By the WahLiao desk · Last verified 1 October 2026
A Super Intelligence chatbot, the kind of AI behind ChatGPT, Claude, Gemini and Copilot, works by prediction. It has learned from an enormous amount of text how language usually runs, and when you ask a question it builds an answer piece by piece, choosing what most plausibly comes next. People then tune it to be helpful and to refuse harmful requests. Unless it is connected to web search, it is not looking anything up, and it has no built-in sense of when it is wrong.
That one idea explains nearly everything: why it can draft a polite letter to your town council in seconds, why it can explain CPF in simpler words than the letter did, and why it can also name an MRT station that does not exist with total confidence.
How it works: quick facts
| Training | The model learns patterns from very large collections of text, code, images or audio |
| Prediction | It writes an answer a small piece at a time, picking what most plausibly comes next |
| Tuning | People rate its answers so it becomes more helpful, polite and safe |
| Search | Some tools can search the web and show sources; open them anyway |
| Mistakes | “Hallucinations”: fluent answers that are wrong or invented |
| Images and voice | Built in a similar way, but learned from pictures and recordings |
The WahLiao Verdict
| Best at | First drafts, summaries, explanations, translation and brainstorming. |
| Weakest at | Exact facts, fresh news without search, and knowing its own limits. |
| The habit to build | Ask for sources, then open them. |
| Never paste in | Passwords, NRIC numbers or anything you would not email a stranger. |
| Remember | Fluent is not the same as correct. |
Step one: training
Before a model can answer anything, it is trained. Engineers feed it a vast amount of text, from books and websites to computer code, and the model gradually adjusts billions of internal settings until it becomes very good at one narrow task: guessing the next piece of a passage. Along the way it absorbs grammar, facts, styles, arguments and the shape of good explanations.
What it does not do is store a tidy encyclopaedia. Knowledge ends up spread across those settings as patterns, which is why a model can recall a fact well in one conversation and garble it in the next, and why its knowledge stops at the point its training data ended.
Step two: prediction
When you type a question, the model breaks it into small chunks called tokens, roughly pieces of words, and produces a reply one token at a time. Each choice depends on everything before it: your question, instructions from the company that runs it, and the words it has already written. Think of your phone’s autocomplete, grown to the size of a library and given the patience to plan a whole paragraph.
Step three: tuning
A raw model will happily continue any text, helpful or not. So companies tune it, largely by having people compare and rate answers, until it behaves like a useful assistant that follows instructions, admits some limits and declines harmful requests. This tuning is why different chatbots have different personalities, even when they learned from similar material.
Why it makes confident mistakes
Because it predicts plausible text rather than retrieving checked facts, a model can produce an answer that sounds right and is wrong. The industry calls these hallucinations, and Singapore’s Model AI Governance Framework for Generative AI, published by IMDA in 2024, lists them among the new risks generative tools bring. The danger zones are precise things: prices, dates, names, legal rules, citations and medical doses. The model does not change its tone when it is unsure. It sounds just as confident when it is inventing.
Two habits help. Ask the tool to give sources, and if it can search the web, open the links it shows. And for anything with consequences, such as a contract, a tax filing or your health, treat its answer as the start of a question to a person or an official page, not the end of one.
When it searches the web
Many chatbots can now search the internet, read the results and summarise them with links. That makes them far better at recent events and current prices, but not infallible: they can misread a page, lean on a weak source or blend two sources into one wrong sentence. The links are there so you can check.
Image makers, voice tools and agents
Image generators learn from huge numbers of pictures paired with descriptions. Given your prompt, most begin with visual noise and refine it, step by step, into an image that matches the words. Voice tools learn the sound and rhythm of speech, and some can imitate a particular person from a short recording, which is why a call in a relative’s voice asking for money should always be checked by calling them back on a number you know.
Agents go a step further than chat. They can take actions, such as browsing, filling in forms or working through a task in several stages. That is useful, and it makes it more important to check what an agent did before anything is paid for or sent.
Good at, bad at
| Good at | Bad at |
|---|---|
| Drafting emails, letters and reports | Exact figures and current prices, unless it searches |
| Explaining a hard idea simply | Knowing when it is wrong |
| Summarising long documents | Citing sources it has not actually read |
| Translating and rewording | Local rules that changed recently |
| Brainstorming options | Judgment calls that need your context |
How AI works: FAQ
How does an AI chatbot like ChatGPT work?
It has learned patterns from a vast amount of text and writes answers by predicting, one small piece at a time, what most plausibly comes next. People then tune it to be helpful and safe.
Does AI look things up?
Only when the tool is connected to web search or to documents you give it. Otherwise it answers from patterns learned in training, which can be out of date.
Why does AI make things up?
Because it generates plausible text rather than retrieving checked facts. These errors are called hallucinations, and they sound as confident as correct answers.
Can AI copy someone’s voice?
Yes. Some voice tools can imitate a person from a short recording. If a call in a familiar voice asks for money or codes, hang up and call back on a number you already know.
Read next
This page belongs to Super Intelligence. Next, see what the tools cost: paying for Super Intelligence in Singapore. To protect your savings from scams, The Ledger explains Money Lock.
Sources checked 1 October 2026: Budget 2026 statement, section C; Morgan Lewis, summary of IMDA’s Model AI Governance Framework for Generative AI (2024). General information. Last updated 1 October 2026.
