Training vs inference: from the internet to a live answer

1. 1The model was frozen months ago, yet the weather answer is right. Tap where you think it came from.
2. 2Training: a huge slice of the internet, most of it thrown away, then guess, miss, nudge, repeated for months on thousands of GPUs. Then the numbers freeze.
3. 3Inference: your question reaches the app, a tool fetches today's weather, the text is pasted next to your question, and the frozen model reads it.

Training and inference are the two halves of every AI product, and this widget plays them one after the other. In the first half you watch a model get made. It starts as a huge slice of the internet: pages, books, code, forums. Most of it is thrown away as duplicates, spam and junk. Then comes the game from the first lesson, in reverse: show a sentence, hide the next word, let the model guess, nudge its numbers when it misses. Repeat trillions of times. When training stops, the numbers are frozen. That is the model that ships.

In the second half you watch a question get answered today. The model alone cannot do it, because its knowledge stops at its training cutoff, months ago. The app notices the question needs fresh facts, runs a tool, and pastes the result next to your question. The frozen model then reads what it was handed and writes the answer. No sensors, no fact database, no weights changing.

The opening question is a real test of intuition: where did the correct weather answer come from? Most people say the model looked it up or somehow knows. Neither is right, and the reveal shows exactly which half of the pipeline did the work.

Almost every confusion about AI products dissolves once these two halves are separate in your head. Training happens once and costs a fortune. Inference happens every time anyone asks anything, and that is where the ongoing bill lives. Retrieval, tools, search and agents all sit on the inference side, bolted onto a frozen model. Fine-tuning and post-training sit on the training side, changing the numbers before they freeze.

Paste this into any HTML page, course platform, wiki or slide tool that accepts an iframe. It runs with no account, sets no cookies of its own, and links back here.

Attribution is built in. If you write about it, a link to this page is all we ask.

Does ChatGPT have a database of facts inside it?

No. There is no fact database inside, only frozen numbers tuned during training. When it cites something current, the app ran a search tool and pasted the results into the conversation for the model to read.

What is the difference between pre-training and post-training?

Pre-training is the next-word game over a huge corpus, which produces a raw guesser. Post-training is a shorter phase where humans write example answers and rank the model's attempts, so the same numbers get tuned to behave like an assistant. Both happen before the freeze.

Can I embed this pipeline widget in my teaching material?

Yes. Copy the iframe snippet on this page into a course page, a wiki or a slide. It needs no account and links back to the full lesson and the training versus inference explainer.