Understand AI from Zero in 2 Hours: A Mental Framework So You Never Feel Overwhelmed Again

2 Hours Understanding AI from Zero: A Framework to Never Feel Overwhelmed Again

Every week there's a new AI tool. Today Claude Code, tomorrow Codex, next week some other strange 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 post goes the opposite direction. I'm not teaching you which buttons to click, I'm building you a mental framework. Once you have the framework, whenever any AI tool comes out in the future, you'll look at it once and understand what it is, what it can do, and whether it's worth learning.

After reading this post, you'll have 4 puzzle pieces: understand large models — write prompts — understand Agents — get hands-on building.

---

Part 1 — Understanding "Large Models" Without Being an Engineer

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

AIGC (Generative AI) is a type of AI that creates text, images, or videos exactly as you describe them, and this is what's exploding right now. AI has existed since the 1950s: facial recognition on phones, TikTok's recommendation algorithm, AlphaGo playing Go — these are all AI, but nobody calls that an "AI craze". The game-changing difference with OpenAI's ChatGPT or Anthropic's Claude is that they speak human language: you chat in everyday language, they understand and create content. For the first time, AI has stepped directly into the lives of ordinary people.

My advice for beginners is to start with text → text, meaning describe in text and receive text back. First, this is the lowest barrier to entry. Second, this is the most common type of application, and third, it's the type with the most free tools available. For example, asking ChatGPT or Google's Gemini to summarize a long email is the simplest text → text exercise.

What makes a "large model" large

Behind every AI app (ChatGPT, Claude, Gemini, DeepSeek...) are large language models (LLMs). To understand the word "large", imagine raising a child from kindergarten to a doctorate:

• Large data = reading entire libraries. The wider the data, the more complete the model's "understanding" 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 are not researchers. The goal of understanding LLMs is to know the boundaries of its abilities, not to dive into deep principles.

How AI works: like a super-memory librarian, but understands nothing

Don't imagine AI "thinking" like a person. Picture it as a librarian with a gigantic memory but doesn't understand the content of the 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-linking game like keyboard suggestions on a phone — but millions of times more complex. It picks the next piece according to the highest probability.
3. It doesn't truly understand what it's saying. No experience, no common sense, no awareness that it might be wrong.

The consequence: it can write beautiful poetry and also "make things up with complete seriousness". Humans are the final quality checkpoint to see if the answer makes sense — not AI.

Strengths and weaknesses: why is it "human-machine collaboration"?

Human-machine collaboration means using AI's strengths to compensate for your weaknesses, not "AI does everything" nor "humans do everything". AI's strengths include processing large volumes of information (reading, summarizing, sorting), broad knowledge across all fields, and generating new content. Five weaknesses you must remember:

• Hallucination: AI predicts probabilities rather than asserting truth, so it can fabricate citations or events.
• Knowledge cutoff: its knowledge stops at training time, it doesn't update itself.
• Amplifying existing biases in the training data.
• Multi-step reasoning has gaps, easy to fail on problems requiring many conditions.
• Context length limitation: talk too long and AI forgets the beginning.

For example, when asking Claude or ChatGPT to cite a study, check the source yourself. So when using it, give clear instructions, provide context, cross-check important information, break complex problems into pieces, and always keep a healthy skepticism.

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

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

Like smartphones: phone brands correspond to AI apps, chip makers correspond to model makers, chip models correspond to model names and versions. Chips matter a lot, but what the phone has depends on the operating system, camera, ecosystem. AI apps are the same: with the same model, two apps can deliver very different experiences because the "shell" differs (web search, file reading, tool calling, remembering preferences...).

Be wary of software "wearing a disguise": a small team builds a shell and plugs in an API, advertising "using top-tier GPT" but actually running a cheap model. Choosing an app: there's no best, only what's best for you — consider your actual needs, acceptable costs, the real model behind it, and whether the team is trustworthy.

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

There are 3 ways to upgrade AI with different difficulty levels: prompt, fine-tune, and RAG. Many people ask "can I train my own model?", and that's a classic misunderstanding. The three ways are:

1. Prompt (instructions): cheapest, anyone can do it, used daily, just learning how to ask better.
2. Fine-tune: lets AI "learn more" to become an expert in one field, by adjusting a small part of parameters with specialized data. This is for technical teams, difficult and expensive.
3. RAG (knowledge base): attach a "living dictionary" to AI; when you ask, AI searches the document repository first then answers. RAG solves outdated knowledge and private business data.

For example, a customer care chatbot that reads the company's product documentation is RAG. How to choose? Prompt for everyone and daily tasks, RAG for businesses needing private data, fine-tune for technical teams when prompts don't cut it. For regular people, fine-tune is the most distant and not needed yet.

How to choose an AI model that suits you by what criteria?

There's no "best" model, only the most suitable one, and there are five criteria to compare. The criteria include:

1. Need to "cross borders": many international models like those from OpenAI, Anthropic, or Google are not easy to access in Vietnam.
2. Price: international models usually start at $20/month, and if you're using APIs then cost becomes a big deal.
3. Speed: the more "deeply thinking" a model is, the slower; simple tasks need a fast model.
4. Context length: determines how deep a discussion you can have before AI forgets. Longer context doesn't mean better memory. A trick is to regularly ask AI to summarize conclusions and save them to a document.
5. Field specialty: each model is strong in different areas — writing, code, reasoning, or analyzing long documents.

The real way to check is to build your own set of "test cases" with a few scenarios you use regularly, then retry whenever a new model appears. At this stage, no single model meets everything, so keep a few models on hand.

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

What is the basic prompt formula?

