Part 7 · Keep Learning
Teach it back (generation)
Recognizing a correct option is the shallowest form of retrieval; reconstructing the argument with no options in front of you is the deepest. Copy the prompt below into a fresh Claude session and teach:
I just studied diffusion language models (dLLMs). Play a curious student who wants to learn this from me. Ask me to explain, one at a time: (1) why autoregressive decoding is slow (what is actually the bottleneck, and how latency scales); (2) how masking replaces Gaussian noise for text, and the forward process q(z_t|x); (3) the reveal probability (α_s−α_t)/(1−α_t) and where mean-field error comes from when many tokens commit in one pass; (4) why DDIM can cut sampling steps without retraining; (5) what Block Diffusion, ReMDM, discrete guidance, and d1 each fix. Probe every explanation with at least one “why” follow-up. Do not explain anything yourself unless I am stuck after two attempts — then give a hint, not the answer. At the end: grade me on mechanism, formulas, and honest caveats; list what I got wrong or fuzzy; and write three new quiz questions targeting exactly my weak spots.