Visualizers
Learn by moving things
Interactive explainers for the ideas behind retrieval, embeddings, and agents. A formula tells you what is true; dragging a vector shows you why. Each one also appears inside the article it belongs to, and runs entirely in your browser, so drag, poke, and break it.
The draft-and-verify cycle
From: Draft, Then Verify →One cycle of speculative decoding. A small draft model guesses the next tokens in one pass; the large target model verifies them together, accepts the correct prefix, and corrects the first miss. Step through it and watch one target pass commit several tokens.
So far, the model has written the text above. Now it needs the next tokens.
Draft better: the previous word fixes the crossbreed
From: Draft Better, Verify Smarter →DSpark's first idea, interactive. Two replies fit after 'Thanks!': 'of course' and 'no problem'. Flip a blind parallel drafter (reaches for the same word and crossbreeds into 'of problem') against DSpark's sequential head, which lets position two see position one and adds a small bias so the words always agree, at the same speed.
Verify smarter: stop when the number drops
From: Draft Better, Verify Smarter →DSpark's second idea, interactive. Each drafted word carries a confidence of surviving the check; multiply them into a running survival that only falls. Change the server load to move the 'worth a check' line and watch the cut move: heavier load raises the bar and stops sooner, so the target never spends its scarce capacity on words that were going to be rejected.