Understanding AI from zero in 2 hours: a mental framework so you never feel overwhelmed again

Every week there's a new AI tool. Today it's Claude Code, tomorrow Codex, next week some weird name. Many people learn AI by chasing tools—and the more they chase, the more overwhelmed they get, because tools are infinite but time is not.

This piece goes the other way. I'm not teaching you which buttons to press. I'm building you a mental framework. Once you have the framework, whenever any AI tool comes out later, you'll take one look and understand what it is, what it can do, whether it's worth learning.

After reading this, you'll have 4 pieces: understanding large models — writing prompts — understanding Agents — getting your hands dirty for real.

---

Part 1 — Understanding "large models" without being an engineer

AI is not new, what's new is "AI that speaks human language"

AI has existed since the 1950s. Face recognition on phones, TikTok's recommendation algorithm, AlphaGo playing go — all of that is AI. But we don't call that "an AI boom."

What's exploding is AIGC — AI-generated content: you say what you want, AI creates exactly that. The critical difference: it speaks human language. You chat in everyday language, it understands, then generates text, images, or video. For the first time, AI is stepping directly into ordinary people's lives.

My advice for beginners: start with text → text (describe with words, get words back). Lowest barrier to entry, most applications, most free tools.

Where does "large model" get its size

Behind every AI app (ChatGPT, Claude, Gemini, DeepSeek...) are large language models (LLM). Understand the word "large" through the image of raising a child from kindergarten to a PhD:

• Large data = reading an entire ocean of books. The wider the data, the more complete the model's "grasp" of the world.
• Large computing resources = enormous tuition fees. Training consumes thousands of GPUs running for months.
• Large number of parameters = brain capacity. The more parameters, the more "intelligent" the model.

We're not researchers. The goal of understanding LLM is to know the boundaries of its capabilities, not to dig into deep principles.

How AI works: like a librarian with superhuman memory, but understanding nothing

Don't imagine AI "thinking" like a human. Picture it as a librarian with a colossal memory but understanding nothing about the content of books:

1. It reads billions of text fragments and remembers the combinations of words that go together. "Rain" often goes with "umbrella."
2. When you ask, it plays a word-chaining game like phone keyboard autocomplete — but a million times more complex. It picks the next piece by highest probability.
3. It doesn't truly understand what it's saying. No lived experience, no common sense, doesn't know on its own that it's wrong.

Consequence: it can write beautiful poetry and also "lie very convincingly." Humans are the final validation layer to see whether the answer makes sense — not AI.

Strengths and weaknesses: why it's "human–machine collaboration"

Not "AI does everything," not "humans do everything," but take what AI does well to compensate for what you do poorly.

Strengths: handling large volumes of information (reading, summarizing, sorting), knowledge spanning all fields, creativity and content generation.

Weaknesses — must remember these by heart:

• Hallucination: it predicts probability, not asserts truth. May fabricate citations, make up events.
• Knowledge cutoff: information stops at the training point, doesn't auto-update.
• Amplifies existing biases in the training data.
• Multi-step reasoning is weak, easy to err on problems requiring balancing multiple conditions.
• Can't grasp deep meaning / sarcasm, and has context length limits — talk too long and it forgets the beginning.

So when using: give clear commands, provide context, cross-check important information, break complex problems into pieces, and always keep a grain of doubt.

The essence of an AI app = Core (model) + Shell (interface)

AI app = Large model (soul) + Product interface (body)

Like a phone: a phone maker is like an AI app, a chip maker is like a model maker, a chip model is like a model name & version. The chip matters a lot, but whether a phone is good depends on the OS, camera, ecosystem. Same with AI apps: the same model, two different apps can give very different experiences because the "shell" differs (web search, file reading, tool calling, preference memory...).

Watch out for software in disguise: a small group builds a shell, plugs in an API, advertises "using premium GPT" but actually runs a cheap model. Choosing an app: there's no best one, only the one that fits you — consider your real needs, acceptable costs, the actual model power behind it, and whether the team is trustworthy.

