Everything about ChatGPT — from what it is and how it was made, to how to use it today, the best prompts for every situation, and how to integrate it into your work. Written for beginners, practitioners, and experts. Official sources only.
ChatGPT is a computer program you can have a conversation with. You type something — a question, a request, an idea — and it replies in natural, human-sounding language. It can write, explain, summarise, translate, answer questions, help with decisions, and much more.
It was made by a company called OpenAI, based in San Francisco. It launched on 30 November 2022 and became the fastest-growing consumer application in history — one million people signed up in the first five days.
You can use it for free at chat.openai.com or on the ChatGPT app on your phone. You do not need to download anything for the web version.
ChatGPT is like having a very knowledgeable friend who is available 24 hours a day, never judges your questions, and can help with almost anything involving words and ideas. Ask it anything. The worst that can happen is it gives you an imperfect answer — and you can ask it to try again.
ChatGPT was made by OpenAI. OpenAI was founded in December 2015 by a group that included Elon Musk (who later left the board), Sam Altman (who became CEO), Greg Brockman, Ilya Sutskever, and others. It started as a non-profit research organisation with the mission to ensure that artificial intelligence benefits all of humanity.
In 2019, OpenAI restructured into a “capped-profit” company and accepted a $1 billion investment from Microsoft. Microsoft has since invested a total of $13 billion and integrated ChatGPT technology into its products including Bing, Word, Excel, Outlook, and Windows.
OpenAI’s headquarters are in San Francisco. As of early 2026, the company employs over 1,000 people and is valued at over $80 billion.
The technology that made ChatGPT possible was published in 2017 — a research paper from Google called “Attention Is All You Need.” It described a new kind of neural network called the transformer. OpenAI’s researchers read this paper and began building on it immediately.
In June 2018, OpenAI released GPT-1 — Generative Pre-trained Transformer, version 1. It had 117 million parameters (think of parameters as the “settings” that determine how the AI behaves) and was trained on a dataset of books. It could generate coherent paragraphs of text, which was impressive at the time but limited by today’s standards.
GPT-2 (February 2019) had 1.5 billion parameters — more than ten times larger than GPT-1. OpenAI was so concerned about potential misuse — fake news, spam, automated propaganda — that they initially refused to release the full model. This was controversial. Many researchers thought OpenAI was being overly cautious. Eventually, the full model was released in November 2019 and the predicted harms did not materialise at the predicted scale.
GPT-3 (May 2020) had 175 billion parameters. OpenAI released it through a limited API — you had to apply for access. The outputs were startling. Developers, writers, and researchers who got access shared examples of GPT-3 writing convincing news articles, answering complex questions, generating code, and composing poetry. For the first time, many people genuinely could not tell whether a piece of text was written by a human or a machine.
But GPT-3 was still a developer tool. It had no friendly interface. Most people never touched it.
The missing ingredient was alignment — making the AI actually helpful rather than just technically capable. In early 2022, OpenAI published InstructGPT, which used a technique called Reinforcement Learning from Human Feedback (RLHF). Human trainers rated the AI’s responses, and those ratings were used to teach the model to give better, more helpful, less harmful answers.
InstructGPT, despite being smaller than GPT-3, was preferred by users in testing. The lesson: size matters less than alignment.
ChatGPT was built on the InstructGPT foundation and wrapped in a simple chat interface. Anyone could use it. No application required. Free to start.
The reception was extraordinary. Within 48 hours, the servers were overwhelmed. Within a week, OpenAI had to implement waitlists. Within two months, 100 million users had signed up — a milestone that took Instagram 2.5 years and TikTok nine months.
Technology journalists called it a “ChatGPT moment” — a reference to the iPhone moment of 2007. Something had fundamentally changed.
GPT-4 launched on 14 March 2023. Unlike its predecessors, GPT-4 could understand images as well as text — you could show it a photo and ask questions about it. Its performance on standardised tests was remarkable: it scored in the 90th percentile on the US Bar Exam (GPT-3.5 had scored around the 10th percentile), the 88th percentile on the LSAT, and 99th percentile on many sections of the GRE.
GPT-4 was available in ChatGPT Plus (the paid tier) and through the API.
