Humans learn language by hearing and reading it — millions of sentences over years. Large language models learn similarly, except they process far more text, far faster, and with a very specific objective: predict what word comes next.
That's it. The core training task is: given these words, what word is most likely to follow? Do this billions of times with billions of examples, and something remarkable emerges.
LLMs are trained on enormous collections of text: - A significant fraction of the publicly accessible internet (Common Crawl dataset) - Books (Project Gutenberg, licensed books, digitized libraries) - Wikipedia and other encyclopedias - Academic papers - Code repositories (GitHub) - News archives
GPT-4 is estimated to have been trained on roughly 1 trillion words — more text than any human could read in thousands of lifetimes.
During training, the model adjusts billions of numerical values called parameters — think of them as the model's memory of patterns it has seen. GPT-3 had 175 billion parameters. Modern frontier models have trillions.
These parameters encode everything the model 'knows': grammar, facts, reasoning patterns, writing styles, and code syntax — all compressed into numerical relationships.
Training on raw internet text produces a model that can complete text statistically — but that's not the same as a helpful assistant. A second phase called fine-tuning teaches the model to follow instructions, answer helpfully, and refuse harmful requests.
OpenAI's key technique for ChatGPT was RLHF (Reinforcement Learning from Human Feedback): human reviewers rated model responses; those ratings trained the model to produce responses humans preferred. This is why ChatGPT feels like it's 'trying to be helpful' even though it has no intent.
Understanding training helps you predict AI behavior: - Topics well-represented in training data → better AI performance - Recent events after the training cutoff → AI doesn't know - Common tasks in training (essay writing, code, translation) → AI performs well - Rare, specialized, or niche knowledge → higher hallucination risk
What it means to 'train' an AI model — the conceptual picture without any math or code
Humans learn language by hearing and reading it — millions of sentences over years. Large language models learn similarly, except they process far more text, far faster, and with a very specific objective: predict what word comes next.
That's it. The core training task is: given these words, what word is most likely to follow? Do this billions of times with billions of examples, and something remarkable emerges.
LLMs are trained on enormous collections of text: - A significant fraction of the publicly accessible internet (Common Crawl dataset) - Books (Project Gutenberg, licensed books, digitized libraries) - Wikipedia and other encyclopedias - Academic papers - Code repositories (GitHub) - News archives
GPT-4 is estimated to have been trained on roughly 1 trillion words — more text than any human could read in thousands of lifetimes.
During training, the model adjusts billions of numerical values called parameters — think of them as the model's memory of patterns it has seen. GPT-3 had 175 billion parameters. Modern frontier models have trillions.
These parameters encode everything the model 'knows': grammar, facts, reasoning patterns, writing styles, and code syntax — all compressed into numerical relationships.
Training on raw internet text produces a model that can complete text statistically — but that's not the same as a helpful assistant. A second phase called fine-tuning teaches the model to follow instructions, answer helpfully, and refuse harmful requests.
OpenAI's key technique for ChatGPT was RLHF (Reinforcement Learning from Human Feedback): human reviewers rated model responses; those ratings trained the model to produce responses humans preferred. This is why ChatGPT feels like it's 'trying to be helpful' even though it has no intent.
Understanding training helps you predict AI behavior: - Topics well-represented in training data → better AI performance - Recent events after the training cutoff → AI doesn't know - Common tasks in training (essay writing, code, translation) → AI performs well - Rare, specialized, or niche knowledge → higher hallucination risk