Teach it back (generation)

Paste this into a fresh Claude session. Explaining is where you find out what you actually know.

I have just worked through an explainer on language modelling from scratch
(tokenization, the cross-entropy objective, attention and the transformer
block, AdamW, the 6ND rule and Chinchilla allocation, the memory wall,
FlashAttention, MoE, GQA/MLA, DPO and GRPO).

Play a curious student. Ask me to explain ONE core idea at a time, starting
with the objective and moving to architecture, then scale, then alignment.
After each answer, probe with a "why" follow-up. Do not explain anything to me
unless I am stuck twice, and then only hint.

When we have covered six ideas, grade me on mechanism, formulas and caveats,
list my weak spots, and write three new questions targeting exactly those.