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 expediter: load the batch to its sweet spot
From: The Expediter →DSpark's last machine, like loading a delivery van. An auto-playing walkthrough: pool every guessed word from every request, sort surest first, and load the batch one at a time. Each word adds a likely delivery but makes the batch slower, so words-per-second climbs to a peak and falls. DSpark stops at the peak, and because survivals only drop, it never needs to keep loading.
The big model checks the guessed words in batches, serving many requests at once. Each word carries an honest chance of passing, from the honesty check.
Do not ship the whole dictionary
From: In the Machine Room →A production trick from DSpark. Training the small guesser needs the big model's answer per word, and the whole answer is a score for every possible word, over a hundred thousand numbers, which floods the link between computers. Auto-playing: it shows the heavy old way, then DSpark sending only the short summary and doing the last step locally, with a size comparison.
To train the small guesser, the big model has to send its answer over to the guesser's computer, for every word it teaches.
A scheduler that cannot wait
From: In the Machine Room →Why the neat sweet-spot rule needs real-world help. Auto-playing: the scheduler assumes GPU speed changes smoothly as the batch grows, but real hardware jumps in steps, a staircase not a ramp, and the pipeline cannot pause to measure it, so DSpark predicts the next batch a step ahead.
Part five picks the batch size by finding where the GPU's speed peaks. To do that, it assumes speed changes smoothly as the batch grows.