Lesson Content

Language models in plain language.

A language model learns by reading an enormous amount of text – books, articles, websites, code, conversations – and learning the patterns in how language is used. At its core, it learns to predict: given this sequence of words, what words are most likely to come next?

When you type a question or request, the model generates a response by making a very large number of these predictions in sequence – each predicted word informed by the context of everything before it. The result is text that follows the patterns of good human writing on similar topics.

This is why AI can write persuasive essays, debug code, explain complex topics, and have coherent conversations: these are all pattern-completion tasks on domains well-represented in training data.

Training data shapes capability and limits.

Everything the model knows comes from training data. This creates two important constraints:

Coverage: AI is better at tasks well-represented in its training data and worse at tasks that are rare or absent. Standard English prose, common programming languages, widely-documented topics – these are well-covered. Highly specialized domains, minority languages, emerging fields, proprietary internal knowledge – these may be poorly covered or absent.

Cutoff: Training data has an end date. The model does not know about events, publications, products, or changes that occurred after its training was completed. For most large commercial models, this cutoff is roughly 12-18 months before current. Check the specific model's knowledge cutoff for accuracy-critical queries.

AI is not a search engine.

This distinction bears repeating at a technical level: a search engine retrieves existing documents from an index. A language model generates new text based on learned patterns. This means AI can produce entirely new content that does not exist anywhere in its training data – and it means AI can confidently state things that are not true, because it is generating plausible text, not retrieving verified facts.

Context is the model's working memory.

A language model processes each conversation by reading the full conversation from the beginning. Everything you wrote is available as context. The model has no memory between separate conversations – each new conversation starts fresh. This is why re-establishing context matters for recurring work, and why Projects and Custom Instructions (which are loaded into every conversation) help.