How large language models like GPT work
GPT does not search a database for answers. It writes them a small piece at a time from patterns learned in training. Once you see that, its strong points, its slips and the best way to ask all make sense.

When you type a question into our chat, GPT does not search a library of stored answers. It writes a new reply, a small piece at a time, based on patterns it absorbed from a huge amount of text during training. That one idea explains a lot: why it is so good with language, why it can explain almost anything, why it sometimes states false things with total confidence, and why the way you ask matters. This guide covers the main parts in plain words, no maths needed, with what each one means for your own chats.
The letters in GPT
GPT stands for generative pre-trained transformer. Generative means it produces new text rather than picking from fixed replies. Pre-trained means it first learned from a very large collection of text before being tuned for conversation. Transformer is the type of neural network inside, a design published by Google researchers in 2017 that is particularly good at relating each word to all the others around it. OpenAI released its first GPT model in 2018; the models in our chat, GPT-6 Luna and GPT-6.1 Terra, are recent members of the same line. The history of GPT page has the full timeline, and what is GPT gives the wider picture.
Tokens: how a model reads
A model does not see letters or whole words. Text is first chopped into tokens. A frequent word usually becomes one token; a long or unusual word gets broken into a few fragments. OpenAI's rule of thumb for English is that a token is roughly three quarters of a word. Other languages often need more tokens for the same meaning.
Tokens explain a few oddities. Models can stumble when counting the letters in a word, because they never see the letters one by one. Limits on input and output are measured in tokens, not characters. And prices for developers, as listed on our GPT API page, are quoted per million tokens.
Prediction, one piece at a time
At its core, the model does one thing: given all the text so far, it estimates which token is likely to come next. It picks one, adds it to the text, and repeats, hundreds or thousands of times, until the answer is complete. That is why replies appear gradually in the chat rather than all at once.
A small amount of randomness in the choice keeps answers from being identical every time. That is what the regenerate button uses: ask again and you get a different route through the same knowledge. Our chat keeps earlier attempts as versions, so you can compare them.
Prediction sounds simple, but to predict well across billions of examples a model has to pick up grammar, facts, styles, reasoning patterns and much more. What looks like understanding emerges from that pressure to predict.
How a model is trained
GPT models are made in stages:
- Pre-training. It processes a vast collection of text, practising a single task: guessing the next token. At this stage it can continue any text but is not yet a good conversation partner.
- Instruction tuning. It is trained on examples of requests and good responses, so it learns to follow instructions and answer questions.
- Feedback and reinforcement learning. People and automated checks rate answers, and the model is adjusted toward responses that are helpful, accurate and safe. Reasoning models are also trained this way to work through problems before answering.
Training ends at some point, and that date is the model's knowledge cut-off. According to OpenAI, GPT-6 Luna's data runs to 18 May 2026 and GPT-6.1 Terra's to 1 June 2026. Nothing after that is known to the model unless you paste it in, and our chat does not browse the web.
The context window: what the model sees
A model has no memory between conversations. Within one chat, it sees a context: the instructions, the messages so far and your newest message, all as tokens. OpenAI gives Luna and Terra a context window of about 1.05 million tokens, enough for several long books.
A free chat like ours sends a much smaller slice: the recent part of the conversation, so answers stay fast. In a very long chat, early messages can fall out of view. If a detail matters, repeat it, or start a new chat with a short summary of where you are. Each chat is separate; the model does not remember what you said in another one.
Reasoning: thinking before answering
Reasoning models generate a stretch of hidden working before the visible reply. That gives them a chance to break a problem into steps and catch their own slips, at the cost of a little time. In our chat, the Think chip switches to GPT-6.1 Terra, which reasons briefly; GPT-6 Luna, the default, answers almost at once. Our guide to reasoning models explains when the pause is worth it.
Where made-up answers come from
Since the model always writes a plausible continuation, it writes one even when the facts are missing. A request for a source on a niche topic gets a reference-shaped answer, whether or not such a source exists. A question about a small town gets details that fit towns in general. The model has no reliable internal signal that says "I don't know this", although newer models are better at admitting uncertainty. That is why checking matters for names, numbers, dates and quotes; our fact-checking guide has a quick routine.
What this means for your prompts
Everything above turns into a few practical rules:
Weak: What does the contract say about ending early? Strong: Here is clause 9 of my rental contract: [paste clause]. In plain English, explain when I can end the contract early, what notice I must give and what it may cost me. Quote the sentence each point is based on.
- Paste the material. The model works far better on text in its context than on what it half-remembers.
- Be specific. Every detail you add narrows the predictions toward what you need.
- Ask for quotes or sources from the text you gave, so you can check each point.
- Start fresh when a long chat has drifted; a short summary in a new chat often works better. The summarizer can write that summary for you.
Want to see it in action? Ask GPT to explain itself in five steps, then ask follow-up questions about the parts that are unclear.
FAQ
Does GPT understand what it writes?
It depends on what you mean by understanding. It clearly captures relationships between ideas well enough to explain, summarise and reason about them, but it has no experience of the world and no reliable way to tell a fact from a plausible guess.
Does it learn from my chats?
Not during the conversation, and not between conversations in our chat. The model's knowledge is fixed at training time. OpenAI states that data sent through its API is not used for training unless the customer opts in.
Why do I get a different answer when I ask again?
The model chooses each next token with a little randomness, so the same question can lead to different wording or structure. For facts, a different answer is a hint to check.
Is GPT the same as ChatGPT?
No. GPT is the family of models made by OpenAI; ChatGPT is OpenAI's own app built on them. gpt.me is an independent site that uses GPT models through an API; our comparison with ChatGPT lists the differences.
Got a question? Ask it now
No account and no download. Open the chat, type what is on your mind, and the answer starts appearing in a second or two.
Start chatting

