Learn
A short list on purpose. Every idea here exists because a symptom on Why needed it to make sense — this is not a glossary.
Context rot — why long inputs are used worse than short ones
Everything inside the context window is available to the model, but it is not used evenly. Published measurements show accuracy falling when the relevant part sits in the middle of a long input, and falling again simply because the input is longer — even when the model can retrieve every relevant fact perfectly. Vendors now call it context rot.
Explains 3 symptoms on this site.
Knowledge cutoff — the date a model stops learning
A knowledge cutoff is the date after which nothing entered the model. It is published by the vendor, it differs per model rather than per brand, and it is softer than a date suggests: Anthropic gives two cutoffs per model, and Google describes one model as reliable to March 2026 in some domains and only to January 2025 in others.
Explains 2 symptoms on this site.
Next-token prediction — how an answer is actually produced
A model does not compose an answer and then write it down. At each position it produces a probability for every possible next piece of text, something draws one, and the whole thing repeats with that piece added. Fluency, variability and confident invention are all consequences of that loop — not separate features.
Explains 2 symptoms on this site.
The context window — the fixed size of one request
A context window is the total size of a single request, counted in tokens: the system prompt, the conversation, your documents and the reply the model is about to write, all in one fixed space. It is not memory and it is not storage. Nothing carries over between requests except the text the application chooses to resend.
Explains 4 symptoms on this site.
The system prompt — the instructions you never wrote
Every chat app puts its own instructions in front of your conversation: how to behave, what it is, what the date is. Anthropic publishes these prompts by model and date, so the block can be measured rather than guessed — and states that they do not apply to its API. Behaviour that differs between the app and the API usually differs here.
Explains 2 symptoms on this site.
Tokenization — what a model actually reads
A tokenizer cuts text into the pieces a model works with, and hands the model a number for each piece. The cut is learned from data rather than from grammar: common words survive whole, rare ones break apart, and the same text splits differently under different dictionaries.
Explains 3 symptoms on this site.