LIVE
Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|Architectural tweaks may break conventional scaling law exponents16/09/26|OpenAI publishes a framework for reporting model misalignment16/09/26 · OpenAI|NVIDIA's Vera Rubin NVL72 Debuts in MLPerf Inference v6.116/09/26 · NVIDIA|OpenAI expands ChatGPT advertising with Sponsored Agents16/09/26 · OpenAI|OpenAI moves into advertising with 'Sponsored Agents'16/09/26 · OpenAI|Google DeepMind Introduces Gemini 3.8 Live and Its Extended Thinking Variant15/09/26 · Google DeepMind|What's at stake in AI's trillion-dollar infrastructure bet15/09/26|A Flaw in Chain-of-Thought Safety Monitoring14/09/26|Stellar Colosseum: A Multi-Agent System for Long-Horizon Mathematical Research14/09/26|Apple Code Hints Siri Could Be Swapped for ChatGPT or Claude14/09/26 · Apple|Anthropic says Houthi-linked actors used Claude Code for missile guidance software13/09/26 · Anthropic|Yoshua Bengio examines why AI agents lie, cheat and coordinate13/09/26|
Beginner⚙️

How ChatGPT works in 5 minutes

ChatGPT is just a juiced-up autocomplete that read the Internet. We explain simply, without jargon, how it really works and what it's for.

8 min readPublished May 4, 2026· Updated September 17, 2026

In One Sentence

ChatGPT is a conversational bot powered by an LLM (a "large language model") that guesses word by word what to answer you, like a supercharged autocomplete that has read a good chunk of the Internet.

🎰
The analogy that works
Imagine a mental slot machine that, for each word, calculates thousands of possibilities and picks the most probable one. Except this machine has read billions of pages, and it's incredibly good at guessing what sounds right. That's ChatGPT.

🚀 Want to compare ChatGPT to its competitors?

30 models compared: Claude, Gemini, Mistral, DeepSeek... Filter by price, performance, language.

See the comparison

How it works, in 4 steps

ChatGPT is essentially 4 phases happening at lightning speed:

The cycle of a ChatGPT response

  1. You write a question

    The system receives your text (the 'prompt').

  2. It breaks into tokens

    Each word is transformed into small numerical chunks the machine can understand.

  3. It predicts the next word

    For each position, it calculates the probability of each possible word and chooses.

  4. It generates the response

    Word by word, until forming a complete and coherent answer.

It's as simple as that. No abstract reasoning, no "thinking". Just massive statistical prediction, but done so well it seems intelligent.

So why does it seem so smart?

Because to predict the next word well, you need to understand an enormous amount implicitly.

The brilliant mathematical trick
To guess the word after "The President of France in 2026 is…", you need to know French politics. To complete "Water boils at 100…", you need to know some physics. By being trained to predict the next word on billions of sentences, the model ended up encoding in its parameters knowledge about grammar, facts, code, logic, languages...

It's a bit like someone spending 10 years completing fill-in-the-blank sentences: by the end, without realizing it, they would have learned history, science, languages, cooking.

What are ChatGPT's strengths compared to Google?

ChatGPT vs Google

 🔍Google💬ChatGPT
Verifiable sourcesYes (direct links)Limited
Real-time informationAlways up to dateLimited to training date
Synthesis / SummaryNo, just linksExcellent
ConversationNo contextRemembers your discussion
Hallucination riskNonePresent
Creative tasks (writing, coding)LimitedExcellent

Practically speaking, what's it for?

ChatGPT is a Swiss Army knife for 3 main use cases:

1. Generate text

Emails, articles, LinkedIn posts, cover letters, meeting summaries, reports... ChatGPT excels at writing.

2. Rephrase / Improve

You give it poorly written text, it corrects it. You want your message in a professional, formal, funny, simple style? It adapts.

3. Understand / Synthesize

You paste a long document, it summarizes it. You ask a complex question, it simplifies. You ask for arguments for/against a topic, it structures them.

