How does AI like ChatGPT actually work, in simple terms?

At its core sits one surprisingly simple idea: it is a prediction machine for text. It was shown a colossal amount of human writing - books, websites, code, conversations - and learned the patterns deeply enough to answer one question extremely well: "given these words, what plausibly comes next?" Everything it does - answering, drafting, translating, coding - is that one trick, applied over and over, word by word. The genuinely surprising discovery of this decade is that predicting text this well turns out to REQUIRE absorbing huge amounts of how the world works - grammar, facts, reasoning patterns, styles - which is why a "next word guesser" can write your email better than most humans. It is not magic, and it is not thinking like you do; it is pattern mastery at a scale no human can hold.

Why does AI sometimes confidently make things up?

Because of what it fundamentally is. A prediction machine produces what is PLAUSIBLE, and truth and plausibility usually - but not always - overlap. When you ask about something it knows well, plausible and true coincide. When you ask about something obscure (a specific court case, a niche statistic, your company’s policy), it still produces plausible-sounding text - fluent, structured, confident - because that is its one move. The industry calls these "hallucinations". Two practical implications for using it: never treat it as a lookup service for facts you cannot verify, and notice that it is most dangerous precisely when it sounds most certain, because confidence is part of the style it learned, not a signal of accuracy.

What is AI genuinely good and bad at right now?

Genuinely strong: transforming things that already exist - summarising, rewriting, translating, drafting from your notes, explaining concepts at any level, and writing working code from clear descriptions (this last one is reshaping an entire profession). Genuinely weak: facts it cannot check (hallucination risk), knowing what it does not know, long chains of precise reasoning without drifting, anything requiring real-world verification, and judgement calls where being wrong has real consequences. The practical rule most professionals converge on: AI for the first draft, human for the decision - which is why every serious workplace use pairs it with a person who can evaluate the output.

Will AI keep getting smarter? Should I be worried?

It has improved rapidly and continues to - each generation is notably better at reasoning, coding, and reliability. Honest experts disagree about the ceiling, so treat anyone claiming certainty (in either direction) as selling something. The practical response does not depend on resolving that debate: in every scenario, the people who do well are the ones fluent at directing and judging these tools, and the people at risk are those who ignore them or trust them blindly. History has run this pattern before - the spreadsheet transformed office work the same way, rewarding the adapters - and our essay "AI is the new Excel" lays out that precedent with numbers. Worry is optional; fluency is not.

How can I go from understanding AI to actually using it well?

A ladder, and you can start today. Rung one: use it daily on real tasks - not tests, real work: drafting messages, planning, summarising documents, learning things - and push past chat into "make me a working tool" requests. Rung two: learn to direct it precisely - specific instructions, examples of what you want, iterating on results; this skill alone separates power users from dabblers. Rung three - where careers change: learn enough building skill to create things WITH it: automations, tools, real software, with AI doing heavy lifting and you doing the judging. That third rung is what Malaysian employers now pay a documented premium for, it takes months rather than years, and the free trial is its zero-cost first step.

All posts

Start here · Curious about tech · AI explained

How does AI actually work? One idea explains almost everything.

You use it, your boss talks about it, the headlines alternate between miracle and menace - and most explanations are either baby-talk or mathematics. Here is the honest middle: the one core idea behind ChatGPT and its siblings, why it produces both brilliance and confident nonsense, and what it actually means for your work. No equations, no hype.

Deric Yee Updated 25 August 2026 9 min read

Glowing abstract lights - a fitting portrait of a neural network

The one idea: a prediction machine for words

Strip away every layer of branding and here is what sits at the centre of ChatGPT, Claude, Gemini, and the rest: a machine trained to answer one question - given these words, what word plausibly comes next? It was built by showing a computer system a colossal share of human writing - books, websites, forums, code - and having it practise that prediction billions upon billions of times, adjusting itself after every miss. Your phone keyboard does a toy version when it suggests your next word. The breakthrough of this decade was discovering what happens when you scale that toy up by a factor of millions: to predict text really well, the machine is forced to absorb how the world works - grammar, facts, argument structures, the difference between a legal contract and a love letter - because all of that shapes what word comes next. When you ask it something, it is not looking up an answer; it is generating the most plausible continuation of a conversation in which your question was just asked, one word at a time, faster than you can read.

Hold onto that phrase - plausible continuation - because it explains both halves of your experience with these tools: why the writing is so fluent (plausibility is literally the thing it is optimised for), and why it can be so confidently wrong. Which brings us to the blind spot.

Why it makes things up - and why that is not a bug being fixed next week

Ask an AI about something well-covered in its training - how photosynthesis works, how to write a resignation letter - and plausible and true point the same way, so you get accuracy. Ask about something obscure or specific - a particular Malaysian court case, your company’s leave policy, a statistic nobody wrote down - and the machine does the only thing it knows: produces the most plausible-sounding text anyway - complete with fluent structure and a confident tone, because confidence is part of the writing style it learned. The industry calls these hallucinations, and they are not carelessness; they are the direct shadow of the core design. Modern systems reduce them (by checking sources, citing documents, saying “I’m not sure” more often), but the deep lesson for any user is permanent: fluency is not evidence. The tool is a brilliant drafter and a dangerous oracle - and knowing which one you are talking to at any moment is the whole skill of using it safely.

