Project
Thinking Canvas
Particle-visualization of an LLM's reasoning process — render the thinking, not just the answer.
A small experiment in making LLM reasoning legible. Most chat interfaces collapse the model’s intermediate steps into “thinking…” spinners; Thinking Canvas turns the reasoning into a visual artifact — particles that emerge, branch, converge, and resolve as the model works through the problem.
Mechanically, it’s honest about what it can and can’t show. While the main model streams its reply (extended thinking on), a companion Haiku call runs in parallel, extracting three concrete visual concepts from the conversation; each resolves through a tiered pipeline — curated animations, icon sets, emoji silhouettes, an embedding match, or an SVG generated on the fly when nothing fits — and ~8,000 physics-simulated particles flow into each shape in turn. So the canvas renders a visual riff on what’s being discussed, paced by the model’s work — not a window into its internals, and it shouldn’t be read as one.
Even that limited signal changes the waiting more than I expected: there’s real anticipation in seeing what the companion model reaches for, and the associations aren’t always what you’d predict. It started as “can I make waiting beautiful?” and ended as a small argument that thinking-time UI is an underexplored product surface.