How AI Works in Practice
Knowing the layers of the technology helps you tell the difference between a real capability and a marketing claim.
Photograph via Unsplash
Lesson 2.1 · Training, Inference, Reinforcement
Three things happen to an AI model in its lifetime. You will hear all three words used loosely. They mean specific things.
Training is the slow, expensive part. The company building the model feeds it enormous quantities of text, images, audio or video, and lets the model adjust its internal weights until it is good at predicting what comes next. Training a frontier model in 2026 costs hundreds of millions of US dollars and uses entire data centres for months. Once a model is trained, the weights are frozen.
Inference is the fast, cheap part. Inference is what happens when you type a prompt and the model gives you back a response. The model passes your input through its frozen weights and produces a probable output. Inference is cheap enough that an answer to your prompt costs the company a fraction of a cent.
Reinforcement is the ongoing fine-tuning. Companies collect feedback from how users react to different responses (the thumbs up, the thumbs down, the regenerate button) and use that feedback to nudge the model towards behaviour humans prefer. This is why ChatGPT, Claude and Gemini have become more conversationally polished since 2023 even when their core training is older.
The implication for your work: when a tool launches a new version, that is usually a fresh training run, which means a fresh knowledge cutoff and sometimes new capabilities. When a tool gets “smarter” without a version bump, that is usually reinforcement.
Lesson 2.2 · The Stack, From Hardware to Apps
The AI world is built in layers. Most of what you encounter sits near the top.
Layer 01: Hardware. Nvidia chips, the undersea cables that move the data, and the cooling systems. Almost nobody outside Nvidia and a few hyperscalers operates here.
Layer 02: Training infrastructure. The data centres run by Microsoft, Google, Amazon, and a handful of others. This is where the models are trained.
Layer 03: Foundation models. GPT, Claude, Gemini, Llama and Mistral are the trained weights. Most are owned by a small number of US-based companies.
Layer 04: Apps. This is where you live: ChatGPT, Midjourney, Suno, Runway, Canva Magic, Adobe Firefly. The apps are usually thin wrappers around foundation models, with a user interface and some workflow logic added.
Money flows downward. The Jamaican creative pays Midjourney US$10 a month. Midjourney pays its foundation model provider for inference. The provider pays the data centre. The data centre pays Nvidia.
Capability flows upward. New foundation model capabilities released at Layer 03 reach you at Layer 04 with a delay of weeks to months, depending on how fast the apps integrate them.
The implication: when you read that “AI” can now do X, ask which layer the announcement is from. A foundation model that can theoretically do something is a long way from your favourite Layer 04 app letting you do it tomorrow.
Lesson 2.3 · Hallucination and Confidence
The single most dangerous quality of current AI models is that they sound the same when they are right and when they are wrong.
Hallucination is the technical term for when a model generates content that is plausible but factually wrong: the chat tool that invents a Caribbean copyright case that never happened, the image generator that adds a sixth finger to a hand, or the code assistant that calls a function that does not exist.
Hallucination is a structural feature of how prediction engines work rather than a bug. The model is trained to produce probable continuations. When the question is in well-trodden territory, probable continuations match reality. When the question is in thin territory (specific Caribbean legal cases, niche software, recent local events), probable continuations diverge from reality, but the model has no internal alarm bell.
The model's confidence is also learned. There is no correlation between the model's tone and the reliability of its output.
The implication: treat any specific factual claim from an AI model as a hypothesis to verify rather than a fact to use. Names, dates, statistics, quotations, citations, and case law all need independent confirmation. The deeper into Caribbean-specific or business-specific territory you go, the higher the hallucination rate.
A working rule: if your client could be embarrassed by a hallucination, verify before sending.
Lesson 2.4 · Why This Stack Discipline Matters
Most of the trouble Caribbean creatives get into with AI in 2026 comes from layer confusion.
A photographer in Liguanea hears that “AI can now generate four-second video clips with consistent characters.” She tries on her usual app and it does not work. She concludes the technology is overhyped. In fact, the new capability sits at Layer 03. Her app has not integrated it yet.
A producer in Kingston hears that ChatGPT is now “agentic” and can take real-world actions. He gives it access to his email and lets it draft replies. It hallucinates a meeting that never happened, and he embarrasses himself with a client. The agentic capability exists at Layer 03, but the version available in his app at Layer 04 still hallucinates routinely.
A bank in New Kingston signs a contract with a vendor selling “AI-powered customer service.” The vendor is reselling Layer 04 access to a foundation model with no domain customisation. Six months later the bank is paying twice the going rate for a wrapped chatbot.
The discipline: when you read AI news, ask three questions. What layer is this announcement from? Does the layer above it support the new capability? When does the layer below it (your daily app) integrate it?