GPT-4o (the “o” stands for “omni”) was released in May 2024. It was designed to be faster and more efficient than GPT-4, and it could natively handle text, images, and audio in a single model rather than using separate systems. The demo showing real-time voice conversation — including the ability to pick up on tone of voice and respond with appropriate emotion — was widely shared and described as the closest thing to the fictional AI “JARVIS” from Iron Man.
GPT-4o became available to free users, making the most capable model accessible without a subscription for the first time.
OpenAI continued to release improvements including GPT-4o mini (a faster, cheaper version), o1 and o3 reasoning models (designed for complex multi-step problems), and deeper integrations with the operating system, browser, and third-party tools. ChatGPT grew to over 200 million weekly active users as of late 2024, according to OpenAI’s public statements.
“My daughter asked me to write a speech for her school play. I had no idea where to start. I typed into ChatGPT: ‘Write a 2-minute speech for a 10-year-old to deliver at a school play about kindness. Make it warm and easy to memorise.’ It gave me three different versions. We picked the one she liked, changed her name to hers, and she learned it in one evening. I was amazed.”
Here are the most common things people use ChatGPT for, with a real example of each:
ChatGPT is excellent at writing first drafts. Emails to colleagues, complaints to companies, cover letters for jobs, birthday messages, school essays, social media posts, thank you notes, blog articles. Give it context — who it’s for, what you want to say, how it should sound — and it writes.
When you search Google, you get a list of links. When you ask ChatGPT, you get an explanation. For questions like “what does this medical term mean?” or “how does my mortgage work?” or “why is my electricity bill so high?” — ChatGPT explains in plain language, at whatever level you ask for.
Paste in a long document — a contract, a report, a news article, a product manual — and ask ChatGPT to summarise it, pull out the key points, or explain what you need to know. This is one of its most practically useful features.
ChatGPT is like a patient tutor who knows every subject. You can ask it to explain something you don’t understand, ask follow-up questions, ask it to simplify, or ask it to give you examples. It never makes you feel stupid for asking basic questions.
Ask ChatGPT to plan a holiday, organise a week of meals, create a study schedule, write a to-do list for a project, or plan a birthday party. It is excellent at taking a vague goal and turning it into a structured plan.
ChatGPT can translate between most major languages — and not just word for word, but naturally, in the right tone. Useful for emails, signs you photographed on holiday, documents, or just understanding something in another language.
ChatGPT is not perfect. Knowing its limitations will save you from problems:
Treat ChatGPT like a very capable colleague who sometimes makes mistakes. Their ideas are always worth hearing. Their facts always need checking on anything important. Their drafts are excellent starting points. They should never be the final authority on anything critical.
For most everyday users, the free tier is sufficient. You can do a lot before hitting usage limits. If you find yourself using it heavily every day and hitting the limits, Plus is worth considering.
Source: openai.com/chatgpt/pricing — April 2026
ChatGPT is only as good as the instructions you give it. The biggest mistake new users make is asking vague questions and being disappointed by vague answers. The solution is simple: be specific.
Role: Who is ChatGPT in this conversation?
Task: What exactly do you want it to produce?
Context: What does it need to know to do this well?
Format: How do you want the output structured?
Constraints: Length, tone, what to avoid.
“Write me an email about a meeting.”
“Write a professional email to my team of 8 people announcing a mandatory all-hands meeting on Friday 14 March at 10am. The purpose is to discuss the Q1 results and our priorities for Q2. Tone: positive and forward-looking. Max 150 words. Include a request to confirm attendance.”
For complex problems, ask ChatGPT to show its reasoning step by step before giving an answer. This produces more accurate results and lets you follow the logic.
Telling ChatGPT to act as a specific expert produces more targeted, specialised responses.
Do not accept the first output if it is not quite right. Ask ChatGPT to revise with specific instructions. This is how you get to excellent output.
Give ChatGPT examples of what you want, and it will follow the pattern precisely.
ChatGPT is not just a website. It is embedded in an expanding ecosystem of tools:
ChatGPT is based on the GPT (Generative Pre-trained Transformer) family of models. The architecture is a decoder-only transformer — in contrast to the encoder-decoder architecture of the original “Attention Is All You Need” paper, GPT models use only the decoder stack, which is optimised for next-token generation rather than sequence-to-sequence tasks.
