What AI really means
The word is everywhere now. In Facebook reels, in newspapers, in office meetings, even in talk at the tea stall. But in most places the word is said as if it were magic, or a robot from a film. The real thing is much plainer.
AI is software. That means a program that runs on a computer, the way the bKash app or Facebook are each a program. The difference is in one place. In ordinary software, a person writes the rules by hand: "in this situation do this, in that situation do that". With AI, people do not write the rules. Instead they put endless examples in front of it, and the program looks at the examples and works out the rules for itself.
Think of a comparison. You give a new shop assistant a written rule, "if someone asks for rice, give it from the sack on the left". That is ordinary software. But if you leave him sitting in the shop for six months, he works out for himself what each customer will ask for. That is the AI way. Nobody wrote the rule. He recognised the pattern.
So there is no need to think of AI as an intelligent creature. It is a pattern-recognising machine. But the power of pattern recognition has grown so large that the results are surprising.
The word "AI" is a kind of umbrella, and many things come under it. In this article we mainly talk about the kind that understands text and produces text, things like ChatGPT or Gemini.
How AI learns
The learning happens in a step called Training. What is done to build an AI like ChatGPT or Gemini can be said in about three lines.
First, a huge amount of text is collected from the internet. Books, websites, forum discussions, code, news archives. So much text that one person could not finish reading it in a whole lifetime.
Then the program is asked to play one game, hundreds of thousands of times. The last word of a sentence is hidden, and the program is told: guess the next word. What comes after "a terrible ... on the roads of Dhaka"? It guesses. If it is wrong, the calculations inside are changed slightly, so that next time it is a little less wrong.
When this game runs tens of millions of times, a strange thing happens. To guess the next word well, you need the grammar of the language, you need knowledge of the subject, you need the flow of reasoning. So while learning patterns, the program builds a blurry map of language, facts and reasoning inside itself. Nobody taught grammar separately. It came from the pressure of guessing words.
In the last step, people sit and look at thousands of answers and give a verdict on which is good and which is bad. This is the step where AI learns to reply politely, to avoid some questions, and to stay away from abuse. It has a side effect too. Since people gave good marks to tidy answers, the Model learned that saying something in a tidy way is the safe choice, so the habit of saying "I do not know" stayed weak.
What a Token really is
So far I said "word" for convenience. The real counting is not done in words. It is done in Tokens.
A Token is a small piece of text. Sometimes a whole word is one Token, and sometimes a word breaks into two or three pieces. A Model never sees a whole sentence. It sees a long row of pieces, and works out what the next piece should be.
Knowing this makes two things suddenly clear.
One, on questions like "how many letters are in this text" or "what is the third letter from the end", AI often gets it wrong. That is because it is not counting letters, it is looking at pieces. Counting letters is a job your phone's ordinary text editor can do better than it.
Two, the same thing written in Bangla script usually uses more Tokens than in English. Because there is more English text on the internet, the way of cutting English words into pieces has become smoother. The result is that in a long Bangla conversation, room runs out a little faster, and since AI services keep their own accounts in Tokens, it counts in the cost as well.
Training and the moment you type are two different things
This is the biggest wrong idea among beginners. Many people think: AI is learning from what I write, so if I correct it, it will get it right next time.
That is not how it works. Training is the factory work. Huge numbers of computers, huge amounts of electricity, a long time. When that step ends, the machine comes out finished, and the numbers inside it no longer change.
When you write in a chat, the factory is not running. The finished machine is simply being run. What you write does not change anything inside it, just as doing a sum on a calculator does not change the circuits inside the calculator.
Then why does it correct itself when you point out a mistake? Because that correction is now written in front of it, and it counts the text in front of it. Inside that same chat it will get it right. If you go to a new chat and ask, you will get the earlier mistake again.
Some apps give a separate feature called "memory", where your preferences are stored like a note and put in front of it in the next chat. That is not the Model learning either. That is note keeping.
What a Model really is
What you get when Training ends is the Model. GPT, Gemini, Claude, Llama: each of these is the name of a Model.
A Model is not a store of facts. Inside there is no table saying "capital of Bangladesh = Dhaka". Inside there are only numbers. Billions upon billions of numbers, called Parameters. During Training, every time it was wrong, it was these numbers that changed a little at a time.
Look at it this way. Training is the years of learning to cook, and the Model is the skill that stays in the cook's head after the learning is done. There is no recipe book separately, but the feel in the hand has formed.
Two important things come out of this. One, what a Model "knows" is blurry, so even when it remembers the main point, it is very normal for it to mix up a specific number or date. Two, Training is done with data up to a certain time, so events after that are not something the Model can be expected to know. This boundary is called the Knowledge cutoff.
