Artificial intelligence
The vocabulary, the credentials, and where the bill arrives
In AI, the vocabulary moves faster than the substance. That gap is where most of the wasted money goes: courses named after a product that will be renamed next year, certificates that prove you can define a term, and a great deal of confident writing that never says what anything costs. The route through it: 13 concepts in the order they make sense, 15 vendor exams with what each one signals, and the two course categories that teach them.
There are two sound ways into this field and one weak one. The builder path is Python, then the maths you can tolerate, then training models on real data. It takes months and it is the only route into research or model engineering. The applied path is learning to use models well: retrieval, evaluation, and the limits of the tools: inside a job you already understand. It takes weeks, and it is where most of the value being captured right now sits. The weak path is collecting fundamentals badges from three vendors and never shipping anything; employers have learned to ask what you built.
The argument in one piece is on the artificial intelligence topic page.
The vocabulary, in three layers
Not alphabetical, and not by popularity. In the order where each term is explained by the ones above it.
Foundation: how a machine learns anything at all
Three terms that have not changed in a decade and explain every headline that has. Read these first; everything else is a special case of them.
- Machine LearningMachine learning is the branch of computing in which a program works out its rule from examples instead of being handed that rule by a programmer, and is judged on how well the rule holds on cases it was never shown.
- Neural NetworksA neural network is a stack of layers of very simple arithmetic units whose connection strengths are adjusted by training until the whole arrangement turns the inputs you have into the outputs you want.
- Deep LearningDeep learning is machine learning built on neural networks of many stacked layers, which work out for themselves which properties of the raw data matter instead of being handed a list by a person.
Architecture: the designs that won
One architecture underlies every chat assistant you have used, and two long-standing problem families explain what it was built to do.
- TransformersA transformer is a neural-network design that takes a whole sequence in at once and, for every element in it, works out how strongly each of the other elements should influence that one.
- Large Language ModelsLarge language models are neural networks trained on a vast quantity of written text to predict the piece of text that comes next, an objective narrow enough to state in a line and broad enough to yield writing, translation, code and summary as side effects.
- Natural Language ProcessingNatural language processing is the area of AI that makes human language readable and writable by machine: sorting text into categories, pulling structured facts out of it, translating it, generating it and measuring whether any of that came out right.
- Computer VisionComputer vision is the branch of artificial intelligence that turns images and video into structured answers: what is in the frame, where it sits, which pixels belong to it, and where it moves next.
Application: where most people meet AI
This is the layer in which most readers earn their living. None of it requires training a model; all of it requires knowing what the model is doing.
- Generative AIGenerative AI is the market's name for models whose output is a piece of content, such as text, images, audio, video or code, as opposed to the much larger body of models that sort, score or forecast things that already exist.
- Prompt EngineeringPrompt engineering is the practice of composing what you send a generative model, the instruction, the supporting material and the worked examples, so that it returns something you can use.
- Retrieval-Augmented GenerationRetrieval-augmented generation is a pattern that searches a body of documents when a question arrives and places the passages it finds into the model's prompt, so the answer is drawn from those sources rather than from training alone.
- Agentic AIAgentic AI is software that wraps a language model in a loop where it can request actions from real tools, read what each action returns, and choose the next move until the job is finished or a rule stops it.
The two that decide whether any of it ships
Every AI project that dies, dies here rather than in the modelling. Learn them early enough to argue with a plan, not late enough to explain a failure.
- MLOpsMLOps is the engineering practice of keeping a trained model useful in production: recording what produced it, serving it, watching for the day its answers stop matching the world, and retraining or reverting before anyone downstream is harmed by the gap.
- Responsible AIResponsible AI is the practice of building and running systems whose decisions you could defend to the person on the receiving end of them, covering fairness, transparency, privacy and accountability.
The credentials, and what each one signals
Every price below is the vendor's list price on the date shown, before tax and country adjustments. The foundational exams certify vocabulary. The associate exams certify plumbing. Neither certifies that you have built anything, and no employer reads them as if they did.