GPT models are trained using self-supervised learning on large text corpora. The training objective is causal language modelling: given a sequence of tokens t₁, t₂, ..., t_{n-1}, predict t_n. The loss function is cross-entropy over the vocabulary. Pre-training data for GPT-3 included: CommonCrawl (filtered), WebText2, Books1, Books2, and English Wikipedia — totalling approximately 570GB of text after filtering.
Pre-training for frontier models (GPT-4 and beyond) uses data that OpenAI has not fully disclosed, citing competitive sensitivity. The GPT-4 technical report notes that they “will not specify the architecture (e.g. model size), hardware, training compute, dataset construction, training method, or similar.”
OpenAI (2023). “GPT-4 Technical Report.” arxiv.org/abs/2303.08774
The alignment methodology that distinguishes ChatGPT from a raw language model is RLHF, described in detail in the InstructGPT paper. The process involves three stages:
Ouyang, L., Wu, J., Jiang, X., Almeida, D., Wainwright, C.L., Mishkin, P., Zhang, C., Agarwal, S., Slama, K., Ray, A., Schulman, J., Hilton, J., Kelton, F., Miller, L., Simens, M., Askell, A., Welinder, P., Christiano, P., Leike, J., & Lowe, R. (2022). “Training language models to follow instructions with human feedback.” arxiv.org/abs/2203.02155
GPT-4 introduced vision capabilities — the model accepts both image and text tokens as input. The vision encoder used in GPT-4 is based on the CLIP (Contrastive Language-Image Pre-Training) approach pioneered by OpenAI in 2021, which trains visual and language representations jointly through contrastive learning.
GPT-4’s performance on standardised benchmarks represents a qualitative shift: 90th percentile on the Uniform Bar Exam (UBE), 99th percentile on the GRE Verbal, 88th percentile on LSAT. The GPT-4 technical report provides extensive benchmark results across academic and professional examinations.
GPT-4o (May 2024) represents a departure from the modular approach of earlier multimodal systems (which used separate encoders for different modalities). GPT-4o is a single end-to-end model trained across text, vision, and audio simultaneously, allowing native audio understanding and generation without the latency and quality loss of a pipeline approach. The model achieves near-human response times in voice conversation (average of 320ms) and can interpret audio tone, laughter, and emotional context.
OpenAI (2024). “Hello GPT-4o.” OpenAI Blog. openai.com/index/hello-gpt-4o
The o1 series (September 2024) introduced a new paradigm: models that spend more inference-time compute “thinking” before producing a response. Rather than generating tokens immediately, o1 produces an internal “chain of thought” before the final answer — similar to how a human might work through a hard problem on scratch paper. This approach dramatically improves performance on mathematical reasoning, competitive coding, and scientific tasks, at the cost of higher latency and compute cost per query.
o1 achieved 83% on AIME 2024 (vs 13% for GPT-4o), 89th percentile on Codeforces, and PhD-level performance on GPQA (Graduate-Level Google-Proof Q&A).
OpenAI (2024). “Learning to Reason with LLMs.” OpenAI Blog. openai.com/index/learning-to-reason-with-llms
OpenAI’s API provides programmatic access to GPT-4o, GPT-4o-mini, o1, o3-mini, and other models. The API uses a REST interface with JSON payloads.
from openai import OpenAI
client = OpenAI(api_key="your-api-key")
response = client.chat.completions.create(
model="gpt-4o",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Explain transformer architecture simply."}
],
temperature=0.7,
max_tokens=500
)
print(response.choices[0].message.content)
Key parameters: temperature (0–2, controls randomness — lower is more deterministic), max_tokens (limits response length), top_p (nucleus sampling, alternative to temperature).
platform.openai.com/docs/overview — OpenAI’s complete developer documentation, including API reference, rate limits, pricing, and model specifications.
OpenAI’s safety work encompasses: RLHF alignment (as above), usage policies enforced through the moderation API, red-teaming before model releases, and ongoing superalignment research (using AI to assist in evaluating AI outputs at superhuman capability levels). The Preparedness Framework (published November 2023) describes OpenAI’s approach to evaluating catastrophic risk from frontier models across four categories: cybersecurity, CBRN (chemical, biological, radiological, nuclear), persuasion/manipulation, and model autonomy.
OpenAI (2023). “OpenAI’s Preparedness Framework (Beta).” openai.com/safety/preparedness