A third thing comes from here too. The more text there is on the internet about a subject, the firmer the Model's guess. Good on the city of Dhaka, thin on the market in your union. Before you ask, ask yourself once: should there be a lot of text about this subject on the internet?
Why it forgets everything when a chat ends
Many people have this experience. Yesterday you had a long talk and explained the whole identity of your shop. Today in a new chat you write "give me another post like yesterday's", and it recognises nothing.
The reason is simple. Inside the Model there is no notebook of memory. Every time it replies, the whole text of that chat so far is put in front of it again, fresh, and it reads all of it and then decides the next piece.
So it is not remembering. It is re-reading every time. The text held in front of it has a maximum size. Once past that size, the earliest parts start to drop out, so in a very long chat it seems to forget the instructions from the beginning.
Knowing this one thing will change how you work.
- Keep one chat for one job, so the text held in front of it stays relevant.
- Write your business's identity in three lines in your phone's notes, and paste it first thing in a new chat.
- When a chat gets very long, say "write the decisions so far in 5 points", then take those five points to a new chat.
What "guessing the next word" means
Many people are surprised to hear that a machine that only guesses the next word can do so much.
The funny thing is, guessing the next word well is a huge job. To say the next word correctly after "Ruman's shop sold 4200 taka yesterday and 3800 taka today, so in two days the total is...", you have to know how to add. To carry on the sentence "the patient has had a fever for three days, so the doctor advised...", you have to understand some of the context.
AI does not stop after placing one word. It reads the whole text again, including the word it just placed, and then decides the word after that. Word after word joined like this, the whole answer stands up. When you see the answer coming in slowly as if being typed, this is exactly what is happening in front of your eyes.
For this same reason, for AI "sure" and "guess" are not different things. In both cases it places the next word in the same way. So a wrong answer sounds just as confident as a correct one.
Another thing can be understood from here too. Each time it picks the next piece, a little randomness is kept on purpose, otherwise the answers would sound mechanical and dull. That is why asking the same question twice gives two different answers. This is not a flaw, it is the design.
Then why does it seem that AI understands
You write in Bangla, "my shop's accounts have got tangled and I do not know what to do". The answer comes tidily, with sympathy, step by step. It is natural to feel that someone on the other side understood you.
Several reasons make it seem so.
- Its command of language is remarkable. A machine that has learned patterns from hundreds of billions of sentences can build sentences cleaner than a person's.
- Because people like polite and sympathetic answers, AI was polished that way in the last step. This sympathy is not acting, and it is not a feeling either. It is a taught pattern.
- It is the habit of our brain to imagine a person behind language. When we see tidy writing, we assume there is thinking inside.
So does AI understand nothing? The honest answer is that even scientists do not agree on this. What is safe in practice is to think of it this way. AI understands the relationship between language and ideas well, but it has no account of true and false. It knows which words fit after which words, but it has no way to verify whether the thing actually happened.
Ordinary software and AI: where is the difference
| Point | Ordinary software | AI Model |
| Who decides the rules | A person, writing by hand | The Model itself, from examples |
| Answer to the same question | Exactly the same every time | May differ a little each time |
| When it is wrong | A specific bug, which can be found and fixed | Hard to explain why it was wrong |
| In an unfamiliar situation | Stops if it was not taught | Guesses something close and says it |
| Trusting the result | Certain if the calculation is right | Needs a look to verify |
The last row is the most important. A calculator will not get a big multiplication wrong. AI can, especially with big numbers. So think of AI as a worker, not a calculator. A worker's job needs a look over it when it is done.
What AI does well and what it cannot
The jobs it does well have one thing in common. In them "roughly right" is enough, and if it is wrong, you can catch it straight away.
- Tidying text, shortening it, changing it into polite language
- Translation from Bangla to English or the other way
- Explaining a hard subject simply
- Making a list of ideas, such as 20 names for a shop
- Writing code and catching mistakes in code
Where it is not right to trust it, there is a common thread too. These need a specific number, name or date, and a mistake causes harm.
