Marimo

Monte Carlo option pricing accelerated on the GPU. Adjust strike, volatility, and path count, watch the pricing surface converge in real time.

Mean-variance portfolio optimization with GPU-accelerated matrix ops. Drag risk tolerance and watch the efficient frontier redraw live.

Parametric 3D forms generated from reactive sliders — rotate, warp, and re-seed the geometry without touching a render pipeline.

Sketch with layered 2D Gaussians — position, covariance, and color all become live, reactive parameters instead of pixels.

An interactive explainer for the Gray-Scott PDE — the same reaction-diffusion system behind the Field Notes plotter pieces, stepped live in the browser.

Binary search traced step by step on a live number line. Every comparison is drawn as it happens, no pseudocode required.

A multi-armed bandit rendered as a live decision canvas — pull arms, watch reward estimates and confidence bounds update in place.

Epsilon-greedy, UCB, and Thompson sampling racing side by side on the same bandit problem — regret curves update as they run.

A random walk through a generated labyrinth, with each step's transition probabilities laid out directly on the maze.

Sketch cities on an Excalidraw-style whiteboard and watch nearest-neighbor and 2-opt tours solve the traveling salesman problem live.

A nearest-neighbor traveling-salesman tour over twelve random cities, scrubbed step by step with a slider — path-so-far and distance stats update live.

Auto-animated superposition of N sine waves with a live phase-vector circle plot — marimo's reactive state model doing the animation loop.

An auto-playing Gaussian PDF/CDF viewer where mean and sigma trace parametric paths — circles, spirals, Lissajous — with a live parameter-space trajectory.

A 2D animated path navigator — seven path shapes with play, reset, and manual override — whose position drives a live Gaussian PDF and parameter trace.

A reactive Bezier curve editor — drag control points and every downstream cell (arc length, curvature, tangents) recomputes instantly.

A minimal demo of marimo's reactive dataflow driving a live animation loop — change one cell, watch the whole graph re-run downstream.

How a single composite benchmark score can hide capability gaps — simulate aggregate vs per-difficulty-band scoring on an LLM-routing benchmark.

Perplexity-difference detection between finetuned and reference models, simulated live — why the metric is simple but noisy.

Slider-driven comparison of naive full-context replay against a fixed context-recycling budget — accuracy and cumulative token cost over a long conversation.

Does test-time-training's perplexity gain reflect genuine behavioral memory? Probe the claim with a support-length slider.

The reasoning-token budget trade-off behind Kimi K1.5's long2short distillation — optimal context budget vs answer quality, simulated.

Expected total generation cost per successful sample under verifier-guided backtracking versus plain best-of-N sampling.

A rule-based deterministic router against a learned router across synthetic workloads — local vs hosted cost trade-offs, visualized.

An intervention spectrum for probing whether roleplaying LLM personas hold genuine internal beliefs or surface mimicry.

How naive grammar constraints distort model intent during constrained decoding — and how adaptive tracking avoids it.