What this means for your work (the Excel parallel)

Once you see AI as a plausibility engine, the workplace picture stops being mysterious. It is superb at transformation - summarise this report, draft this email from my notes, translate this, explain this contract clause simply, write code that does X - because transformation keeps it anchored to material you gave it. It is unreliable as a fact source you cannot verify, and it cannot own a decision - it does not know your context, cannot be accountable, and will not notice what it got wrong. So every serious professional use converges on the same shape: AI drafts, a human judges. That division of labour is precisely why this technology is compared to the spreadsheet - a tool that did not delete office jobs but redrew them, punishing those who refused to learn it and promoting those who adapted early. We tell that story, with the actual employment numbers from the spreadsheet era and this one, in AI is the new Excel - the essay to read after this one if the what-about-my-job question is already forming.

From understanding it to being good with it

Understanding how the machine works puts you - genuinely - ahead of most of its users. Converting that into advantage is a ladder. Use it daily on real things, and push past chatting: ask it to build you a working tool (a budget tracker, a study planner) and watch it happen - the moment most people first feel what this era makes possible. Learn to direct it precisely - specific asks, examples, iteration - which is a learnable craft, not a knack. And if the curiosity has teeth, take the rung that changes careers: learn enough building skill to create real things with AI as your workforce- the combination Malaysian employers now pay RM 6,000–9,000/month for at entry level (the data). The zero-cost on-ramp is the free trial - real projects, a patient AI tutor (now you know how it works), a live human instructor - and the plain-language map of what building leads to is what can I build with AI.

FAQ

How does AI like ChatGPT actually work, in simple terms?
At its core sits one surprisingly simple idea: it is a prediction machine for text. It was shown a colossal amount of human writing - books, websites, code, conversations - and learned the patterns deeply enough to answer one question extremely well: "given these words, what plausibly comes next?" Everything it does - answering, drafting, translating, coding - is that one trick, applied over and over, word by word. The genuinely surprising discovery of this decade is that predicting text this well turns out to REQUIRE absorbing huge amounts of how the world works - grammar, facts, reasoning patterns, styles - which is why a "next word guesser" can write your email better than most humans. It is not magic, and it is not thinking like you do; it is pattern mastery at a scale no human can hold.
Why does AI sometimes confidently make things up?
Because of what it fundamentally is. A prediction machine produces what is PLAUSIBLE, and truth and plausibility usually - but not always - overlap. When you ask about something it knows well, plausible and true coincide. When you ask about something obscure (a specific court case, a niche statistic, your company’s policy), it still produces plausible-sounding text - fluent, structured, confident - because that is its one move. The industry calls these "hallucinations". Two practical implications for using it: never treat it as a lookup service for facts you cannot verify, and notice that it is most dangerous precisely when it sounds most certain, because confidence is part of the style it learned, not a signal of accuracy.
What is AI genuinely good and bad at right now?
Genuinely strong: transforming things that already exist - summarising, rewriting, translating, drafting from your notes, explaining concepts at any level, and writing working code from clear descriptions (this last one is reshaping an entire profession). Genuinely weak: facts it cannot check (hallucination risk), knowing what it does not know, long chains of precise reasoning without drifting, anything requiring real-world verification, and judgement calls where being wrong has real consequences. The practical rule most professionals converge on: AI for the first draft, human for the decision - which is why every serious workplace use pairs it with a person who can evaluate the output.
Will AI keep getting smarter? Should I be worried?
It has improved rapidly and continues to - each generation is notably better at reasoning, coding, and reliability. Honest experts disagree about the ceiling, so treat anyone claiming certainty (in either direction) as selling something. The practical response does not depend on resolving that debate: in every scenario, the people who do well are the ones fluent at directing and judging these tools, and the people at risk are those who ignore them or trust them blindly. History has run this pattern before - the spreadsheet transformed office work the same way, rewarding the adapters - and our essay "AI is the new Excel" lays out that precedent with numbers. Worry is optional; fluency is not.
How can I go from understanding AI to actually using it well?
A ladder, and you can start today. Rung one: use it daily on real tasks - not tests, real work: drafting messages, planning, summarising documents, learning things - and push past chat into "make me a working tool" requests. Rung two: learn to direct it precisely - specific instructions, examples of what you want, iterating on results; this skill alone separates power users from dabblers. Rung three - where careers change: learn enough building skill to create things WITH it: automations, tools, real software, with AI doing heavy lifting and you doing the judging. That third rung is what Malaysian employers now pay a documented premium for, it takes months rather than years, and the free trial is its zero-cost first step.

Back to all posts What it means for jobs: the Excel story →

You now know how the machine thinks.
The next step is making it work for you.

Real projects with AI as your workforce and a live instructor as your guide - the free trial converts understanding into capability. One signup, no card.

Start the free trial The Excel parallel, in full