A visual, infographic-style article that untangles the buzzwords once and for all. Follow the full nesting — AI, machine learning, deep learning, transformers, LLMs and generative AI — with one concrete Microsoft example at every layer, so you can place any tool or claim you encounter on the right shelf instantly.
These five terms are nested subsets, not competing technologies. Artificial intelligence is the umbrella: any system behaving intelligently. Machine learning sits inside it: systems that learn from data. Deep learning sits inside ML: learning with multi-layer neural networks. Inside deep learning, the transformer architecture (introduced in the 2017 paper 'Attention Is All You Need') is the design that made training at scale practical. Inside transformers, an LLM is a transformer trained on massive text. And generative AI is the application layer that uses LLMs (and image/audio models) to create new content. Each inner circle inherits everything from the outer ones.
Job interviews, vendor decks and AI strategy meetings all mix these words carelessly, and the person who can draw the hierarchy on a whiteboard instantly stands out. More practically, the hierarchy tells you what to expect from a tool. A classic ML-based spam filter explains itself with feature weights; a transformer-based Copilot drafts content but needs grounding and verification. When you know which layer a product lives on, you know its failure modes, its data requirements and how to talk to it — the entire rest of this curriculum depends on this map.
Walk down the nesting. Machine learning replaced hand-written rules with patterns learned from labelled or unlabelled data. Deep learning replaced hand-crafted features with stacked neural layers that learn features themselves — that is what makes it 'deep'. Transformers replaced recurrence with self-attention, letting a model weigh every word against every other word in parallel, which unlocked training on the entire internet's text. Pre-train a transformer on trillions of tokens and you get an LLM such as GPT-4o, Claude or Gemini. Wrap an LLM in a chat interface with images and audio and you get generative AI products like ChatGPT and Copilot.
AI (any machine that acts smart)
└─ Machine Learning (learns from data)
└─ Deep Learning (neural network layers)
└─ Transformers (self-attention)
└─ LLM (GPT-4o, Claude, Gemini)
└─ Generative AI (creates new content)Place real Microsoft technologies on the tree. Outlook's spam filter: machine learning (classic, non-deep). Defender for Endpoint behaviour scoring: deep learning. GitHub Copilot code completions: an LLM (transformer) fine-tuned on code. Microsoft 365 Copilot: an LLM plus generative AI application layer, grounded with Microsoft Graph. Azure Anomaly Detector: deep learning time-series models. One diagram, every product placed — this is the whiteboard moment interviewers remember.
On paper, write the six labels and nest them: AI → ML → Deep Learning → Transformers → LLM → Generative AI. Then ask ChatGPT to "draw the same hierarchy in ASCII with one Microsoft example per layer" and compare. Correct your version where the AI exposes a gap.
Ask the AI: "Classify these into the hierarchy with one line of justification each: Excel forecast sheet, Defender for Endpoint, Windows Hello, GitHub Copilot, Outlook spam filter, Azure Anomaly Detector, a GPO, Microsoft 365 Copilot, Azure Face API, Intune remediation scripts." Notice how rule-based tools (GPOs, scripts) sit outside the tree entirely.
Ask: "Explain self-attention in transformers to a Windows admin using the sentence 'The app failed because its detection rule did not match' — show which words attend to which." You will see how the model links "failed" to "detection rule" without any hand-written grammar.
Open the public GPT-4o or Claude model card / documentation page. Find three facts: which company trains it, what the context window is, and what modalities it accepts. Reading vendor documentation critically is a Level 1 skill most admins skip.
Explain the full hierarchy out loud to a colleague (or a voice recorder) in under a minute without notes. If you stumble on a layer, revisit it — this Feynman technique is also your interview answer, ready-made.
The hierarchy also maps risk. Classic ML tools process narrow data you defined; generative AI products take free-text input and can surface training-data content, so data-handling rules differ by layer. Know which layer each tool in your tenant uses before approving it: an LLM chat feature needs prompt-logging and data-boundary review that a spam filter never did. Azure OpenAI, for example, gives you content filtering, region pinning and no-training-on-your-data guarantees — controls you should verify layer by layer.
They are nested subsets. AI is the umbrella field of intelligent behaviour in software. Machine learning is AI that learns from data rather than explicit rules. Deep learning is ML using many-layered neural networks. Transformers are a deep learning architecture built on self-attention, and an LLM is a transformer pre-trained on massive text. So an LLM is also DL, ML and AI — each term narrows the previous one.
Self-attention let models process every token in relation to every other token in parallel, instead of step-by-step like recurrent networks. That made training on enormous datasets efficient on modern GPUs, and it captures long-range context far better. Every major LLM — GPT, Claude, Gemini, Llama — is a transformer; the 2017 paper it comes from is titled 'Attention Is All You Need'.
Copilot is an application built on an LLM — so it is generative AI, on a transformer, using deep learning, which is a form of machine learning, inside the AI umbrella. Its distinguishing feature is grounding: it combines the LLM with your organisation's Microsoft Graph content under your existing permissions, which is what separates it from a generic chatbot.
Create a one-page infographic for your team: the nested hierarchy on the left, one Microsoft product example per layer on the right, plus one sentence per layer on 'what can go wrong'. Ask an LLM to critique it: 'Here is my AI hierarchy cheat sheet for IT admins — find any misplaced example or missing layer.' Iterate once and pin it in your team's channel.
No. For small, structured datasets — think a few thousand rows of ticket metadata — a decision tree or gradient-boosted model is often more accurate, faster, cheaper and easier to explain. Deep learning wins on unstructured data at scale: text, images, audio. Knowing which to reach for is exactly why the hierarchy matters.
A transformer is a neural network design whose key trick is self-attention: while processing any word, it can directly weigh the relevance of every other word in the text. This parallelism makes training on huge datasets efficient and gives models their impressive grasp of context. GPT literally stands for Generative Pre-trained Transformer.
You need the map, not the calculus. Knowing that an LLM is a transformer predicting tokens tells you why context windows, hallucinations and grounding behave the way they do — enough to use AI safely and evaluate tools. Training your own models is where deeper knowledge starts, and that is optional for most IT roles.