3 ways to "upgrade" AI, and when to use which

Many people ask: "Can I train my own model?" — Classic misunderstanding. There are 3 ways to make an impact, different levels:

1. Prompt (command): cheapest, everyone can do it, use daily. Just learn how to ask better.
2. Fine-tune (tuning): let AI "learn more" to become expert in one field, by adjusting a small portion of parameters with specialized data. For tech teams, hard, expensive.
3. RAG (knowledge base): give AI a "living dictionary" on the side. When you ask, it searches the document store then answers. Solves two things: outdated knowledge, and private business data. AI search and customer care chatbots are RAG.

Which to pick? Simple: they solve different problems. Prompt for everyone, daily tasks; RAG for businesses/devs needing new knowledge or private data; Fine-tune for tech teams when prompts never hit the mark. For ordinary people, fine-tune is farthest away — not needed yet.

Pick a model that fits you: look at 5 criteria

No "best" model, only the one that fits best. Five criteria to compare:

1. Whether you need to "cross borders" — many international models aren't easy to access in Vietnam.
2. Cost — foreign models usually start at 20 USD/month; if you use API, cost is a big deal.
3. Speed — the more a model "thinks deeply," the slower; simple work needs a fast model.
4. Context length — for me this is the most important criterion: it decides how deep you can go in a conversation before AI forgets. Note: long context doesn't mean good memory — "storing more" is different from "finding what matters in that pile." Tip: periodically ask AI to summarize conclusions and save to a document.
5. Field strength — different models excel differently at writing, code, reasoning, long document analysis.

And the real way to check: build your own "test case" set — a few scenarios you use regularly. Whenever a new model comes out, test it on those same problems. If it works, keep it; if not, ignore the hype. Picking AI tools should have the "spirit of buying fish": keep multiple models in hand, match each job to an optimal tool. At this stage, no single model can do it all.

Part 2 — Prompt: the bridge between you and the AI "brain"

The basic formula

Quality prompt = Clear instruction + Necessary context

Instruction is the soul: say clearly what you want the AI to do, in what way, to what standard. Context is background information that helps the AI understand correctly. For example, instead of "write me a phone introduction post", say: "Gaming phone for young people, around 3 million đồng price range, strengths are fast charging and heat dissipation; write an introduction post emphasizing game performance and fair price, youthful tone but not over the top, under 800 words."

Is "prompt engineering dead"?

No. Reasoning models (like the DeepSeek R1 line) made many people say just "be clear and that's enough". True, but: (1) not everyone can articulate clearly, (2) AI is not a person anyway, being clear with it still needs some technique. Prompts aren't dead — they're "democratized": ordinary people can use natural language, while professionals still use more sophisticated prompts. And the more powerful the model, the more we don't need to write one perfect prompt from the start — we can have back-and-forth conversations like talking with a colleague.

6 techniques for effective prompts

1. Task — Task: state the specific job clearly. "Write a popular science article about the harms of plastic waste", not "write something".
2. Context — Context: let the AI know why you're doing this, for whom.
3. Examples — Examples: give a few samples so it understands your "taste" correctly. This is my favorite technique, works great for creative work (scripts, captions).
4. Persona — Role: "Assume you're a middle school science teacher, explain in simple language".
5. Format — Format: request Markdown, tables, or JSON to reuse easily.
6. Tone — Tone: formal for reports, natural for social media.

Not required to use all 6. Every technique serves one purpose: getting the right result.

My method: "Ask high, pull low"

Newcomers often learn all the techniques but still go blank when facing the chat box. How to get unstuck:

• Step 1 — Ask high: don't ask right away "give me a gym plan". Ask the framework first: "I want to do X but don't have a starting point, can you help me list the framework I need to consider for this?"
• Step 2 — Pull low: based on the AI's framework, fit your real situation into it to make a specific prompt.
• Step 3 — Optimize continuously: if not satisfied, adjust again, keep talking until it works.
• Step 4 — Cast framework for reuse: summarize it yourself into a prompt template for next time.

Magic sentence: "I want to do XX but don't have a starting point, can you help me build a framework for this?"

Two "lazy but smart" tips

• Start by mimicking: like learning to write means reading good writing — look at prompts from skilled people, analyze their structure, understand why they did it that way, then adapt to your needs.
• Use AI to write prompts: my main approach is to "say roughly": "I want to do XXX, write me a prompt." Or tell it to mimic: "Copy the structure and tone of this prompt, write me one about YYY." Don't demand perfection right away — let the AI give you a 60-point version and iterate gradually.

