crediblejhj

Thank you for your detailed reply:)

Great post. I'm a first-year CS student without much math knowledge (also poor English). I'd like to check whether I got the key idea right. There are layers and each layer stores key, value pairs (normalized in this article). If I query Q then Q will be compared with every stored key "Q@A + Q@B + Q@C + ... = value". If A is the answer then Q@A will be 1, others will be interferences. D dimension only has d orthogonal directions, so if I store more keys than the dimension, I can't avoid remaining interferences. What we want is reducing them by going through function F. And this article introduces heuristic capacity results with different memory store functions (linear -> ReLU -> self attention). Is it roughly right?

self attention is the best way to suppress interferences?

What is it?

I ran an experiment to evolve what I call “transparent intelligence” using fewer than 100 neural network weights. Using a genetic algorithm over 3000+ generations, survival improved from 0% to around 50%. However, true discrimination behavior — distinguishing food from poison in the environment — never emerged. After extensive parameter tuning and repeated trials, I concluded that the limitation was likely not hyperparameter tuning, but the representation and the structure of the search space itself. I’ve documented the full process, including the failed attempts and analysis.

Haha thanks for sharing your story.

Hi there! I'm so glad to meet someone my age working on a similar path.

Your project sounds really fascinating as well. To be honest, I wasn't familiar with the concept of "porting" until I looked it up through Gemini just now—it's such an interesting approach! You mentioned that you couldn't quite get it to work despite countless efforts, but I believe that only proves how incredibly difficult and ambitious your attempt was.

During this project, I also faced constant failures and had to compromise on many things because of the sheer complexity. My engine doesn't support JS yet, and only handles a handful of CSS properties—and even those aren't perfect. But as people like us always do, we learn and grow through these failures. I truly hope you’ve gained a lot of growth and insight through your project as well.

About LLMs: I used Claude. The key was using it as a mentor, not a code generator:

-Asked WHY things work, not just HOW

-Got help understanding concepts

-Debugging assistance when stuck

But I designed the architecture and solved problems myself.

Thanks again for the star and your kind words! Let’s keep in touch and keep building.

I used Claude AI as a learning assistant throughout the project.

I had AI help me with:

- Initial layout calculation logic (the layout concept was new to me)

- Catching edge cases in the parser

- Recent refactoring for code quality

But I designed and coded myself:

- The overall rendering pipeline architecture

- CSS parser from scratch

- Debugging

- All architectural decisions

As for understanding - I don't know what understanding exactly is. But I think it means how familiar you become with a concept. By that definition, yes, I understand how the rendering pipeline, parsers, layout tree etc work.

First time using C++, so I needed guidance on concepts I'd never seen before. But the learning, problem-solving, and persistence were mine.

Thanks for asking!