A short glossary for the terms the narrator uses.
- token
- A chunk of text the model reads as one unit. In English one token averages about three quarters of a word.
- embedding
- The long list of numbers that places a token in a space where words with related meanings sit close together.
- attention
- The step where each token weighs every other token in the passage to work out what it means here.
- context window
- Everything the model can see while it writes, your question and the earlier conversation included.
- temperature
- A setting for how much chance goes into picking the next token. Turn it down for predictable answers.
- RLHF
- Reinforcement learning from human feedback. People compare two answers and the model learns to prefer the kind they pick.
- test-time compute
- Extra tokens a model spends thinking before it answers. Harder problems tend to reward more of them.