vs multimodal LLMs, and how to evaluate accuracy on your own image types.

You should read this if: Your product involves processing images, documents, or video and you need to understand the AI options available.

What is Vibe Coding?

Building software by describing intent and letting AI write the code. Covers the current landscape of vibe coding tools (Cursor, Claude Code, v0, Bolt), the mental model shift from writing code to directing a system, and the patterns that separate effective vibe coders from frustrated ones. Explains why Levels 0-3 vocabulary makes AI direction precise and effective.

You should read this if: You want to build a prototype, an internal tool, or a product MVP without hiring a full development team, and you want to understand how to do it reliably rather than getting stuck on the first error message.

Why this matters in practice

AI is infrastructure, not magic: A language model is an HTTP API. It takes a JSON body with a prompt and returns a JSON body with a completion. It runs on GPU servers in a cloud region. It has rate limits, token costs, and latency characteristics. Treating it as infrastructure, rather than magic, is what separates products that scale from prototypes that do not.

Hallucination is a design problem: LLMs predict plausible next tokens based on training data. When the training data does not cover a topic accurately, the model generates plausible-sounding text that is factually wrong. This is not a bug to report to the vendor. It is a known property to design around: retrieval, grounding, and evals are the design solutions. Level 4 covers all three.

The vocabulary gap costs money: When a product team cannot evaluate AI vendor claims, they sign contracts for capabilities the model does not reliably have. When a product team cannot write a precise system prompt, they get inconsistent outputs they cannot debug. The concepts in Level 4 close that gap.

Vibe coding scales further than you expect: The constraint on vibe coding is not the AI. It is the quality of architectural direction. A person who understands what a database schema is, what an API endpoint does, and how a server handles requests can direct an AI coding tool to build a production-quality feature. That same person without Levels 0-3 will build a fragile prototype and not know why it fails.

The workshop connection

The three-workshop method used at ai-workshops.online maps directly to the concepts in this level.

Workshop 1 (Discovery) is use case prioritisation: which AI capabilities map to which business problems, and which are technically feasible with current models. That evaluation requires understanding what language models can and cannot do reliably, covered in “What is AI?”

Workshop 2 (Concept) is architecture: where does the model fit in the system, what data sources feed it, how does retrieval work, and where do evals run. Reading that architecture requires Level 3 vocabulary. Designing it requires Level 4 knowledge of production AI patterns.

Workshop 3 (Prototype) is building: a working demo using the vibe coding approach, with real infrastructure behind it. The prototype is built using the techniques in “What is Vibe Coding?” and deployed on the infrastructure from Level 3.

The wiki is the theory. The workshops are the practice. Level 4 is where both converge.

What comes next

You have completed the foundations curriculum. The next step is depth.

Production AI
Building RAG Systems LLMOps Step-by-step guides to production AI implementation
Infrastructure
From Zero to Production Capacity Planning for AI How to deploy and scale what you build
Glossary
Transformer Architecture Attention Mechanism Foundation Models Precise definitions for the technical vocabulary of AI systems

Deep dives:

  • Building RAG Systems : document ingestion, chunking, embeddings, and retrieval in production
  • LLMOps : the operational discipline of running language models reliably at scale
  • From Zero to Production : the full path from prototype to deployed product

Further reading