Numbers worth knowing by heart
The figures and rules of thumb you want in your head the morning of an interview, and to hand the rest of the time. Arithmetic you can apply rather than vendor trivia that will be wrong in six months — and each card says when it was last checked.
Transformer memory arithmetic
20 rowsWhat a model costs in bytes: weights, KV cache, activations and optimiser state, and how to tell in advance whether it fits on the card you have.
Training hyperparameters worth knowing
19 rowsStarting values that are right often enough to be defaults, and what each one interacts with when you change it.
Latency and cost budgets
16 rowsWhat a person notices, what an LLM call actually costs in time, and the arithmetic that turns hardware price and throughput into a price per million tokens.
GPU specifications
15 rowsMemory, bandwidth and rough compute for the cards people actually rent, and why bandwidth is the number that decides single-stream generation speed.
Metrics: which one, and when
20 rowsThe classification and regression metrics, what each one hides, and how to choose between them from the cost of being wrong rather than from habit.
Evaluation set sizes and statistics
21 rowsHow many cases you need before a difference is real, what a percentage point means at each size, and when to believe a judge.
Context limits and token arithmetic
21 rowsHow to estimate a token count from something you can see, what fits in each common window, and how the window behaves as it fills.