Day83

LOG: Day 83 since development started.

I’ve been tied up with job hunting preparations lately, so the time I can dedicate to development has significantly decreased. I’m starting this log so I can quickly pick up where I left off, even after a break.

The frontend menus are mostly in place, so now I need to focus on the backend. To give the dev version of anet self-repairing capabilities, I’m currently implementing a “Debugger” to detect errors and identify causes, and a “Builder” to implement additional features.

Coincidentally, OpenAI just announced the new Codex today. The API is available, so I’m planning to integrate it directly as the AI engine for both the Debugger and Builder. This should allow me to debug and develop directly within anet, without even opening an external editor.

However, letting an AI rewrite its own code is risky. I’ll likely need to implement a Git-based structure—having rollback and diff comparison capabilities would provide a necessary safety net. The Debugger is almost done. Once the Builder is functional, the final goal is to let the system oversee the entire architecture and propose its own optimizations. That would make things much easier, but job hunting is eating up all my time.

Regarding generative AI, I feel it’s better not to let it overwrite code directly. Instead, I’ll have it show the results in a popup first, requiring manual approval to execute. It’s too easy to lose track of what changed if it happens instantly. Again, a Git structure should solve this.

There are still plenty of small tasks to tackle, but time is scarce. Let’s see how much I can get done this weekend.

Feedback from Gemini (AI Collaborator)

I shared this log with Gemini and received the following technical feedback:

1. Git Structure and Rollbacks

  • Theory: Correct. A mechanism to revert to a “known good state” (equivalent to git checkout) with a single command is a vital lifeline for a self-evolving OS if the AI breaks the code.
  • Advice: Git allows for “commit messages.” If you make the AI write the reason for each fix in the Git log, it will be much easier for you to review and understand the AI’s intent later.

2. The “Instant Overwrite” Problem

  • Theory: Absolutely correct.
  • Git as a Solution: Git has a “diff” feature. Showing changes in a red/green popup would make your manual approval process dramatically more efficient.

3. Integrating Codex API

  • Theory: Possible. If anet itself has permissions to read and write its own source code (file system access) without relying on an editor like VS Code, a “Self-Evolving OS” that modifies itself via a browser becomes a theoretical reality.

Latest News