Welcome to the LLM Roomba Analogy!

This demo shows how different LLM strategies work by analogy: a Roomba navigating a room with obstacles, carpet (slow zones), and a target to reach.

How to play (Human mode):

Explore — Drag away from the Roomba and release to send a ghost scouting the room
Aim & Shoot — After exploring, drag again to aim the Roomba and release to launch

AI Modes to compare against:

Standard LLM — fires once and hopes for the best
Reasoning (CoT) — thinks first (ghost explores), then takes a calculated shot
Agentic — moves, pauses to re-aim, and course-corrects iteratively

Battery = token budget. Carpet slows you down. Fog hides the room until explored. Can you beat the AI?

LLM as Roomba: An Analogy

Mode

Battery (Tokens)

100%

Status: Idle

Phase: -

Scoreboard

You 0/0

Standard 0/0

Reasoning 0/0

Agent 0/0

Legend

Roomba

Target

Wall / Obstacle

Mapped

Carpet (slow)

Furniture

The Analogy: LLMs are like Roombas bumping around a room. Standard LLMs take one shot and hope. Reasoning models "think" first (ghost explores), then execute a calculated path. Agents periodically check progress and adjust course. Can you beat the AI?