This video details an experiment in using agentic coding, specifically with the Kimmy K2 AI model via the Moonshot AI API and the Rode framework, to develop a web-based open-world game called "Block Walkers." The core concept involves client-side AI-driven NPC conversations utilizing WebGPU and the Transformers.js library with a small Quen language model. The initial attempt to generate the game code proved largely unsuccessful, characterized by extreme slowness in AI responses, context length limitations, and the generation of overly complex, non-functional code. A subsequent attempt, guided by a simplified prompt, still faced significant delays and ultimately failed to produce a fully working game directly from Kimmy K2. The creator eventually refined and completed the prototype with the assistance of Google Gemini 2.5 Pro, resulting in a basic but functional demonstration of client-side AI interactions. This project is now positioned as a foundational "quilting" benchmark for future iterative development by other AI models, serving to explore the capabilities and current limitations of large language models in intricate software engineering tasks, particularly regarding efficiency and complete task execution.
Summarizing the AI Agentic Coding Experiment for "Block Walkers"
0:00 - Project Introduction: The video introduces an experiment to build an open-world web-based game, "Block Walkers," featuring NPCs with AI-driven conversations, leveraging WebGPU and Transformers.js to run AI models entirely client-side.
0:17 - Agentic Coding Setup: The creator aims to use Kimmy K2 (via Moonshot AI API) within the Rode agentic coding framework to generate the game's code, specifically for NPC conversations powered by a tiny Quen model.
1:01 - Initial Attempt & Failure: The first attempt (overnight, 12 hours) failed, generating an excessive amount of non-functional code, including an entire game engine and test scripts, described as "ridiculous" abstractions.
1:33 - Performance Issues: A significant problem identified was Kimmy K2's "extremely slow" response speed through its official API, making agentic coding tasks requiring many tokens impractical and "taxing on one's patience."
2:14 - Moonshot API Costs & Setup Tips: The 12-hour failed attempt cost approximately $3.50. A crucial practical tip for using Moonshot AI with Root Code is to explicitly enable streaming in the settings, which is not documented by Moonshot.
3:38 - Simplified Prompt: For a second attempt, the prompt was stripped down to focus on core game mechanics (open-world city block, NPC collision for chat), low-poly visuals, and the requirement for Transformers.js with the Quento 0.5B Onyx model, explicitly demanding a non-overcomplicated codebase.
5:07 - Continued Challenges: Despite the simplified prompt, Kimmy K2's 128K context length, while reasonable, offered less "comfort and overhead" than larger models like Gemini for complex builds. The slow generation speed remained a major hurdle, with debugging taking up to 10 minutes just to start reading files.
7:41 - Gemini's Intervention: Kimmy K2 successfully "architected out" and generated a large portion of the code, but the game never properly ran on its own. Google Gemini 2.5 Pro was subsequently brought in to "tidy things up" and make the game functional.
8:38 - New "Quilting" Benchmark: The resulting, partially functional prototype serves as a new "agentic coding benchmark." It will be hosted on GitHub and used in future videos for other AI models to iterate upon and add features, embodying a "quilting" concept of collaborative development.
9:21 - Client-Side AI Demonstration: The prototype demonstrates successful client-side NPC interactions using WebGPU to load a very tiny, quantized Quen model, allowing conversational intelligence to run directly in the user's Chrome browser without server-side processing.
10:31 - Future Development Plans: The creator plans to upgrade the Quen model (e.g., to 300.6b), implement voice interaction (e.g., with Whisper for speech-to-text), improve graphics beyond simple blocks, and add more complex gameplay logic.
13:05 - Usability & Accessibility: The use of WebGPU and Transformers.js is highlighted for its ability to simplify deployment for less technical users, avoiding the need for complex local AI setups like Olama. The project will be available on GitHub.