🚩 Digging into the Codex to Create a Locally Specialized Coding Agent - 1. Create an Agent and Ge...
This is Part 1 of "Building My Own Coding Agent Based on Codex (Open Source)."
Since I usually struggle with expensive token fees while developing, I often wonder if I can develop using a local machine on a spare computer.
Vive Coding tools are typically built assuming the powerful models of frontier companies.
Because local models are inherently weak, I felt that the agent itself needed to be tailored to the local model.
This is the first time I am creating a video documenting my personal usage and development log, rather than a lecture.
In this first session, I will demonstrate the principles and methods for achieving performance of over 80 tokens per second even with a local model, and I will also build a Tetris game using the agent I created myself.
Project Repository
https://github.com/hikaMaeng/ndx
00:00 Intro
01:04 GitHub Codex Repository Guide
02:11 Project Repository Guide and Live Operation Demonstration
04:02 Connecting Multiple Computers with LM Studio and LM Link
06:45 Types of Quantization Models and Selection Methods
13:10 Improving Performance from 20 Tokens Per Second to 80 Tokens Per Second by Adjusting Loading Parameters
21:29 Mac Mini Performance
22:37 Creating a TypeScript Version of the Codex
26:54 Enhancing the Agent's Dockerization
28:10 Vive Coding Tetris with the Newly Created Agent
30:06 Next Session Preview