task-discretization

Shows a continuous task trajectory being segmented by a discretization mapping f into an ordered sequence of atomic subtasks (τ1..τn). Each τ box has an explicit input/output interface and a measurable completion signal (metric bar). As subtasks complete in order, a reconstructed path in the bottom panel grows to approximate the original continuous task, illustrating composability and observability for learning.

canvasclick to interact

t=0s

practical uses

01.Designing LLM agent workflows as verifiable steps with clear inputs/outputs
02.Creating reward/metric signals for training or evaluation (RLHF/RLAIF, automated graders)
03.Breaking down continuous user goals into testable, automatable pipeline stages

technical notes

Pure Canvas2D; responsive scaling via scale=min(w,h)/240; blocky grid snapping at 4px*scale; animation cycles ~3.8s with discrete step indexing plus eased within-step progress; no external dependencies.

← deep-learning softmax-and-logits →