How many parameters does a frontier model typically have?	~200 billion (the chapter’s stat; the range is 100B–1T depending on the model).
How much memory do 200B parameters occupy in BF16?	~400 GB — two bytes per parameter.
What is the fundamental operation of a transformer linear layer?	y = Wx + b, followed by a nonlinearity (GELU, SiLU, or ReLU).
What is self-attention, expressed as a formula?	softmax(QKᵀ/√d)V — a matmul of queries and keys, a softmax, then a matmul with values.
How many transformer blocks does a typical frontier model have?	~80.
What paper introduced the transformer’s self-attention mechanism?	‘Attention Is All You Need’ by Vaswani et al. (2017).
What does BF16 sacrifice compared with FP32, and what does it keep?	It keeps FP32’s exponent range; it sacrifices mantissa precision (uses FP16’s narrower mantissa).
What is FP8 used for, and what did Hopper add to support it?	FP8 is the new precision floor for inference (and increasingly training); Hopper added native FP8 tensor cores.
What is FlashAttention?	A kernel that rearranges the attention computation to be memory-efficient on real hardware, exploiting cache hierarchy to avoid materializing the full QKᵀ matrix in HBM.
What is the vocabulary size and embedding dimension of a typical frontier model?	~100,000 tokens in the vocabulary; each mapped to a ~16,000-dimensional embedding vector.