ChatGPT: the interface that changed everything

OpenAI released ChatGPT on 30 November as a free "research preview." Two weeks later, it has taken over every conversation I have, in research circles and outside them. Sam Altman posted that it crossed a million users in about five days. I can't think of any piece of software that spread this fast.

I've been playing with it a lot, so here is my attempt to explain what it is, why it suddenly feels so different, and where it still falls short.

What is it, in plain words?

At its core, ChatGPT is a language model: a program trained to predict the next word in a piece of text. Give it "The capital of France is" and it predicts "Paris." Do this one word at a time, feeding each prediction back in, and you get whole paragraphs, essays, or code.

That sounds too simple to be useful, but to predict the next word well across millions of documents, a model has to pick up grammar, facts, styles of writing, and even some reasoning patterns. All of that gets stored in the model's parameters, which are billions of numbers adjusted during training.

ChatGPT is built on a model from OpenAI's GPT-3.5 series. So the raw ability isn't new. GPT-3 has been around since 2020, and developers have been building on it through OpenAI's API.

So what changed?

Two things: how the model was tuned, and how it is presented.

1. It was trained to be an assistant

A raw language model is not trying to help you. It just continues text. Ask it a question, and it might answer, or it might write five more questions, because that's what often follows a question on the internet.

To fix this, OpenAI used a method called reinforcement learning from human feedback (RLHF). Their blog post describes three steps:

Show it good examples. Human trainers wrote conversations, playing both the user and the assistant. The model was fine-tuned on these to learn the format of being helpful.
Teach it what "better" means. The model wrote several answers to the same prompt, and trainers ranked them from best to worst. A separate reward model learned to predict those rankings.
Practise against the judge. The language model was then trained with reinforcement learning to produce answers the reward model scores highly.

Think of it like training a new employee: first you show them examples of good work, then you give feedback on their drafts, and they slowly learn your taste. The same team used this approach earlier this year for InstructGPT.

The result is a model that answers the question you asked, admits mistakes when you point them out, pushes back on false assumptions, and declines some harmful requests. None of that is perfect, but it changes the experience completely.

2. It's a chat box

This is the part I find most interesting. Before ChatGPT, using GPT-3 meant an API key, a "playground" with sliders for things like temperature, and carefully engineered prompts. Now it's a text box. You type like you'd text a friend, and it remembers what you said earlier in the conversation, so you can say "make it shorter" or "now explain it to a 10-year-old."

It reminds me of how the web browser didn't invent the internet but made it usable for everyone. The technology was there. The interface let normal people in.

What I've noticed so far

Where it shines:

Explaining concepts. I asked it about the Penrose process and frame dragging, and the explanations were surprisingly clear for a first pass.
Small coding tasks. Little Python scripts and plotting code often run on the first try.
Rewriting and summarising. Emails, abstracts, and making messy notes readable.

Where it worries me:

It is confidently wrong. It can invent references that don't exist or get maths subtly wrong, and it sounds exactly as sure when it's wrong as when it's right. OpenAI says this openly in their own list of limitations.
It depends on phrasing. Ask the same question two ways and you can get two different answers.
It doesn't know recent events. Its training data stops in 2021.
It's already causing problems. Stack Overflow temporarily banned answers generated by ChatGPT a few days ago, because so many plausible but wrong answers were being posted.

My rule for now: treat it like a very well-read friend who sometimes bluffs. Useful for a first draft or explanation, never the final word.

Why I think this matters

For the first time, talking to a computer in plain language feels normal. My guess is that every big tech company will want its own version within months, and new kinds of work will appear around checking, evaluating, and building on top of these models.

I don't think we go back from here.