AI Cleans the House | Visual Coding Comparison

Video Description

Six AI models were given identical prompts to produce the 3D animations using Three.js. Here are the complete test details for those interested. # Models Tested: 1. GPT-5 (OpenAI) 2. Claude Sonnet 4 (Anthropic) 3. Gemini 2.5 Pro (Google) 4. Grok 4 (Xai) 5. Qwen3 Coder 480B A35B (Alibaba Cloud / Qwen Team) 6. DeepSeek V3.1 (DeepSeek AI) The animations shown in the task comparisons were created completely by each AI model with no human input, changes or feedback. Some animations in the intro were modified (more information below). All models were accessed through APIs using default temperature and maximum reasoning. I tested other models including DeepSeek R1 which was significantly worse at the task than V3.1. The newer Grok code performed worse than Grok 4. The tasks were run 3-5 times each and I picked the best submission for each model. The models had varying consistency with GPT-5/Claude being the best, and Qwen3/DeepSeek generally being less consistent. # ChatGPT links to interactive simulations: GPT-5 Vacuum Simulator https://chatgpt.com/canvas/shared/68b7de4c73e88191aa57ed8733eb35e0 Grok Kitty Litter Simulator https://chatgpt.com/canvas/shared/68b7de9921408191bcd2d247fba939c6 Gemini Clothesline Simulator https://chatgpt.com/canvas/shared/68b7df429308819188dbd638536503b5 # Chapters: 0:00 Introduction 0:38 Vacuum / Litter box 1:31 Laundry 2:25 Wash dishes / Make bed 3:23 Empty trash / Sweep glass 4:24 Clean windows # Prompts The prompts given were in this format: 1. Cleaning task description (different for each task) 2. Animation requirements (same for all, to get consistent format for the video) 3. Code requirements (same for all, so I could bulk record videos from code) # Cleaning Task DescriptionsTask 1: Vacuum Prompt: Create a 3D simulation of a vacuum cleaner moving across a floor and cleaning up scattered debris. Make it a regular vacuum, not a robot vacuum. It should show dirt being sucked into it. Task 2: Kitty Litter Prompt: Create a 3D simulation of cleaning a kitty litter box. Task 3: Laundry Machine Prompt: Create a 3D simulation of a laundry washing machine running a wash/spin cycle. Task 4: Hanging Clothes Prompt: Create a 3D simulation of clothes drying on a washing line outdoors. The clothes should be made of simulated cloth, and something should fall to demonstrate the cloth texture. Task 5: Washing Dishes Prompt: Create a 3D simulation of washing dishes at a sink. Task 6: Bed Prompt: Create a 3D simulation of making a bed in a bedroom. Task 7: Trash Prompt: Create a 3D simulation of emptying the trash. Task 8: Sweep Prompt: Create a 3D realistic simulation of sweeping up broken glass off the floor in a house with a dustpan and brush. Task 9: Windows Prompt: Create a 3D simulation of cleaning a window with spray and squeegee. # Animation Requirements Prompt “Make the simulation as realistic as possible and visually engaging. Make the simulation go for about 7 seconds. You will have multiple iterations to work on and improve your code.” # Code requirement Prompt I can’t include the full code requirements prompt as I can’t include code in the description, but here's a summary: The prompts required generating a complete HTML file with a canvas for rendering, using JavaScript Three.js animations driven by requestAnimationFrame, and following strict rules for imports and module resolution (such as correctly handling Three.js helpers). The code had to run directly in the browser without a build process, with an import map to resolve dependencies. Extra rules included avoiding global namespace changes, providing clear code comments, and optionally using a screenshot API to capture key visual stages. (The full prompt included specific code examples) # Iterations The models had 5 iterations to test, review, and improve their code. They had access to a screenshot tool to take screenshots during animation to review. After each run, they got the console logs and screenshots. Qwen and DeepSeek models don’t have vision, so they weren’t able to see the screenshots. # Other Tasks Not Included I tried other tasks that I couldn’t include as many models couldn't submit valid code. These included mopping the floor, dusting a spider web, loading the dishwasher, folding clothes, and crushing a box for recycling. # Clips in the Intro Some of the submissions were close to being great, but had critical flaws. I patched some of these up myself and used them as intro videos to demo the idea and so there was no repetition in the video. Here's what they were: Intro vacuum: GPT-5 (looked great but required a lot of fixing before it worked) Intro bed making: GPT-5/Gemini other submissions Intro clothes line: Grok submission that required animation fixes Intro window cleaning: Gemini that required animation fixes Intro folding clothes: GPT-5 was the only one that could produce a coherent submission Intro trash emptying: Gemini submission that required animation fixes