The AI hierarchy and the five-stage LLM mental model
Place LLMs inside deep learning, deep learning inside machine learning, machine learning inside AI — then open the LLM up into five stages.
The portal's foundation is two nested pictures. First the hierarchy: artificial intelligence is the broad family; machine learning learns patterns from data; deep learning stacks layered networks for perception; large language models are deep-learning models trained on language; generative AI is the application layer that creates rather than classifies. The portal is explicit that it teaches *using* generative models, not training networks.
Second, how an answer actually gets made — five stages: tokenization (input cut into tokens, 'like a chef cutting a vegetable'), embeddings (tokens as numbers where similar concepts sit near each other, like map locations), attention (weighing what matters — 'a torch on the correct row in a dark cinema'), prediction (estimating the next token), and variation (repeated sampling produces a full answer, differently worded each run). Variation is expected, not an error — but important claims still require verification.
The portal's suggested exercise: 'Explain LLMs like I'm five', then 'explain the same to a CFO' — watching what the audience layer changes is the fastest way to feel context at work.
This is the same ground Catalyst's Basecamp 1 and 5 covered by video, here in a quotable written form — the KB's most re-usable plain-language explanation of what a model is doing.
An LLM retrieves the memorized answer to your question.
It repeatedly predicts a likely next token in context — a 'smart parrot' with autocomplete mechanics, which is why grounding and verification matter.
Go deeper
In one line: Hierarchy: AI ⊃ machine learning ⊃ deep learning ⊃ LLMs, with generative AI as the creation-focused application layer. LLM answer production in five stages: tokenization, embeddings, attention, prediction, variation — probabilistic sampling means same prompt, differently-worded answers, which is expected behavior, not failure.
Answer from memory first — the recall attempt is what makes it stick. Then reveal.
Recite the five stages and one analogy for each.
Tokenization (chef cutting vegetables), embeddings (nearby places on a map), attention (torch on the right cinema row), prediction (next-token estimate), variation (re-sampled wording each run).
Two runs of the same prompt disagree in wording. Broken?
No — sampling is probabilistic, so variation is expected. What it does mean: verify material claims rather than trusting fluency.