The basic prompt formula is: quality prompt = clear command + necessary context. The command is the soul of the prompt: it says clearly what you want the AI to do, in what way, and to what standard. Context is the background information that helps the AI understand correctly.

For example, instead of "write me a phone introduction post", write: "Gaming phone for young people, price around 3 million đồng, strengths are fast charging and heat dissipation; write an introduction post focusing on gaming performance and reasonable price, youthful tone but not exaggerated, under 800 words." The second sentence has command, context, tone and length limit, so ChatGPT or Claude both give much better results.

Is "prompt engineering dead"?

No. Reasoning models (like DeepSeek R1) make many people say you only need to "say it clearly". True, but: (1) not everyone can say things clearly, (2) AI is not a person anyway, so saying things clearly to it still needs some technique. Prompt is not dead — it's "democratized": ordinary people can use it with natural language, while professionals still use more sophisticated prompts. And the stronger the model, the more we don't need to write one perfect prompt right from the start — you can have back-and-forth conversations like talking to a colleague.

What are the 6 techniques for writing effective prompts?

The six techniques for writing effective prompts are task, context, examples, persona, format and tone. According to longjing's curriculum, specifically:

1. Task: say clearly the specific work, like "write a popular science article about the harms of plastic waste", not "write something".
2. Context: let the AI know why you're doing this and for whom.
3. Examples: give a few samples so the AI understands the right "taste" you want; this trick works great for scenarios and captions.
4. Persona: "Suppose you are a middle school science teacher, explain in simple language".
5. Format: ask for Markdown, tables or JSON so it's easy to reuse.
6. Tone: formal for reports, natural for social media.

You don't have to use all 6. Every technique serves one goal: getting the right result.

My method: "Ask high, pull low"

Beginners often learn enough techniques but sit in front of the chat box with a blank mind. The fix:

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

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

Two tricks of "lazy but smart"

• Start with imitation: like learning to write, you have to read sample texts — look at prompts from skilled people, analyze the structure, understand why they did it that way, then edit it to fit your needs.
• Use AI to write prompts: my main approach is "just say it": "I want to do XXX, write me a prompt." Or tell it to copy: "Imitate the structure and tone of this prompt, write one for me about YYY." Don't demand perfection right away — let AI give you a 60-point version and iterate gradually.

Part 3 — What is Agent (AI Agents)

Core Difference

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

Why are there so many Agent tools?

Like Lark / DingTalk / WeCom all being "team collaboration software" just from different companies — when a product form gets market acceptance, companies flood in to make their own. Many tools have different names but solve the same problem, interact the same way and share similar underlying logic.

A 3-Dimensional Framework to Identify Any New Agent

The 3-dimensional framework is three questions that help identify any new Agent: which company made it, how you access it and what work it can do. First is which company made it, for example OpenAI, Anthropic, big companies or open-source communities; the company that makes the model and the company that makes the Agent product are not necessarily the same. Next is how you access it, which includes four forms: chat app (ChatGPT), desktop interface (easiest for regular people), command line CLI (Claude Code, Codex CLI) and through chat apps like Telegram or Zalo. Finally is what work it can do: general office Agent (write documents, analyze data, make slides), programming Agent (read code, fix bugs, run tests) or specialized Agent. These forms are not mutually exclusive, and one product can have multiple entry points.

Underlying Logic: Execution Architecture + AI Brain

Every Agent is fundamentally the same structure: execution framework (read files, run commands, call tools) + a large model as the brain. The difference is only in two points: what tools the execution framework can call, and which model is plugged in as the brain.

Tools determine what you can do. The brain determines how well you can do it.

Which Agent is best?

This question is meaningless outside specific needs. The right approach has three steps. First, identify your actual work. Next, choose a tool but don't rush to buy a plan; take 2–3 real tasks and compare. Finally, evaluate using five questions: Does the Agent complete the work or just give suggestions, how accurate are the results and how much do you need to fix, how much time does it take or what's the standard, does it support the software you need, and what permissions are needed with what data risks. A truly good Agent is one that completes work steadily and economically, not one that looks smart. For example, if you want Claude Code or Codex to clean up a folder of 100 files, test that exact task before paying.

Part 4 — Getting Started: A Roadmap for Complete Beginners

Don't exhaust yourself learning each tool's interface — buttons move tomorrow. Learn the common way of working across all tools, through 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 task you normally do yourself (clean files, make Excel, write reports). Don't demand complex automation right away.
3. Understand the execution process — watch how it plans, calls tools, reads files, handles errors. Through this you grasp the relationship between model – tools – permissions – context.
4. Dive deeper based on actual needs — want to build software then learn programming Agent; want your Agent to have extra skills then look into Skills, MCP, tool connections; need remote execution/scheduling then research further. No need is no need to force learning.

When starting, just pay attention to 3 things: choose a workspace (one folder to keep everything organized), grant permissions (beginners should give broad permissions to avoid clicking approval for each step — with sensitive data then reconsider), choose a model (start with free/cheap models, upgrade when you see it's not enough; crude but effective rule: higher price usually means higher effectiveness).

Conclusion: AI isn't about learning it well, it's about using it that makes you good

Four pieces of the puzzle are enough:

• Large models: understand the essence, how they work, strengths and weaknesses.
• Prompt: techniques and methods for talking to AI.
• Agent: a framework for recognizing 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 prompting 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 trap of learning endlessly without finishing. Jump in, use it, make mistakes, fix them, solve the work sitting on your desk. Every time you get stuck, every time you adjust, you'll understand AI one layer deeper.

From today, make AI your everyday tool: write weekly reports, clean up documents, learn something new, make plans, analyze data — try asking AI first for anything.

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