AI UNDERDOGSDAILY PICK
AI UNDERDOGS
零依赖，从像素画到光线追踪
Zero Libraries. Pixel to Raytracing
fersab/pixelforge
别人用 AI 写 CRUD
这位老哥让它从零手撸光线追踪——没用任何框架
Everyone else is using AI to build wrappers
This guy made Claude Code write a full
rendering pipeline from scratch — no WebGL, no
Three.js, nothing
三个渲染器，一个项目
Three Renderers, One Project
Pixelforge 是一个纯 ES6 项目
里面塞了三个完整的渲染管线——软件光栅化器
CPU 光线追踪器
还有一个带 BVH 加速和 2×2 抗锯齿的 GPU 光追器
Pixelforge packs three full rendering pipelines into one
pure ES6 project — a software rasterizer, a
CPU raytracer, and a GPU raytracer with BVH
acceleration and 2×2 antialiasing
★ SIGNAL 1
故意选最难的路
The Hard Path, on Purpose
他完全可以套个 Three.js 十分钟搞定
但他偏不——从 Canvas 的像素操作开始
一帧一帧自己算
这不是做产品
这是搞 demoscene 的心态
He could've grabbed Three.js and called it a
day. Instead he starts from raw pixel manipulation
on Canvas and builds every stage by hand
This isn't product thinking — it's demoscene energy
★ SIGNAL 2
BVH 不是入门教程
BVH Is Not Tutorial Stuff
BVH，包围体层级
是光线追踪的核心加速结构——这玩意儿是正经计算机图形学的硬核数学
不是 YouTube 教程里画个球就完事的那种
BVH — Bounding Volume Hierarchy — is the
core acceleration structure behind real raytracing. This is
serious CG math, not the 'draw a sphere'
tutorial stuff
作者自己说的——「今天你看到的 AI 项目大多是 LLM 包装
CRUD 应用
或者几个框架拼在一起
我想让 Claude Code 去做真正的算法苦力活
那种让你回到 demoscene 时代的底层图形编程」
In his own words: 'Most AI-assisted projects today
are LLM wrappers, CRUD apps, or frameworks stitched
together with API calls. I wanted to push
Claude Code in a different direction — heavy
algorithmic work, real-time math, the kind of low-level
graphics programming that takes you back to the
demoscene'
AI UNDERDOGS
三个渲染器，直接在浏览器里跑
Three renderers, runs right in your browser
fersab/pixelforge
关注 · 每天发现更多 AI 神作
fersab.github.io/pixelforge