- Today's market price, today's news, a new government rule
- Medicine doses or diagnosing a disease
- Legal sections, case details, explaining a land deed
- Someone's personal information or an organisation's internal accounts
- Large sums of accounting, where one taka off is a problem
Which tool for which job
AI is not a substitute for every tool. Picking the wrong tool is the biggest waste of time for beginners.
| What you need | The best tool | Why |
| Today's gold price or a train time | A search engine or the official site | Specific and changing information, and the Model may have it out of date |
| The total of 37 sales | A calculator or a spreadsheet | The sum must be exact, a guess will not do |
| A summary and advice from a list of 37 sales | AI | Recognising patterns and putting them in words is its strength |
| Fixing the English of a letter | AI | If it is wrong, you can read it yourself and catch it |
| Explaining a land deed or a case | A professional | Someone has to take responsibility, and no Model takes it |
| Learning something new, understanding an idea | AI, then a book or a course | Good at explaining ideas, but specific facts have to be checked |
| Why a fever for three days | A doctor | The price of a mistake is paid with the body |
What to do in the first seven days, step by step
Enough theory. Now a format to start with in hand.
- Open a free account. A phone is enough, you do not need a computer. Start in the browser, or download the official app.
- Pick a first job whose result you can verify yourself. Copy one of your own old Facebook posts, and write, "rewrite this in 3 lines, more attractive".
- If you do not like the answer, do not stop, have it fixed. "Make it shorter", "use fewer English words", "say the price bigger". Two or three rounds of polishing will change the result.
- Write a three-line identity and keep it in your phone's notes. Who you are, what your business or studies are, who your readers are. Paste this at the start of every new chat.
- For one week, give it one real job every day. Not made-up questions, your real work. This is where the learning happens.
- Once a day, try to catch it in a mistake on purpose. Ask a question you know the answer to, then check the answer. Seeing for yourself where it is weak is the best kind of learning.
- At the end of the week, count up. In which jobs did time get saved and in which did it grow instead. This list is your own rulebook for use.
A saree shop in Mymensingh, a full example
Suppose there is a small saree shop in Mymensingh, called "Nilanjana Saree Ghar". The owner posts photos on Facebook every day, but when she sits down to write, half an hour goes. On the first day she wrote,
Write a post for a saree shop.
What came was nice to look at, but not useful. Shiny advertising language mixed with English, no price anywhere, no name of the area anywhere, and a row of emoji at the end. Her customers do not talk in this language.
On the second day she explained the whole thing and wrote,
I run a saree shop in Mymensingh town, called "Nilanjana Saree Ghar". A new stock of Tangail handloom sarees has come, priced from 1800 to 4500 taka. The customers are mainly women aged 30 to 50, mostly local. Write a post for Facebook. Conditions: at most 4 lines, simple spoken Bangla, as few English words as possible, keep the price range exactly, and the last line must say "send a message in the inbox". Give 3 options in different tones.
This time three options came. One very emotional, one straightforward, one starting with a question. She picked the straightforward one, changed one word, and checked the price figures against her own notebook.
Three things worth noticing happened here.
First, the first attempt failed, but the reason was not AI's weakness. The first instruction had nothing in it: where the shop is, who the customers are, what the price is. When there is a gap, it fills the gap, and the gap it fills is like the average internet, not like your Mymensingh.
Second, the long instruction took about five minutes to write, but she saved it in her notes. The next day she only has to change the product and the price, and that takes one minute. The cost is once, and the gain is every day.
Third, in one place it made a mistake. In one option it had written on its own "delivery all over the country", but she does not deliver. If nobody had pointed it out, it would have been posted, and people would have come to the inbox with wrong expectations. This is why the last pair of eyes has to be a person's.
A coaching centre in Rajshahi, a freelancer in Chattogram and a student in Dhaka
A coaching centre in Rajshahi. It teaches class nine science, with 22 in a batch. A test has to be given every week, and setting the questions takes the teacher's Sunday evening. Now she pastes in the chapter from the book and writes, "make 15 multiple-choice questions from this text, 4 options each, the answers in a separate list at the very end, use no information from outside the book". In 20 minutes she gets 15 questions, keeps 9 of them as they are, polishes 4 a little, and drops 2 because their answers are not in the book. The evening is saved, but the dropping is a job she still has to do herself.
A freelancer in Chattogram. He has to talk in English with foreign clients, and when he writes a hard message about rates, keeping the tone right gets difficult. Now he writes his point in Bangla, then says "write this in professional but friendly English, at most 120 words, no apologetic tone". Before sending the message he reads it once himself, because AI sometimes puts in a promise he did not make, such as delivery on a specific date.
A college student in Dhaka. She could not understand a chapter written in English. She pasted the chapter and wrote, "explain in simple Bangla, then ask me 5 questions so I can tell whether I have really understood". The last part is the real one. After reading the explanation it feels like you understood, but when the questions come, you see where the gaps remain. But she does not copy the text she has to hand in, because that teaches nothing and also gets caught.