ChatGPT versions

OpenAI offers several versions, with different capabilities:

Main versions (2026)

 🆓FreeChatGPT Plus (~$20/month)
Model usedRecent GPT but not the bestMost recent models (GPT-5)
Daily limitsRestrictiveVery generous
Image generationLimitedUnlimited
Vision / Code interpreterLimitedComplete
Access to custom GPTsLimitedComplete
Files (PDF, image)LimitedUnlimited

The 5 limitations you absolutely need to know

1. Hallucinations
ChatGPT can invent facts, sources, citations. Worse: it does so confidently. Always verify critical information.
2. No real-time knowledge
Its knowledge stops at its training date (usually several months behind). For news, sports, finance: it doesn't know.
3. Bad at complex maths
It's okay for simple arithmetic, but can mess up slightly advanced calculations. Always verify with a calculator.
4. Training data biases
Since it learned from the Internet, it reproduces the biases : cultural, social, ideological, it found there. Not neutral, despite OpenAI's efforts.
5. Privacy
Everything you write can be used to improve the model (unless you disable this setting). Never paste sensitive info, medical details, or passwords.

How to use ChatGPT effectively

The 4 reflexes to adopt
- Be specific: "Write a formal email to my manager to cancel tomorrow's meeting due to a family emergency" > "Write an email" - Give context: "You are a lawyer specializing in French employment law..." - Iterate: if the answer isn't great, ask "Redo it shorter / simpler / more formal" - Verify: especially for numbers, dates, sources, laws

ChatGPT vs the others: who's the best?

Quick clarification: ChatGPT is not the best LLM on all topics. It's the most well-known, but not necessarily the most performant.

Average score on public benchmarks (MMLU, HumanEval, MATH)

Claude Opus 4.795/100
ChatGPT (GPT-5)92/100
Gemini 3 Pro90/100
DeepSeek V388/100
Mistral Large 384/100

Depending on your use:

  • Code & reasoning: Claude > ChatGPT > Gemini
  • Versatility: ChatGPT and Claude are neck and neck
  • Multimodal (image, video): Gemini is very strong
  • Open source: Mistral, Llama, DeepSeek
  • Native French: Mistral and Claude are excellent

Going further on training

📚How did ChatGPT learn to speak?

ChatGPT's training happens in 3 phases:

1. Pre-training

The model reads billions of pages from the Internet, books, code. It learns to predict the next word. This is the longest step (several months) and the most expensive (hundreds of millions of dollars).

2. Supervised fine-tuning

It's shown human examples of good responses: "If asked X, answer Y". This teaches it to be helpful, not just imitate the Internet.

3. RLHF (Reinforcement Learning from Human Feedback)

Humans compare two responses from the model and say which is better. The model learns to favor preferred responses. This is what makes ChatGPT polite and helpful rather than cynical like certain parts of Reddit.

Conclusion: ChatGPT is not "naturally intelligent". It has been carefully educated to resemble a helpful assistant.

The metaphor that summarizes everything

📚
The amnesiac librarian
Imagine a librarian who has read every book in the world, but has lost memory of where they learned what. When you ask a question, they answer brilliantly, synthesizing their readings. But sometimes they mix up passages and invent things believing they really read them. That's ChatGPT.

Key takeaways

ChatGPT is:

  • ✅ An ultra-powerful assistant for generating, rephrasing, synthesizing text
  • ✅ A huge productivity boost if used well
  • ❌ Not an infallible oracle: it makes mistakes, especially on precise facts
  • ❌ Not a search engine: it doesn't know what's happening in real time
  • ❌ Not confidential by default: don't share your secrets with it

Most important: it's a tool, not a brain. Always verify, always keep your critical thinking.

🧠 Quiz
Question 1 of 3

What does ChatGPT do for each word it generates?

Going further

Tags
ChatGPTLLMVulgarisationOpenAI

Read next