What is AI, ML, and LLM?

Explained in plain language, without jargon or formulas.

[🇷🇺 Russian version: what-is-ai.ru.md]

← Level Zero · How Neural Networks Work →

Contents

1. AI, ML, LLM — What’s the Difference?
2. What is a Language Model (LLM)?
3. Where Does the Model Get Its Answers?
4. Parameters: What Do 7B, 14B, 70B Mean?
5. Open-Source vs Proprietary Models
6. What’s Next

1. AI, ML, LLM — What’s the Difference?

These three acronyms are often confused. Let’s break them down.

┌──────────────────────────────────────┐
│  AI (Artificial Intelligence)        │
│  Artificial Intelligence — general   │
│  concept: a machine that "thinks"    │
│                                      │
│  ┌──────────────────────────────┐    │
│  │  ML (Machine Learning)       │    │
│  │  Machine Learning — AI that  │    │
│  │  learns from data            │    │
│  │                              │    │
│  │  ┌──────────────────────┐    │    │
│  │  │  LLM (Large Language │    │    │
│  │  │  Model)              │    │    │
│  │  │  Large Language      │    │    │
│  │  │  Model — what you    │    │    │
│  │  │  chat with           │    │    │
│  │  └──────────────────────┘    │    │
│  └──────────────────────────────┘    │
└──────────────────────────────────────┘

AI — the broadest concept. Includes chess programs, face recognition on your phone, and voice assistants.

ML — a way to create AI: instead of manually programming rules, you “feed” the program examples and let it find patterns itself.

LLM — a specific type of ML that works with text. ChatGPT, Claude, Qwen, DeepSeek — these are all LLMs.

For this handbook: we talk almost exclusively about LLMs — language models you can run on your own computer.

2. What is a Language Model (LLM)?

Imagine a “smart autocomplete” that works not with words, but with entire texts.

You start a phrase — the model continues. You ask a question — the model “completes” the answer. Everything it does is predict the next word (technically, token) based on previous ones.

Analogy: An LLM is a person who has read almost the entire internet and can now finish your thought. Not because they “understand,” but because they’ve seen similar texts millions of times.

Important: the model doesn’t “think” or “understand” in the human sense. It’s an extremely complex probability calculator: sequence of words in → most probable continuation out. But due to complexity (billions of parameters), the result looks like the model actually understands.

3. Where Does the Model Get Its Answers?

The model does not have internet access when answering. All its “knowledge” is what it memorized during training.

The process looks like this:

Stage 1: Training

The model is “fed” massive amounts of text — books, articles, websites, code. Trillions of words. It learns to predict the next word again and again until it gets good enough at it.

Analogy: imagine you’ve never seen chess. You’re shown a million games, and you start guessing what move is usually made in a given position. You don’t know the rules — you’ve just “played out” the statistics.

Training a large model takes months and costs millions of dollars (electricity, server rental). This is done by big companies: Meta, Google, Alibaba, DeepSeek.

Stage 2: Inference

This is what you do when typing a question in chat. The model is already trained — it just applies its knowledge. This is fast and free (only electricity).

Training → model became smart. Inference → it answers questions.

Stage 3: Fine-tuning

If you need the model to understand your narrow domain — you can take a ready model and “fine-tune” it on your data.

Analogy: the model graduated regular school. You send it to professional development courses in your specialty. This is much faster and cheaper than teaching from scratch.

4. Parameters: What Do 7B, 14B, 70B Mean?

When looking at a model, you see: Qwen 3.5 7B, Llama 3.1 70B. The number with B — number of parameters (B = billion).

Parameters are the model’s “neurons”. More parameters = potentially smarter model, but also more resources needed.

Parameters	Analogy	Where to Run	Examples
1–3B	🐭 mouse brain	Any laptop, 8 GB RAM	Phi-3-mini, Qwen 2.5 1.5B
7–9B	🐕 sweet spot	MacBook / PC, 16 GB RAM	Qwen 3.5 7B, Llama 3.1 8B
14–30B	🐒 needs hardware	32 GB RAM or GPU	Qwen 3.5 14B, DeepSeek-R1 14B
70B+	human	Server, 64 GB+ RAM	Llama 3 70B, DeepSeek-R1 671B

Golden rule: more parameters = smarter model, but proportionally more memory needed. On MacBook Air 16 GB — your ceiling is 7–9B. On MacBook Pro 48 GB — you can run 30B.

Important nuance: 70B model is not 10x smarter than 7B. It’s 20–30% smarter but needs 10x more resources. For most everyday tasks, 7–9B models are more than enough.

5. Open-Source vs Proprietary Models

For local running we only care about open-source models. Their weights are open — you can download the model file and run it anywhere.

Leading open-source models as of July 2026:

• Qwen 3.5 (Alibaba) — best for Russian language
• Llama 3.1 / 4 (Meta) — benchmark, good English
• DeepSeek-R1 / V3 — excellent reasoning
• Mistral — efficient models
• Gemma 4 (Google) — gaining popularity
• Phi-4 (Microsoft) — small and smart

6. What’s Next

In section: what-is-ai · how-models-work · cloud-vs-local · hardware-guide · glossary · faq · learning-path · setup-windows · setup-linux
Related sections: Local Models · AI Agents · Use Cases
Navigation: ← Zero Level · ↑ Back to main · 🇷🇺 Русский