"AI model" isn't one thing — it's a whole toolbox. A model built to write essays is a bad fit for outlining objects in a photo, and a model built to click through a website is a bad fit for answering trivia. Below are eight architectures you'll keep running into, explained the way you'd explain them to a curious beginner: what they actually do, a simple analogy to anchor it, and where you'd actually see one in the wild.
🧠 LLM — Large Language Model
What it does: reads and writes text the way a well-read person does — it predicts, one piece at a time, what word should come next, and strings enough of those predictions together to hold a conversation, explain a concept, or write code.
Think of it like: a friend who has read an enormous library and can talk fluently about almost anything you bring up, even if they've never seen your exact question before.
Where you'll see it: chatbots, writing assistants, coding help, summarizing documents. GPT, Claude, and Gemini are all LLMs — the generalists of the AI world.
🎨 LCM — Latent Consistency Model
What it does: turns a text prompt into an image in just a handful of steps, instead of the 20–50 steps a typical image-generating model needs to slowly refine noise into a picture.
Think of it like: a sketch artist who can draw a recognizable portrait in three quick strokes, instead of one who needs fifty careful passes to get the same likeness.
Where you'll see it: real-time image generation tools, live drawing apps, and anywhere speed matters more than squeezing out the last bit of detail.
🤖 LAM — Large Action Model
What it does: goes past just talking — it plans a sequence of steps and then actually carries them out, clicking buttons, filling in forms, and completing tasks across real apps and websites on your behalf.
Think of it like: a capable assistant you hand a task to — "book this flight" — who doesn't just tell you how to do it, but goes and does it, step by step, checking the outcome along the way.
Where you'll see it: AI agents that operate a browser, automate multi-step workflows, or complete a purchase or a form submission without a human clicking each step.
🔀 MoE — Mixture of Experts
What it does: this one isn't a task type like the others — it's a design trick for building a huge model that stays fast. Instead of using every single part of the network on every input, a small "router" sends each piece of input to just a few specialist sub-networks, leaving the rest switched off for that input.
Think of it like: a school with a hundred subject-specialist teachers, where the principal sends each question straight to the one or two teachers who actually know that topic, instead of making every teacher answer every question.
Where you'll see it: under the hood of some of today's largest language models — Mixtral and DeepSeek-V3 both use this trick to hold enormous knowledge without needing enormous compute for every single request.
👁️ VLM — Vision-Language Model
What it does: looks at an image and reads text at the same time, then reasons about the two together — describing a photo, answering a question about a chart, or comparing what's written against what's shown.
Think of it like: a museum guide who can look at a painting and a placard side by side and explain how the two connect, rather than only being able to read the placard or only look at the painting.
Where you'll see it: "upload a photo and ask about it" features, chart and document understanding, and most of what people casually call a "multimodal" assistant.
⚡ SLM — Small Language Model
What it does: the same basic idea as an LLM — predicting and generating text — but built deliberately small, so it runs quickly, cheaply, and often directly on a phone or laptop instead of a data center.
Think of it like: hiring a focused specialist for one specific job instead of flying in a world expert for it — the specialist is faster, cheaper, and plenty good enough for that one job.
Where you'll see it: on-device assistants, offline apps, and narrow tasks (like autocomplete or simple classification) where calling a giant model would be overkill.
🧩 MLM — Masked Language Model
What it does: trains by having some words in a sentence hidden ("masked") and learning to guess what's missing from the surrounding context — which teaches it to deeply understand how language fits together, rather than mainly to generate long new text.
Think of it like: a fill-in-the-blank worksheet — getting good at guessing the missing word forces you to genuinely understand the sentence around it, not just memorize the sentence as a whole.
Where you'll see it: search engines, text classification, and the embeddings that power semantic search — BERT is the best-known example of this style of model.
✂️ SAM — Segment Anything Model
What it does: looks at an image and traces the exact outline of objects in it, pixel by pixel — not just "there's a dog in this photo," but precisely which pixels belong to the dog and which belong to the background.
Think of it like: someone tracing every object in a photo with scissors along its exact edge, instead of just circling roughly where it is.
Where you'll see it: precise photo-editing tools ("cut out this object"), medical-image analysis, and any task that needs an exact boundary rather than a rough label.
💡 Key takeaway: Don't always choose the biggest model. Choose the architecture that matches the task.
Comments
Post a Comment