Ask HN: 寻找“Transformer 替代方案”的发布路径？
2 分•作者: adinhitlore•5 个月前
我有一个副业项目，花了大约 1000 个小时，设定了两个目标：
1. 在 CPU 上比 Transformer 更快；
2. 比 Transformer 更智能。
下面是一些截图（黑色/红色部分暂时被屏蔽）：
[图片链接]
[图片链接]
[图片链接]
总结：这到底是什么？
两种架构：
1. 线性 RNN，解决了当前领先的 RNN Transformer 替代方案（RWKV、Mamba）中存在的长程记忆问题，此外，它对 CPU 友好，完全用 C 语言从头编写，但规模不大：大约 4000 行代码。
2. 两个 SNN 实验程序（最初用 C 语言编写，后来移植到 C# 和 F#），结果比预期的要好，但遗憾的是，目前来说：比线性 RNN 架构要笨（我需要更多测试）。
问题是：该如何处理它们？谷歌 Gemini Pro 3.1/Sonnet 4.6 建议我申请专利、保护知识产权，并估计价值数百万美元，但这显然是个错误：我已经将所有代码上传到 Claude/Gemini 进行分析，但考虑到该项目大约 70% 是用“氛围代码”编写的，如果我像个守门人一样，那就太自命清高了。
问题是：我不想获得数百万美元，但与此同时，我看到了免费开源发布的一些问题：
* 完全不一致，我不相信“AGI 炒作”，但可能存在潜在风险，例如在网络安全方面；
* 坦率地说，我讨厌 Xai 和马斯克，而且由于可能对运行 AI 模型作为 B2C 解决方案感兴趣的公司大约有 20 家，其中一家将是 Xai。
* 非常规的实现：全部用 C 语言编写，并移植到 C#/F#。没有 Python 或 Rust，这意味着一些不熟悉这些语言的 ML 人员可能会遇到问题，因此我将不得不不停地提供支持，这很耗时，而且说实话，一旦开源，我就必须免费提供支持。
* 即使它有潜力，也可能默默无闻地消失在 GitHub 上，除非你中彩票，否则自然流量很少起作用。
顺便说一句，这**不是**炫耀，我坚信有比我更好的程序员，比我更了解 ML 的人，比我更好的数学家，但坦率地说，我拥有一种特殊的毅力和傲慢的结合，这在技术/发明/新颖性方面大有帮助。
就像我说的，这是数百小时工作的结果，并辅以多年在其他领域的编程经验，这**不是**那种“Claude，给我 AGI”式的周末尝试。
所有项目都能在零警告的情况下编译，在逻辑上似乎有效，并且明显比 Transformer 更快，具有泛化和创建新/独特内容的能力。缺少的部分是扩展和在经典基准上的测试。
我缺乏的是对技术应用的理解。10 倍！
查看原文
So, a side project I've spent/wasted ~1000 hours on, with 2 goals set in mind:<p>1. faster than transformers on CPU;
2. smarter than transformers.<p>couple of screenshots below (the black/red part are censored on purpose...for now):<p>https://i.imgur.com/r0equ55.png
https://i.imgur.com/fohRbIr.png
https://i.imgur.com/5Xx1RGX.png<p>Summary: what the hell is this?<p>Two architectures -<p>1. Linear RNN which solves the long memory problem in current front-runner RNN transformer alternatives (RWKV, Mamba), in addition to being cpu friendly and entirely in C from scratch, but not too big: ~4000 lines.<p>2. 2 SNN experimental programs (in C originally but also ported to C# and F#) that turned out to be better than expected but unfortunately for the time being: dumber than the linear RNN one (i need more tests).<p>The question is: what to do with them? google gemini pro 3.1/sonnet 4.6 told me to patent, IP, estimating value in the many millions and while this is clearly a mistake: I've uploaded all the code to claude/gemini for analysis though seeing how the project is ~70% vibecoded I think it would be snobby to act like a gatekeeper.<p>The thing is: I don't want millions but at the same time i see several issues with fee open source rollout:<p>* completely unalighned, i don't believe in the "agi hype" but potential risks may exist, such as in cybersecurity;
* I frankly hate Xai and Musk and since the companies who may be interested in running AI models as b2c solution are likely ~20, one of them will be xai.
* Very unorthodox implementation: All in C with ports in c#/f#. No python or rust, which would mean likely some people unfamiliar with these languages in ML running into issues so i'd have to support nonstop which is time consuming and let's face it i'll have to do it for free once it's open source.
* It may die completely unheard of somewhere on GitHub even if it has potential, organic traffic rarely works unless you hit the lottery.<p>This is NOT a flex btw, I'm convinced there are programmers better than me, people who understand ML better than me, mathematicians better than me though frankly I posses special kind of persistence combined with arrogance which goes a long way in terms of technology/inventions/novelty.<p>Like i said this is the results of hundreds of hours work spiced up by many years programming experience in other areas, this wasn't one weekend "claude, give me agi" kind of shot.<p>All of the projects compile with zero warnings, logically seem to work and are visibly faster than transformers with obvious ability to generalize and create new/unique content. The missing part is scaling and benchmarking on classic benchmarks.<p>What I lack is understanding adoption of technology.10x!