A small entrepreneur, on the other hand, builds a website for their business by giving instructions in plain words, which can be done without knowing code using a tool like BanglaCodes. A site published for free works fully and even installs on a phone; only if the messages from the contact form must really reach you do you need lifetime hosting. Knowing this on day one saves surprise later.
Notice that in every example the person takes the final decision. This is the safest way to use AI so far.
Mistakes beginners make
| Mistake | What happens | What to do |
| Giving a one-line instruction | It fills the gaps and gives an average answer | Add context, the reader and the size |
| Treating the first answer as final | The weakest draft gets used | Give corrections like "make it shorter" two or three times |
| Not verifying numbers and dates | A confident mistake gets printed | Check every number and name in the answer separately |
| Ten kinds of jobs in one chat | Old context leaks into new answers | Open a new chat when the subject changes |
| Writing secret information in the chat | The information stays with the company | Replace NID, bank details or a customer list before writing |
| Thinking of it as a calculator | Quiet mistakes in big sums | Do the sums in a spreadsheet, get the explanation from AI |
| Pressing it with "that is wrong" | It apologises and moves to a wrong answer | Say "where did you get that, how sure are you" |
When not to use it
This part is more important than any list of features. In some places the honest answer is that this is the wrong tool.
Where the price of a mistake is paid with the body. Medicine doses, a child's fever, pregnancy problems, changing the medicine for an old illness. AI can explain ideas, but it has not seen your body.
Where you have to take responsibility and sign. A land deed, a reply to a legal notice, tax filing, an audit opinion. It is not only legal risk. The result of a mistake here is carried for years. AI can give a draft, but it cannot take responsibility.
Where today's specific number is everything. Today's dollar rate, today's market price, an exam date, a train timetable. There are official sources for these, and that is what you need.
Where your own voice is the real thing. A condolence message, a letter of apology, words written to a loved one, a wedding speech. Four untidy lines in your own language are better than ten tidy ones, because the person who reads is looking not for language but for you.
Where someone else's secret information is involved. A list of customers' phone numbers, a patient's file, an employee's salary papers. These are not yours, so the right to take the risk is not yours either.
Where learning is the whole point. Exam answers, assignments, the first steps in a new skill. If you copy the answer, the paper will be handed in, but nothing will go into your head.
Questions and answers
Does AI look up answers from the internet? Generally no. The Model builds answers from the patterns it learned in Training. But many AI apps now have a separate search feature, and when it is on, it really does look at the web. If you see links in the answer, search was used, and if there are no links, assume it is speaking from memory.
Does AI understand Bangla well? Much better than before, and enough for everyday work. But there is far more English text on the internet than Bangla, so there is more risk of mistakes in the Bangla language and in facts specific to Bangladesh. On complicated technical subjects, asking in English and then having the answer translated into Bangla gives a better result.
Does AI keep learning from what I say? The Model does not change while you chat. The things you write are used inside that conversation, such as answering while remembering earlier messages. Some apps separately offer a feature that remembers your preferences, and that is not the Model learning, it is note keeping. Remember that what you write may be kept by the company, so do not write an NID number or bank details in a chat.
If the same question gets two different answers, which is right? The mismatch is the real news. Where the two answers agree, that is comparatively safe, and where a number or name changes, it is guessing.
Is what is said about AGI or superintelligence true? It is still an argument, not a sure prediction. Some say it will come within a few years, some say it is far away. Listen carefully to those who state a date firmly. Working out what you can do with today's AI will be more useful.
Do I need a computer to start using AI? No. A browser or an app on your phone is enough. Most of the big AIs have a free tier, and for everyday work that is enough. Do not start by thinking about cost. Use it for two weeks first and see which of your jobs it really helps.
Does AI run inside my phone, or somewhere else? The big Models do not run on the phone. They run on the company's servers. What you write goes there over the internet, and the answer comes back. That is why it does not work without internet, and that is also why your words leave your phone.
Is using AI dishonest in some way? It depends on what the job is. Having it draft a letter or translate is as normal as using a calculator. But where the thing you hand in is counted as your own work, such as an exam paper, find out the rule from the institution.
In short
- AI is software that works not from hand-written rules but by recognising patterns from huge numbers of examples.
- An AI like ChatGPT does not look up answers. It guesses one piece at a time and builds up the whole answer.
- A Model is the sum of billions of numbers stored when Training ends, not a store of facts, so mixing up specific numbers and dates is normal.
- The Model does not change from what you write. It re-reads the whole chat every time to answer, so in a new chat it does not know you.
- AI is excellent for jobs where "roughly right" is enough, and where specific numbers, law or health questions are involved, a person must verify.
- To start you need no computer or technical knowledge. A phone and curiosity are enough.