Part 3 — What is Agent (AI automation agent)

Core difference

Agent is AI that can self-plan steps, call tools, and execute toward your goal. Regular chat: AI gathers and organizes data, then you dump it into Excel yourself. Agent: you state the goal — "survey market X, make an Excel table" — it finds, analyzes, creates the file and hands it to you. The final step that you used to do manually, Agent now does within its granted permissions.

Why are Agent tools mushrooming everywhere?

Like Lark / DingTalk / WeCom all being "team collaboration software" just from different companies — when a product form gains market acceptance, companies rush in to make their version. Many tools have different names but solve the same problem, work the same way, and run on similar logic underneath.

The 3-dimensional framework — the most important part of this post

When you encounter any new Agent, ask 3 questions:

1. Who made it? (OpenAI, Anthropic, big companies, open-source community...) Note: the company that builds the model and the company that builds the Agent product are not necessarily the same.
2. How do you use it? Four modes: chat app (ChatGPT, chatbot), desktop interface (easiest for regular people), command line CLI (Claude Code, Codex CLI — suited for programming but usable without knowing code), through chat app (send commands remotely via Telegram/Zalo...). These modes don't exclude each other — one product can have multiple entry points.
3. What can it do? General office Agent (write documents, analyze data, make slides), coding Agent (read code, fix bugs, run tests — nowadays also handles non-coding tasks), specialist Agent (tied to data and workflows of a specific field).

The underlying logic: Execution architecture + AI brain

Every Agent really has the same structure: execution framework (read files, run commands, call tools) + a large language model as the brain. They differ in just two ways: what tools the execution framework can call, and which model is plugged in as the brain.

Tools determine what you can do. Brain determines how well you can do it.

Which Agent is best?

Divorced from your specific needs, the question is meaningless. Just: identify your actual work first → pick a tool → don't rush to buy a plan, take a few real tasks and test it. Look at: does it complete the work or just offer suggestions? is the result accurate, how much do you have to fix? how much time/cost does it take? does it support the software/platform you need? what permissions does it need, are there data risks? Good actually = completes work steadily and economically, not just looking smart.

---

Part 4 — Getting started: roadmap for zero beginners

Don't pour effort into learning each tool's interface — today the button is here, tomorrow it moves. Learn how to work with AI in general across all tools, in 4 stages:

1. Know how to express your needs to AI — already covered in the prompt section.
2. Use one Agent with a graphical interface to do one real task — make it do exactly one thing you used to do yourself (clean files, make Excel, write a report). Don't demand complex automation right away.
3. Understand the execution process — watch it plan, call tools, read files, handle errors. This way you grasp the relationship between model – tools – permissions – context.
4. Go deep based on real needs — want to build software then learn coding Agent; want Agent to have extra specialist skills then explore Skills, MCP, tool connections; need to run remotely/on schedule then research further. No need if there's no need.

When starting, just pay attention to 3 things: choose a workspace (one folder where all your work files sit neatly), grant permissions (beginners should give broad permissions to avoid clicking approve at each step — with sensitive data, reconsider), pick a model (start with free/cheap models, upgrade only when you see it's not enough; rough rule but works: higher price usually means better results).

Conclusion: AI isn't about learning to be good at it, it's about using it to get good

Four puzzle pieces are enough:

• Large models: understand the nature, how they work, strengths and weaknesses.
• Prompts: techniques and ways to talk with AI.
• Agents: a framework to recognize every new tool.
• Practice: complete your first real task.

Three core things to remember:

• AI is a tool, not magic — know its boundaries.
• The essence of prompts is stating your needs clearly — not memorizing formulas.
• Current agents aren't truly "intelligent" — but they already solve many real tasks, what matters is finding the right place to use them.

Don't fall into the loop of endless learning. Dive in, use it, make mistakes, fix them, solve the work sitting on your desk. Each time you struggle, each time you adjust, you'll understand AI one level deeper.

From today, make AI a daily tool: write weekly reports, organize documents, learn something new, make plans, analyze data — try asking AI first for anything.

This post was inspired by the curriculum "Understanding AI from Zero in 2 Hours" by the author 龙井 (longjing), rewritten and adapted for the Vietnamese context.