- Course completion · FoundationalAI Fundamentals: Foundations for Understanding AI0 USD (as of 26 September 2026)
- AIF-C01 · FoundationalAWS Certified AI Practitioner100 USD (as of 12 September 2026)
- Course completion · FoundationalData Fundamentals0 USD (as of 26 September 2026)
- Course completion · FoundationalGoogle Data Analytics Professional Certificate49 USD (as of 26 September 2026)
- Lab-based assessment · FoundationalMicrosoft Applied Skills: Get started developing agents in Microsoft Foundry0 USD (as of 26 September 2026)
- AI-901 · FoundationalMicrosoft Certified: Azure AI Fundamentals99 USD (as of 8 September 2026)
- DP-900 · FoundationalMicrosoft Certified: Azure Data Fundamentals99 USD (as of 8 September 2026)
- 1Z0-1122-26 · FoundationalOracle Cloud Infrastructure AI Foundations Associate0 USD (as of 26 September 2026)
- DEA-C01 · AssociateAWS Certified Data Engineer – Associate150 USD (as of 12 September 2026)
- Graded projects · AssociateData Analysis with Python Certification0 USD (as of 26 September 2026)
- Data Engineer Associate · AssociateDatabricks Certified Data Engineer Associate200 USD (as of 12 September 2026)
- Course completion · AssociateGoogle Advanced Data Analytics Professional Certificate49 USD (as of 26 September 2026)
- Course completion · AssociateGoogle Business Intelligence Professional Certificate49 USD (as of 26 September 2026)
- DP-700 · AssociateMicrosoft Certified: Fabric Data Engineer Associate165 USD (as of 8 September 2026)
- PL-300 · AssociateMicrosoft Certified: Power BI Data Analyst Associate165 USD (as of 8 September 2026)
Pick the vendor your employer already runs, or the one your target job postings name. The fundamentals exams cover substantially the same ideas in different dialects: the ideas transfer, the dialects do not. The full exam catalogue, including the security, networking and project-management ladders, is on the certifications overview.
Where AI meets the rest of the site
AI is less a field of its own than a pressure applied to the other four. These are the places where that pressure turns into a decision someone has to make.
Hosting & cloud
This is where an AI project stops being free. A model that runs on someone else's API is a per-token bill that scales with success; a model you host yourself is a GPU you rent by the hour whether it is busy or not. Neither is the cheap option by default, and the decision is the same one the hosting ladder describes: the lowest rung that meets the real need.
Data & analytics
Every serious practitioner in this field spends most of their time on data, not on models. That is why the associate-level exams above are data-engineering exams rather than AI exams, and why a retrieval system fails on the documents long before it fails on the model.
Programming
Assistants changed how code gets written and not what makes it correct. The skills that decide whether an AI-assisted codebase survives its second year, reading code you did not write, reviewing, testing, knowing what to refuse, are the same ones that decided it before.
Digital assets
Generated images are the live licensing question of the moment: what you may sell, what a marketplace now requires you to declare, and what a client's legal team will ask for. The licence vocabulary that answers it is the same one that governs stock and templates.
Courses from the AI categories
From the generative-AI and data-science course categories; each card leads to the course page with the provider link and the affiliate disclosure.
Questions people arrive with
- Do I need to learn maths to work with AI?
- Only for one of the two paths. If you want to build or train models, yes: linear algebra and statistics at working level, and there is no shortcut past it. If you want to use models well, which is where most of the value companies are currently capturing sits, no. The applied path is retrieval, evaluation, prompting and knowing the failure modes, and it takes weeks rather than months. Decide which one you are on before you buy a course, because the two share almost no material.
- Is an AI certificate worth anything?
- A fundamentals badge certifies vocabulary, and employers read it that way: it says you can hold a conversation about the platform, not that you have shipped anything. That is useful for someone moving into an AI-adjacent role from marketing, support or analysis, and close to useless on its own in an application for an engineering job. The associate-level data exams are a different proposition: they certify plumbing that a team needs. The common failure is collecting three fundamentals badges from three vendors and no project that ever ran.
- Which vendor should I certify with?
- Whichever cloud your employer already runs, or the one named most often in the job postings being applied to. The fundamentals exams from Microsoft and AWS cover substantially the same ideas in two dialects; the ideas transfer and the dialects do not. Picking on merit rather than on your target employer's stack is the most common wasted exam fee in this field.
Last reviewed 3 October 2026 · Getting Digital