Step by Step
T
Token — the unit of text an LLM actually processes
Roughly ¾ of a word on average. LLMs don't read whole words or letters — they break text into tokens and process those.
Example: the word "unbelievable" might be split into multiple tokens like "un," "believ," and "able" depending on the tokenizer.
T
Temperature — controls randomness in output
A setting from 0.0 to 1.0+ that controls how predictable vs. creative the model's word choices are. Lower temperature = more deterministic and repeatable; higher temperature = more varied and surprising.
Example: temperature 0.0 for a customer support bot (consistent, safe answers) vs. temperature 0.9 for creative fiction writing (varied, surprising phrasing).
C
Context window — how much the model can "see" at once
The maximum number of tokens the model can consider in a single conversation, including everything you've typed and everything it's generated so far.
Example: a 200K token context window lets you paste in an entire book and ask questions about it, while a small context window would force you to summarize instead.
H
Hallucination — confidently stated false information
A fundamental limitation, not a rare bug: LLMs predict statistically plausible next tokens, not verified facts, so they can produce fluent, confident-sounding statements that are simply untrue.
Example: an LLM might cite a specific academic paper with a plausible-sounding title and author that doesn't actually exist — stated with the same fluent confidence as a true fact.
Applied Walkthrough
1
A user asks an LLM a very specific factual question about a minor historical event, and the model responds with total confidence.
2
The answer sounds fluent and specific — dates, names, details — but is entirely fabricated.
3
This is hallucination: the model is predicting plausible-sounding next tokens, not retrieving verified facts from a database.
4
Because hallucination is a structural feature of how LLMs generate text (not a bug that gets patched out), the correct response is always to verify important facts independently rather than trust fluency as a proxy for accuracy.
Exam Application
Exams test whether you understand these terms are properties of how the model works, not optional add-ons. A question describing a chatbot that gives wildly different answers to the same prompt each time is testing your understanding of temperature. A question about a model losing track of earlier parts of a long conversation is testing context window limits. A question about confidently stated wrong facts is testing hallucination.
⚠ Common Trap
The most common trap is assuming hallucination only happens with obscure or unusual questions. In reality, LLMs can hallucinate confidently on any topic, including ones that sound like common knowledge — fluency and confidence in the model's tone are not reliable signals of factual accuracy. Never treat a fluent answer as inherently a correct one.
✓ Quick Self-Check
1. Roughly how much text does one token represent?
Roughly ¾ of a word on average — tokens are sub-word units, not whole words.
Tap to reveal / hide
2. What does a temperature of 0.0 produce compared to a temperature of 1.0?
Temperature 0.0 produces deterministic, predictable output; temperature 1.0+ produces more varied, creative, random output.
Tap to reveal / hide
3. What happens if a conversation exceeds the model's context window?
The model can no longer "see" the earliest parts of the conversation — it effectively loses track of information beyond that window.
Tap to reveal / hide
4. Is hallucination a rare software bug or a fundamental limitation?
A fundamental limitation — LLMs generate statistically plausible text, not verified facts, so confident false statements are structurally possible at any time.
Tap to reveal / hide
5. Why can't fluent, confident phrasing be trusted as a sign of factual accuracy?
Because LLMs are optimized to produce plausible-sounding next tokens, and fluency comes from the same process whether the underlying claim is true or hallucinated.
Tap to reveal / hide